diff --git a/LICENSE.txt b/LICENSE.txt
new file mode 100644
index 0000000..261eeb9
--- /dev/null
+++ b/LICENSE.txt
@@ -0,0 +1,201 @@
+ Apache License
+ Version 2.0, January 2004
+ http://www.apache.org/licenses/
+
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+ 1. Definitions.
+
+ "License" shall mean the terms and conditions for use, reproduction,
+ and distribution as defined by Sections 1 through 9 of this document.
+
+ "Licensor" shall mean the copyright owner or entity authorized by
+ the copyright owner that is granting the License.
+
+ "Legal Entity" shall mean the union of the acting entity and all
+ other entities that control, are controlled by, or are under common
+ control with that entity. For the purposes of this definition,
+ "control" means (i) the power, direct or indirect, to cause the
+ direction or management of such entity, whether by contract or
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
+ outstanding shares, or (iii) beneficial ownership of such entity.
+
+ "You" (or "Your") shall mean an individual or Legal Entity
+ exercising permissions granted by this License.
+
+ "Source" form shall mean the preferred form for making modifications,
+ including but not limited to software source code, documentation
+ source, and configuration files.
+
+ "Object" form shall mean any form resulting from mechanical
+ transformation or translation of a Source form, including but
+ not limited to compiled object code, generated documentation,
+ and conversions to other media types.
+
+ "Work" shall mean the work of authorship, whether in Source or
+ Object form, made available under the License, as indicated by a
+ copyright notice that is included in or attached to the work
+ (an example is provided in the Appendix below).
+
+ "Derivative Works" shall mean any work, whether in Source or Object
+ form, that is based on (or derived from) the Work and for which the
+ editorial revisions, annotations, elaborations, or other modifications
+ represent, as a whole, an original work of authorship. For the purposes
+ of this License, Derivative Works shall not include works that remain
+ separable from, or merely link (or bind by name) to the interfaces of,
+ the Work and Derivative Works thereof.
+
+ "Contribution" shall mean any work of authorship, including
+ the original version of the Work and any modifications or additions
+ to that Work or Derivative Works thereof, that is intentionally
+ submitted to Licensor for inclusion in the Work by the copyright owner
+ or by an individual or Legal Entity authorized to submit on behalf of
+ the copyright owner. For the purposes of this definition, "submitted"
+ means any form of electronic, verbal, or written communication sent
+ to the Licensor or its representatives, including but not limited to
+ communication on electronic mailing lists, source code control systems,
+ and issue tracking systems that are managed by, or on behalf of, the
+ Licensor for the purpose of discussing and improving the Work, but
+ excluding communication that is conspicuously marked or otherwise
+ designated in writing by the copyright owner as "Not a Contribution."
+
+ "Contributor" shall mean Licensor and any individual or Legal Entity
+ on behalf of whom a Contribution has been received by Licensor and
+ subsequently incorporated within the Work.
+
+ 2. Grant of Copyright License. Subject to the terms and conditions of
+ this License, each Contributor hereby grants to You a perpetual,
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+ copyright license to reproduce, prepare Derivative Works of,
+ publicly display, publicly perform, sublicense, and distribute the
+ Work and such Derivative Works in Source or Object form.
+
+ 3. Grant of Patent License. Subject to the terms and conditions of
+ this License, each Contributor hereby grants to You a perpetual,
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+ (except as stated in this section) patent license to make, have made,
+ use, offer to sell, sell, import, and otherwise transfer the Work,
+ where such license applies only to those patent claims licensable
+ by such Contributor that are necessarily infringed by their
+ Contribution(s) alone or by combination of their Contribution(s)
+ with the Work to which such Contribution(s) was submitted. If You
+ institute patent litigation against any entity (including a
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
+ or a Contribution incorporated within the Work constitutes direct
+ or contributory patent infringement, then any patent licenses
+ granted to You under this License for that Work shall terminate
+ as of the date such litigation is filed.
+
+ 4. Redistribution. You may reproduce and distribute copies of the
+ Work or Derivative Works thereof in any medium, with or without
+ modifications, and in Source or Object form, provided that You
+ meet the following conditions:
+
+ (a) You must give any other recipients of the Work or
+ Derivative Works a copy of this License; and
+
+ (b) You must cause any modified files to carry prominent notices
+ stating that You changed the files; and
+
+ (c) You must retain, in the Source form of any Derivative Works
+ that You distribute, all copyright, patent, trademark, and
+ attribution notices from the Source form of the Work,
+ excluding those notices that do not pertain to any part of
+ the Derivative Works; and
+
+ (d) If the Work includes a "NOTICE" text file as part of its
+ distribution, then any Derivative Works that You distribute must
+ include a readable copy of the attribution notices contained
+ within such NOTICE file, excluding those notices that do not
+ pertain to any part of the Derivative Works, in at least one
+ of the following places: within a NOTICE text file distributed
+ as part of the Derivative Works; within the Source form or
+ documentation, if provided along with the Derivative Works; or,
+ within a display generated by the Derivative Works, if and
+ wherever such third-party notices normally appear. The contents
+ of the NOTICE file are for informational purposes only and
+ do not modify the License. You may add Your own attribution
+ notices within Derivative Works that You distribute, alongside
+ or as an addendum to the NOTICE text from the Work, provided
+ that such additional attribution notices cannot be construed
+ as modifying the License.
+
+ You may add Your own copyright statement to Your modifications and
+ may provide additional or different license terms and conditions
+ for use, reproduction, or distribution of Your modifications, or
+ for any such Derivative Works as a whole, provided Your use,
+ reproduction, and distribution of the Work otherwise complies with
+ the conditions stated in this License.
+
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
+ any Contribution intentionally submitted for inclusion in the Work
+ by You to the Licensor shall be under the terms and conditions of
+ this License, without any additional terms or conditions.
+ Notwithstanding the above, nothing herein shall supersede or modify
+ the terms of any separate license agreement you may have executed
+ with Licensor regarding such Contributions.
+
+ 6. Trademarks. This License does not grant permission to use the trade
+ names, trademarks, service marks, or product names of the Licensor,
+ except as required for reasonable and customary use in describing the
+ origin of the Work and reproducing the content of the NOTICE file.
+
+ 7. Disclaimer of Warranty. Unless required by applicable law or
+ agreed to in writing, Licensor provides the Work (and each
+ Contributor provides its Contributions) on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ implied, including, without limitation, any warranties or conditions
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+ PARTICULAR PURPOSE. You are solely responsible for determining the
+ appropriateness of using or redistributing the Work and assume any
+ risks associated with Your exercise of permissions under this License.
+
+ 8. Limitation of Liability. In no event and under no legal theory,
+ whether in tort (including negligence), contract, or otherwise,
+ unless required by applicable law (such as deliberate and grossly
+ negligent acts) or agreed to in writing, shall any Contributor be
+ liable to You for damages, including any direct, indirect, special,
+ incidental, or consequential damages of any character arising as a
+ result of this License or out of the use or inability to use the
+ Work (including but not limited to damages for loss of goodwill,
+ work stoppage, computer failure or malfunction, or any and all
+ other commercial damages or losses), even if such Contributor
+ has been advised of the possibility of such damages.
+
+ 9. Accepting Warranty or Additional Liability. While redistributing
+ the Work or Derivative Works thereof, You may choose to offer,
+ and charge a fee for, acceptance of support, warranty, indemnity,
+ or other liability obligations and/or rights consistent with this
+ License. However, in accepting such obligations, You may act only
+ on Your own behalf and on Your sole responsibility, not on behalf
+ of any other Contributor, and only if You agree to indemnify,
+ defend, and hold each Contributor harmless for any liability
+ incurred by, or claims asserted against, such Contributor by reason
+ of your accepting any such warranty or additional liability.
+
+ END OF TERMS AND CONDITIONS
+
+ APPENDIX: How to apply the Apache License to your work.
+
+ To apply the Apache License to your work, attach the following
+ boilerplate notice, with the fields enclosed by brackets "[]"
+ replaced with your own identifying information. (Don't include
+ the brackets!) The text should be enclosed in the appropriate
+ comment syntax for the file format. We also recommend that a
+ file or class name and description of purpose be included on the
+ same "printed page" as the copyright notice for easier
+ identification within third-party archives.
+
+ Copyright [yyyy] [name of copyright owner]
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License.
diff --git a/README.md b/README.md
index e6fb349..ad4ff1b 100644
--- a/README.md
+++ b/README.md
@@ -4,7 +4,7 @@
-Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation
+Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation (NeurIPS 2025)
[Zhe Kong*](https://scholar.google.com/citations?user=4X3yLwsAAAAJ&hl=zh-CN) · [Feng Gao*](https://scholar.google.com/citations?user=lFkCeoYAAAAJ) ·[Yong Zhang](https://yzhang2016.github.io/)✉ · [Zhuoliang Kang](https://scholar.google.com/citations?user=W1ZXjMkAAAAJ&hl=en) · [Xiaoming Wei](https://scholar.google.com/citations?user=JXV5yrZxj5MC&hl=zh-CN) · [Xunliang Cai](https://openreview.net/profile?id=~Xunliang_Cai1)
@@ -79,13 +79,20 @@ We propose **MultiTalk** , a novel framework for audio-driven multi-person conve
> - ⏱️ **Long Video Generation**: Support video generation up to 15 seconds
## 🔥 Latest News
-
+* Dec 16, 2025: 🚀 We are excited to announce the release of **[LongCat-Video-Avatar](https://github.com/MeiGen-AI/LongCat-Video-Avatar)**, a unified model that delivers expressive and highly dynamic audio-driven character animation, supporting native tasks including Audio-Text-to-Video, Audio-Text-Image-to-Video, and Video Continuation with seamless compatibility for both single-stream and multi-stream audio inputs. The release includes our Technical Report, [code](https://github.com/meituan-longcat/LongCat-Video), [model weights](https://huggingface.co/meituan-longcat/LongCat-Video-Avatar), and [project page](https://meigen-ai.github.io/LongCat-Video-Avatar/).
+* Aug 19, 2025: 🔥🔥 We released [**InfiniteTalk**](https://github.com/MeiGen-AI/InfiniteTalk), a novel new paradigm for video dubbing. InfiniteTalk supports infinite-length video-to-video generation and image-to-video generation. Models, code, gradio, and comfyui have all been released.
+* July 11, 2025: 🔥🔥 `MultiTalk` supports INT8 [quantization](https://github.com/huggingface/optimum-quanto) and [SageAttention2.2](https://github.com/thu-ml/SageAttention), and updates the CFG strategy (2 NFE per step) for FusionX LoRA,
+* July 01, 2025: 🔥🔥 `MultiTalk` supports input audios with TTS, [FusioniX](https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/blob/main/FusionX_LoRa/Wan2.1_I2V_14B_FusionX_LoRA.safetensors) and [lightx2v](https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32.safetensors) LoRA acceleration (requires only 4~8 steps), and Gradio.
* June 14, 2025: 🔥🔥 We release `MultiTalk` with support for `multi-GPU inference`, `teacache acceleration`, `APG` and `low-VRAM inference` (enabling 480P video generation on a single RTX 4090). [APG](https://arxiv.org/abs/2410.02416) is used to alleviate the color error accumulation in long video generation. TeaCache is capable of increasing speed by approximately 2~3x.
* June 9, 2025: 🔥🔥 We release the [weights](https://huggingface.co/MeiGen-AI/MeiGen-MultiTalk) and inference code of **MultiTalk**
* May 29, 2025: We release the [Technique-Report](https://arxiv.org/abs/2505.22647) of **MultiTalk**
* May 29, 2025: We release the [project page](https://meigen-ai.github.io/multi-talk/) of **MultiTalk**
## 🌐 Community Works
+- [Wan2GP](https://github.com/deepbeepmeep/Wan2GP): thank [deepbeepmeep](https://github.com/deepbeepmeep) for providing the project [Wan2GP](https://github.com/deepbeepmeep/Wan2GP) that enables Multitalk on very low VRAM hardware (8 GB of VRAM) and combines it with the capabilities of Vace.
+- [Replicate](https://replicate.com/zsxkib/multitalk): thank [zsxkib](https://github.com/zsxkib) for pushing MultiTalk to Replicate platform, try it! Please refer to [cog-MultiTalk](https://github.com/zsxkib/cog-MultiTalk) for details.
+- [Gradio Demo](https://github.com/MeiGen-AI/MultiTalk): thank [fffiloni](https://github.com/fffiloni) for developing this gradio demo on Hugging Face. Please refer to the [issue](https://github.com/MeiGen-AI/MultiTalk/issues/39) for details.
+- [ComfyUI](https://github.com/kijai/ComfyUI-WanVideoWrapper/tree/multitalk): thank [kijai](https://github.com/kijai) for integrating MultiTalk into ComfyUI-WanVideoWrapper. [Rudra](https://github.com/Rudra-ai-coder) found something interesting that MultiTalk can be combined with Wanx T2V and VACE in the [issue](https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/635).
- [Google Colab example](https://colab.research.google.com/drive/185OyRIpJDlpnRjhBRb7FnaRlq11BLZTa?usp=sharing), an exmaple for inference on A100 provided by [Braffolk](https://github.com/Braffolk).
## 📑 Todo List
@@ -96,13 +103,14 @@ We propose **MultiTalk** , a novel framework for audio-driven multi-person conve
- [x] Multi-GPU Inference
- [ ] Inference acceleration
- [x] TeaCache
- - [ ] int8 quantization
+ - [x] int8 quantization
- [ ] LCM distillation
- [ ] Sparse Attention
- [x] Run with very low VRAM
-- [ ] TTS integration
-- [ ] Gradio demo
+- [x] TTS integration
+- [x] Gradio demo
- [ ] ComfyUI
+- [ ] 1.3B model
## Quick Start
@@ -117,10 +125,11 @@ pip install -U xformers==0.0.28 --index-url https://download.pytorch.org/whl/cu1
```
#### 2. Flash-attn installation:
```
+pip install misaki[en]
pip install ninja
pip install psutil
pip install packaging
-pip install flash_attn
+pip install flash_attn==2.7.4.post1
```
#### 3. Other dependencies
@@ -146,12 +155,15 @@ sudo yum install ffmpeg ffmpeg-devel
| --------------|-------------------------------------------------------------------------------|-------------------------------|
| Wan2.1-I2V-14B-480P | 🤗 [Huggingface](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P) | Base model
| chinese-wav2vec2-base | 🤗 [Huggingface](https://huggingface.co/TencentGameMate/chinese-wav2vec2-base) | Audio encoder
+| Kokoro-82M | 🤗 [Huggingface](https://huggingface.co/hexgrad/Kokoro-82M) | TTS weights
| MeiGen-MultiTalk | 🤗 [Huggingface](https://huggingface.co/MeiGen-AI/MeiGen-MultiTalk) | Our audio condition weights
Download models using huggingface-cli:
``` sh
huggingface-cli download Wan-AI/Wan2.1-I2V-14B-480P --local-dir ./weights/Wan2.1-I2V-14B-480P
huggingface-cli download TencentGameMate/chinese-wav2vec2-base --local-dir ./weights/chinese-wav2vec2-base
+huggingface-cli download TencentGameMate/chinese-wav2vec2-base model.safetensors --revision refs/pr/1 --local-dir ./weights/chinese-wav2vec2-base
+huggingface-cli download hexgrad/Kokoro-82M --local-dir ./weights/Kokoro-82M
huggingface-cli download MeiGen-AI/MeiGen-MultiTalk --local-dir ./weights/MeiGen-MultiTalk
```
@@ -189,6 +201,8 @@ Our model is compatible with both 480P and 720P resolutions. The current code on
--size multitalk-720: generate 720P video.
--use_apg: run with APG.
--teacache_thresh: A coefficient used for TeaCache acceleration
+—-sample_text_guide_scale： When not using LoRA, the optimal value is 5. After applying LoRA, the recommended value is 1.
+—-sample_audio_guide_scale： When not using LoRA, the optimal value is 4. After applying LoRA, the recommended value is 2.
```
#### 1. Single-Person
@@ -205,7 +219,6 @@ python generate_multitalk.py \
--mode streaming \
--use_teacache \
--save_file single_long_exp
-
```
##### 2) Run with very low VRAM
@@ -223,7 +236,6 @@ python generate_multitalk.py \
--num_persistent_param_in_dit 0 \
--use_teacache \
--save_file single_long_lowvram_exp
-
```
##### 3) Multi-GPU inference
@@ -240,9 +252,21 @@ torchrun --nproc_per_node=$GPU_NUM --standalone generate_multitalk.py \
--mode streaming \
--use_teacache \
--save_file single_long_multigpu_exp
-
```
+##### 4) Run with TTS
+```
+python generate_multitalk.py \
+ --ckpt_dir weights/Wan2.1-I2V-14B-480P \
+ --wav2vec_dir 'weights/chinese-wav2vec2-base' \
+ --input_json examples/single_example_tts_1.json \
+ --sample_steps 40 \
+ --mode streaming \
+ --num_persistent_param_in_dit 0 \
+ --use_teacache \
+ --save_file single_long_lowvram_tts_exp \
+ --audio_mode tts
+```
#### 2. Multi-Person
@@ -272,7 +296,6 @@ python generate_multitalk.py \
--num_persistent_param_in_dit 0 \
--use_teacache \
--save_file multi_long_lowvram_exp
-
```
##### 3) Multi-GPU inference
@@ -287,16 +310,142 @@ torchrun --nproc_per_node=$GPU_NUM --standalone generate_multitalk.py \
--sample_steps 40 \
--mode streaming --use_teacache \
--save_file multi_long_multigpu_exp
+```
+
+##### 4) Run with TTS
+
+```
+python generate_multitalk.py \
+ --ckpt_dir weights/Wan2.1-I2V-14B-480P \
+ --wav2vec_dir 'weights/chinese-wav2vec2-base' \
+ --input_json examples/multitalk_example_tts_1.json \
+ --sample_steps 40 \
+ --mode streaming \
+ --num_persistent_param_in_dit 0 \
+ --use_teacache \
+ --save_file multi_long_lowvram_tts_exp \
+ --audio_mode tts
+```
+
+
+#### 3. Run with FusioniX and CausVid(Require only 4~8 steps)
+
+[FusioniX](https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/blob/main/FusionX_LoRa/Wan2.1_I2V_14B_FusionX_LoRA.safetensors) require 8 steps and [lightx2v](https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32.safetensors) requires only 4 steps.
+
+```
+python generate_multitalk.py \
+ --ckpt_dir weights/Wan2.1-I2V-14B-480P \
+ --wav2vec_dir 'weights/chinese-wav2vec2-base' \
+ --input_json examples/single_example_1.json \
+ --lora_dir weights/Wan2.1_I2V_14B_FusionX_LoRA.safetensors \
+ --lora_scale 1.0 \
+ --sample_text_guide_scale 1.0 \
+ --sample_audio_guide_scale 2.0 \
+ --sample_steps 8 \
+ --mode streaming \
+ --num_persistent_param_in_dit 0 \
+ --save_file single_long_lowvram_fusionx_exp \
+ --sample_shift 2
+```
+
+or
+
+```
+python generate_multitalk.py \
+ --ckpt_dir weights/Wan2.1-I2V-14B-480P \
+ --wav2vec_dir 'weights/chinese-wav2vec2-base' \
+ --input_json examples/multitalk_example_2.json \
+ --lora_dir weights/Wan2.1_I2V_14B_FusionX_LoRA.safetensors \
+ --lora_scale 1.0 \
+ --sample_text_guide_scale 1.0 \
+ --sample_audio_guide_scale 2.0 \
+ --sample_steps 8 \
+ --mode streaming \
+ --num_persistent_param_in_dit 0 \
+ --save_file multi_long_lowvram_fusionx_exp \
+
+```
+
+#### 4. Run with the quantization model (Only support run with single gpu)
```
+python generate_multitalk.py \
+ --ckpt_dir weights/Wan2.1-I2V-14B-480P \
+ --wav2vec_dir 'weights/chinese-wav2vec2-base' \
+ --input_json examples/multitalk_example_2.json \
+ --sample_steps 40 \
+ --mode streaming \
+ --use_teacache \
+ --quant int8 \
+ --quant_dir weights/MeiGen-MultiTalk \
+ --num_persistent_param_in_dit 0 \
+ --save_file multi_long_lowvram_exp_quant
+```
+
+Run with lora:
+
+```
+python generate_multitalk.py \
+ --ckpt_dir weights/Wan2.1-I2V-14B-480P \
+ --wav2vec_dir 'weights/chinese-wav2vec2-base' \
+ --input_json examples/multitalk_example_1.json \
+ --quant int8 \
+ --quant_dir weights/MeiGen-MultiTalk \
+ --lora_dir weights/MeiGen-MultiTalk/quant_models/quant_model_int8_FusionX.safetensors \
+ --sample_text_guide_scale 1.0 \
+ --sample_audio_guide_scale 2.0 \
+ --sample_steps 8 \
+ --mode streaming \
+ --num_persistent_param_in_dit 0 \
+ --save_file multi_long_lowvram_fusionx_exp_quant \
+ --sample_shift 2
+```
+
+#### 5. Run with Gradio
+
+
+
+```
+python app.py \
+ --lora_dir weights/Wan2.1_I2V_14B_FusionX_LoRA.safetensors \
+ --lora_scale 1.0 \
+ --num_persistent_param_in_dit 0 \
+ --sample_shift 2
+```
+
+or
+
+```
+python app.py --num_persistent_param_in_dit 0
+```
+
+or
+
+```
+python app.py \
+ --quant int8 \
+ --quant_dir weights/MeiGen-MultiTalk \
+ --lora_dir weights/MeiGen-MultiTalk/quant_models/quant_model_int8_FusionX.safetensors \
+ --sample_shift 2 \
+ --num_persistent_param_in_dit 0
+```
## 🚀Computational Efficiency
+
+#### 1) Non quantitative results
+
The results are evaluated on A100 GPUs for multi-person generation. Single-person generation uses less memory and provides faster inference.
TeaCache is capable of increasing speed by approximately 2~3x.
+#### 2) Quantitative results
+
+
+ 
+
+
## 📚 Citation
diff --git a/app.py b/app.py
new file mode 100644
index 0000000..4a05a22
--- /dev/null
+++ b/app.py
@@ -0,0 +1,787 @@
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import argparse
+import logging
+import os
+os.environ["no_proxy"] = "localhost,127.0.0.1,::1"
+import sys
+import json
+import warnings
+from datetime import datetime
+
+import gradio as gr
+warnings.filterwarnings('ignore')
+
+import random
+
+import torch
+import torch.distributed as dist
+from PIL import Image
+import subprocess
+
+import wan
+from wan.configs import SIZE_CONFIGS, SUPPORTED_SIZES, WAN_CONFIGS
+from wan.utils.utils import cache_image, cache_video, str2bool
+from wan.utils.multitalk_utils import save_video_ffmpeg
+from kokoro import KPipeline
+from transformers import Wav2Vec2FeatureExtractor
+from src.audio_analysis.wav2vec2 import Wav2Vec2Model
+
+import librosa
+import pyloudnorm as pyln
+import numpy as np
+from einops import rearrange
+import soundfile as sf
+import re
+
+def _validate_args(args):
+ # Basic check
+ assert args.ckpt_dir is not None, "Please specify the checkpoint directory."
+ assert args.task in WAN_CONFIGS, f"Unsupport task: {args.task}"
+
+ # The default sampling steps are 40 for image-to-video tasks and 50 for text-to-video tasks.
+ if args.sample_steps is None:
+ args.sample_steps = 40
+
+ if args.sample_shift is None:
+ if args.size == 'multitalk-480':
+ args.sample_shift = 7
+ elif args.size == 'multitalk-720':
+ args.sample_shift = 11
+ else:
+ raise NotImplementedError(f'Not supported size')
+
+ args.base_seed = args.base_seed if args.base_seed >= 0 else random.randint(
+ 0, 99999999)
+ # Size check
+ assert args.size in SUPPORTED_SIZES[
+ args.
+ task], f"Unsupport size {args.size} for task {args.task}, supported sizes are: {', '.join(SUPPORTED_SIZES[args.task])}"
+
+
+def _parse_args():
+ parser = argparse.ArgumentParser(
+ description="Generate a image or video from a text prompt or image using Wan"
+ )
+ parser.add_argument(
+ "--task",
+ type=str,
+ default="multitalk-14B",
+ choices=list(WAN_CONFIGS.keys()),
+ help="The task to run.")
+ parser.add_argument(
+ "--size",
+ type=str,
+ default="multitalk-480",
+ choices=list(SIZE_CONFIGS.keys()),
+ help="The buckget size of the generated video. The aspect ratio of the output video will follow that of the input image."
+ )
+ parser.add_argument(
+ "--frame_num",
+ type=int,
+ default=81,
+ help="How many frames to be generated in one clip. The number should be 4n+1"
+ )
+ parser.add_argument(
+ "--ckpt_dir",
+ type=str,
+ default='./weights/Wan2.1-I2V-14B-480P',
+ help="The path to the Wan checkpoint directory.")
+ parser.add_argument(
+ "--quant_dir",
+ type=str,
+ default=None,
+ help="The path to the Wan quant checkpoint directory.")
+ parser.add_argument(
+ "--wav2vec_dir",
+ type=str,
+ default='./weights/chinese-wav2vec2-base',
+ help="The path to the wav2vec checkpoint directory.")
+ parser.add_argument(
+ "--lora_dir",
+ type=str,
+ nargs='+',
+ default=None,
+ help="The path to the LoRA checkpoint directory.")
+ parser.add_argument(
+ "--lora_scale",
+ type=float,
+ nargs='+',
+ default=[1.2],
+ help="Controls how much to influence the outputs with the LoRA parameters. Accepts multiple float values."
+ )
+ parser.add_argument(
+ "--offload_model",
+ type=str2bool,
+ default=None,
+ help="Whether to offload the model to CPU after each model forward, reducing GPU memory usage."
+ )
+ parser.add_argument(
+ "--ulysses_size",
+ type=int,
+ default=1,
+ help="The size of the ulysses parallelism in DiT.")
+ parser.add_argument(
+ "--ring_size",
+ type=int,
+ default=1,
+ help="The size of the ring attention parallelism in DiT.")
+ parser.add_argument(
+ "--t5_fsdp",
+ action="store_true",
+ default=False,
+ help="Whether to use FSDP for T5.")
+ parser.add_argument(
+ "--t5_cpu",
+ action="store_true",
+ default=False,
+ help="Whether to place T5 model on CPU.")
+ parser.add_argument(
+ "--dit_fsdp",
+ action="store_true",
+ default=False,
+ help="Whether to use FSDP for DiT.")
+ parser.add_argument(
+ "--save_file",
+ type=str,
+ default=None,
+ help="The file to save the generated image or video to.")
+ parser.add_argument(
+ "--audio_save_dir",
+ type=str,
+ default='save_audio/gradio',
+ help="The path to save the audio embedding.")
+ parser.add_argument(
+ "--base_seed",
+ type=int,
+ default=42,
+ help="The seed to use for generating the image or video.")
+ parser.add_argument(
+ "--input_json",
+ type=str,
+ default='examples.json',
+ help="[meta file] The condition path to generate the video.")
+ parser.add_argument(
+ "--motion_frame",
+ type=int,
+ default=25,
+ help="Driven frame length used in the mode of long video genration.")
+ parser.add_argument(
+ "--mode",
+ type=str,
+ default="streaming",
+ choices=['clip', 'streaming'],
+ help="clip: generate one video chunk, streaming: long video generation")
+ parser.add_argument(
+ "--sample_steps", type=int, default=None, help="The sampling steps.")
+ parser.add_argument(
+ "--sample_shift",
+ type=float,
+ default=None,
+ help="Sampling shift factor for flow matching schedulers.")
+ parser.add_argument(
+ "--sample_text_guide_scale",
+ type=float,
+ default=5.0,
+ help="Classifier free guidance scale for text control.")
+ parser.add_argument(
+ "--sample_audio_guide_scale",
+ type=float,
+ default=4.0,
+ help="Classifier free guidance scale for audio control.")
+ parser.add_argument(
+ "--num_persistent_param_in_dit",
+ type=int,
+ default=None,
+ required=False,
+ help="Maximum parameter quantity retained in video memory, small number to reduce VRAM required",
+ )
+ parser.add_argument(
+ "--use_teacache",
+ action="store_true",
+ default=False,
+ help="Enable teacache for video generation."
+ )
+ parser.add_argument(
+ "--teacache_thresh",
+ type=float,
+ default=0.2,
+ help="Threshold for teacache."
+ )
+ parser.add_argument(
+ "--use_apg",
+ action="store_true",
+ default=False,
+ help="Enable adaptive projected guidance for video generation (APG)."
+ )
+ parser.add_argument(
+ "--apg_momentum",
+ type=float,
+ default=-0.75,
+ help="Momentum used in adaptive projected guidance (APG)."
+ )
+ parser.add_argument(
+ "--apg_norm_threshold",
+ type=float,
+ default=55,
+ help="Norm threshold used in adaptive projected guidance (APG)."
+ )
+ parser.add_argument(
+ "--color_correction_strength",
+ type=float,
+ default=1.0,
+ help="strength for color correction [0.0 -- 1.0]."
+ )
+
+ parser.add_argument(
+ "--quant",
+ type=str,
+ default=None,
+ help="Quantization type, must be 'int8' or 'fp8'."
+ )
+ args = parser.parse_args()
+ _validate_args(args)
+ return args
+
+
+def custom_init(device, wav2vec):
+ audio_encoder = Wav2Vec2Model.from_pretrained(wav2vec, local_files_only=True).to(device)
+ audio_encoder.feature_extractor._freeze_parameters()
+ wav2vec_feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(wav2vec, local_files_only=True)
+ return wav2vec_feature_extractor, audio_encoder
+
+def loudness_norm(audio_array, sr=16000, lufs=-23):
+ meter = pyln.Meter(sr)
+ loudness = meter.integrated_loudness(audio_array)
+ if abs(loudness) > 100:
+ return audio_array
+ normalized_audio = pyln.normalize.loudness(audio_array, loudness, lufs)
+ return normalized_audio
+
+def audio_prepare_multi(left_path, right_path, audio_type, sample_rate=16000):
+ if not (left_path=='None' or right_path=='None'):
+ human_speech_array1 = audio_prepare_single(left_path)
+ human_speech_array2 = audio_prepare_single(right_path)
+ elif left_path=='None':
+ human_speech_array2 = audio_prepare_single(right_path)
+ human_speech_array1 = np.zeros(human_speech_array2.shape[0])
+ elif right_path=='None':
+ human_speech_array1 = audio_prepare_single(left_path)
+ human_speech_array2 = np.zeros(human_speech_array1.shape[0])
+
+ if audio_type=='para':
+ new_human_speech1 = human_speech_array1
+ new_human_speech2 = human_speech_array2
+ elif audio_type=='add':
+ new_human_speech1 = np.concatenate([human_speech_array1[: human_speech_array1.shape[0]], np.zeros(human_speech_array2.shape[0])])
+ new_human_speech2 = np.concatenate([np.zeros(human_speech_array1.shape[0]), human_speech_array2[:human_speech_array2.shape[0]]])
+ sum_human_speechs = new_human_speech1 + new_human_speech2
+ return new_human_speech1, new_human_speech2, sum_human_speechs
+
+def _init_logging(rank):
+ # logging
+ if rank == 0:
+ # set format
+ logging.basicConfig(
+ level=logging.INFO,
+ format="[%(asctime)s] %(levelname)s: %(message)s",
+ handlers=[logging.StreamHandler(stream=sys.stdout)])
+ else:
+ logging.basicConfig(level=logging.ERROR)
+
+def get_embedding(speech_array, wav2vec_feature_extractor, audio_encoder, sr=16000, device='cpu'):
+ audio_duration = len(speech_array) / sr
+ video_length = audio_duration * 25 # Assume the video fps is 25
+
+ # wav2vec_feature_extractor
+ audio_feature = np.squeeze(
+ wav2vec_feature_extractor(speech_array, sampling_rate=sr).input_values
+ )
+ audio_feature = torch.from_numpy(audio_feature).float().to(device=device)
+ audio_feature = audio_feature.unsqueeze(0)
+
+ # audio encoder
+ with torch.no_grad():
+ embeddings = audio_encoder(audio_feature, seq_len=int(video_length), output_hidden_states=True)
+
+ if len(embeddings) == 0:
+ print("Fail to extract audio embedding")
+ return None
+
+ audio_emb = torch.stack(embeddings.hidden_states[1:], dim=1).squeeze(0)
+ audio_emb = rearrange(audio_emb, "b s d -> s b d")
+
+ audio_emb = audio_emb.cpu().detach()
+ return audio_emb
+
+def extract_audio_from_video(filename, sample_rate):
+ raw_audio_path = filename.split('/')[-1].split('.')[0]+'.wav'
+ ffmpeg_command = [
+ "ffmpeg",
+ "-y",
+ "-i",
+ str(filename),
+ "-vn",
+ "-acodec",
+ "pcm_s16le",
+ "-ar",
+ "16000",
+ "-ac",
+ "2",
+ str(raw_audio_path),
+ ]
+ subprocess.run(ffmpeg_command, check=True)
+ human_speech_array, sr = librosa.load(raw_audio_path, sr=sample_rate)
+ human_speech_array = loudness_norm(human_speech_array, sr)
+ os.remove(raw_audio_path)
+
+ return human_speech_array
+
+def audio_prepare_single(audio_path, sample_rate=16000):
+ ext = os.path.splitext(audio_path)[1].lower()
+ if ext in ['.mp4', '.mov', '.avi', '.mkv']:
+ human_speech_array = extract_audio_from_video(audio_path, sample_rate)
+ return human_speech_array
+ else:
+ human_speech_array, sr = librosa.load(audio_path, sr=sample_rate)
+ human_speech_array = loudness_norm(human_speech_array, sr)
+ return human_speech_array
+
+def process_tts_single(text, save_dir, voice1):
+ s1_sentences = []
+
+ pipeline = KPipeline(lang_code='a', repo_id='weights/Kokoro-82M')
+
+ voice_tensor = torch.load(voice1, weights_only=True)
+ generator = pipeline(
+ text, voice=voice_tensor, # <= change voice here
+ speed=1, split_pattern=r'\n+'
+ )
+ audios = []
+ for i, (gs, ps, audio) in enumerate(generator):
+ audios.append(audio)
+ audios = torch.concat(audios, dim=0)
+ s1_sentences.append(audios)
+ s1_sentences = torch.concat(s1_sentences, dim=0)
+ save_path1 =f'{save_dir}/s1.wav'
+ sf.write(save_path1, s1_sentences, 24000) # save each audio file
+ s1, _ = librosa.load(save_path1, sr=16000)
+ return s1, save_path1
+
+
+
+def process_tts_multi(text, save_dir, voice1, voice2):
+ pattern = r'\(s(\d+)\)\s*(.*?)(?=\s*\(s\d+\)|$)'
+ matches = re.findall(pattern, text, re.DOTALL)
+
+ s1_sentences = []
+ s2_sentences = []
+
+ pipeline = KPipeline(lang_code='a', repo_id='weights/Kokoro-82M')
+ for idx, (speaker, content) in enumerate(matches):
+ if speaker == '1':
+ voice_tensor = torch.load(voice1, weights_only=True)
+ generator = pipeline(
+ content, voice=voice_tensor, # <= change voice here
+ speed=1, split_pattern=r'\n+'
+ )
+ audios = []
+ for i, (gs, ps, audio) in enumerate(generator):
+ audios.append(audio)
+ audios = torch.concat(audios, dim=0)
+ s1_sentences.append(audios)
+ s2_sentences.append(torch.zeros_like(audios))
+ elif speaker == '2':
+ voice_tensor = torch.load(voice2, weights_only=True)
+ generator = pipeline(
+ content, voice=voice_tensor, # <= change voice here
+ speed=1, split_pattern=r'\n+'
+ )
+ audios = []
+ for i, (gs, ps, audio) in enumerate(generator):
+ audios.append(audio)
+ audios = torch.concat(audios, dim=0)
+ s2_sentences.append(audios)
+ s1_sentences.append(torch.zeros_like(audios))
+
+ s1_sentences = torch.concat(s1_sentences, dim=0)
+ s2_sentences = torch.concat(s2_sentences, dim=0)
+ sum_sentences = s1_sentences + s2_sentences
+ save_path1 =f'{save_dir}/s1.wav'
+ save_path2 =f'{save_dir}/s2.wav'
+ save_path_sum = f'{save_dir}/sum.wav'
+ sf.write(save_path1, s1_sentences, 24000) # save each audio file
+ sf.write(save_path2, s2_sentences, 24000)
+ sf.write(save_path_sum, sum_sentences, 24000)
+
+ s1, _ = librosa.load(save_path1, sr=16000)
+ s2, _ = librosa.load(save_path2, sr=16000)
+ # sum, _ = librosa.load(save_path_sum, sr=16000)
+ return s1, s2, save_path_sum
+
+def run_graio_demo(args):
+ rank = int(os.getenv("RANK", 0))
+ world_size = int(os.getenv("WORLD_SIZE", 1))
+ local_rank = int(os.getenv("LOCAL_RANK", 0))
+ device = local_rank
+ _init_logging(rank)
+
+ if args.offload_model is None:
+ args.offload_model = False if world_size > 1 else True
+ logging.info(
+ f"offload_model is not specified, set to {args.offload_model}.")
+ if world_size > 1:
+ torch.cuda.set_device(local_rank)
+ dist.init_process_group(
+ backend="nccl",
+ init_method="env://",
+ rank=rank,
+ world_size=world_size)
+ else:
+ assert not (
+ args.t5_fsdp or args.dit_fsdp
+ ), f"t5_fsdp and dit_fsdp are not supported in non-distributed environments."
+ assert not (
+ args.ulysses_size > 1 or args.ring_size > 1
+ ), f"context parallel are not supported in non-distributed environments."
+
+ if args.ulysses_size > 1 or args.ring_size > 1:
+ assert args.ulysses_size * args.ring_size == world_size, f"The number of ulysses_size and ring_size should be equal to the world size."
+ from xfuser.core.distributed import (
+ init_distributed_environment,
+ initialize_model_parallel,
+ )
+ init_distributed_environment(
+ rank=dist.get_rank(), world_size=dist.get_world_size())
+
+ initialize_model_parallel(
+ sequence_parallel_degree=dist.get_world_size(),
+ ring_degree=args.ring_size,
+ ulysses_degree=args.ulysses_size,
+ )
+
+
+ cfg = WAN_CONFIGS[args.task]
+ if args.ulysses_size > 1:
+ assert cfg.num_heads % args.ulysses_size == 0, f"`{cfg.num_heads=}` cannot be divided evenly by `{args.ulysses_size=}`."
+
+ logging.info(f"Generation job args: {args}")
+ logging.info(f"Generation model config: {cfg}")
+
+ if dist.is_initialized():
+ base_seed = [args.base_seed] if rank == 0 else [None]
+ dist.broadcast_object_list(base_seed, src=0)
+ args.base_seed = base_seed[0]
+
+ assert args.task == "multitalk-14B", 'You should choose multitalk in args.task.'
+
+
+
+ wav2vec_feature_extractor, audio_encoder= custom_init('cpu', args.wav2vec_dir)
+ os.makedirs(args.audio_save_dir,exist_ok=True)
+
+
+ logging.info("Creating MultiTalk pipeline.")
+ # wan_i2v = None
+ wan_i2v = wan.MultiTalkPipeline(
+ config=cfg,
+ checkpoint_dir=args.ckpt_dir,
+ quant_dir=args.quant_dir,
+ device_id=device,
+ rank=rank,
+ t5_fsdp=args.t5_fsdp,
+ dit_fsdp=args.dit_fsdp,
+ use_usp=(args.ulysses_size > 1 or args.ring_size > 1),
+ t5_cpu=args.t5_cpu,
+ lora_dir=args.lora_dir,
+ lora_scales=args.lora_scale,
+ quant=args.quant
+ )
+
+ if args.num_persistent_param_in_dit is not None:
+ wan_i2v.vram_management = True
+ wan_i2v.enable_vram_management(
+ num_persistent_param_in_dit=args.num_persistent_param_in_dit
+ )
+
+
+
+ def generate_video(img2vid_image, img2vid_prompt, n_prompt, img2vid_audio_1, img2vid_audio_2,
+ sd_steps, seed, text_guide_scale, audio_guide_scale, mode_selector, tts_text, resolution_select, human1_voice, human2_voice):
+ input_data = {}
+ input_data["prompt"] = img2vid_prompt
+ input_data["cond_image"] = img2vid_image
+ person = {}
+ if mode_selector == "Single Person(Local File)":
+ person['person1'] = img2vid_audio_1
+ elif mode_selector == "Single Person(TTS)":
+ tts_audio = {}
+ tts_audio['text'] = tts_text
+ tts_audio['human1_voice'] = human1_voice
+ input_data["tts_audio"] = tts_audio
+ elif mode_selector == "Multi Person(Local File, audio add)":
+ person['person1'] = img2vid_audio_1
+ person['person2'] = img2vid_audio_2
+ input_data["audio_type"] = 'add'
+ elif mode_selector == "Multi Person(Local File, audio parallel)":
+ person['person1'] = img2vid_audio_1
+ person['person2'] = img2vid_audio_2
+ input_data["audio_type"] = 'para'
+ else:
+ tts_audio = {}
+ tts_audio['text'] = tts_text
+ tts_audio['human1_voice'] = human1_voice
+ tts_audio['human2_voice'] = human2_voice
+ input_data["tts_audio"] = tts_audio
+
+ input_data["cond_audio"] = person
+
+ if 'Local File' in mode_selector:
+ if len(input_data['cond_audio'])==2:
+ new_human_speech1, new_human_speech2, sum_human_speechs = audio_prepare_multi(input_data['cond_audio']['person1'], input_data['cond_audio']['person2'], input_data['audio_type'])
+ audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ audio_embedding_2 = get_embedding(new_human_speech2, wav2vec_feature_extractor, audio_encoder)
+ emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ emb2_path = os.path.join(args.audio_save_dir, '2.pt')
+ sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
+ sf.write(sum_audio, sum_human_speechs, 16000)
+ torch.save(audio_embedding_1, emb1_path)
+ torch.save(audio_embedding_2, emb2_path)
+ input_data['cond_audio']['person1'] = emb1_path
+ input_data['cond_audio']['person2'] = emb2_path
+ input_data['video_audio'] = sum_audio
+ elif len(input_data['cond_audio'])==1:
+ human_speech = audio_prepare_single(input_data['cond_audio']['person1'])
+ audio_embedding = get_embedding(human_speech, wav2vec_feature_extractor, audio_encoder)
+ emb_path = os.path.join(args.audio_save_dir, '1.pt')
+ sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
+ sf.write(sum_audio, human_speech, 16000)
+ torch.save(audio_embedding, emb_path)
+ input_data['cond_audio']['person1'] = emb_path
+ input_data['video_audio'] = sum_audio
+ elif 'TTS' in mode_selector:
+ if 'human2_voice' not in input_data['tts_audio'].keys():
+ new_human_speech1, sum_audio = process_tts_single(input_data['tts_audio']['text'], args.audio_save_dir, input_data['tts_audio']['human1_voice'])
+ audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ torch.save(audio_embedding_1, emb1_path)
+ input_data['cond_audio']['person1'] = emb1_path
+ input_data['video_audio'] = sum_audio
+ else:
+ new_human_speech1, new_human_speech2, sum_audio = process_tts_multi(input_data['tts_audio']['text'], args.audio_save_dir, input_data['tts_audio']['human1_voice'], input_data['tts_audio']['human2_voice'])
+ audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ audio_embedding_2 = get_embedding(new_human_speech2, wav2vec_feature_extractor, audio_encoder)
+ emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ emb2_path = os.path.join(args.audio_save_dir, '2.pt')
+ torch.save(audio_embedding_1, emb1_path)
+ torch.save(audio_embedding_2, emb2_path)
+ input_data['cond_audio']['person1'] = emb1_path
+ input_data['cond_audio']['person2'] = emb2_path
+ input_data['video_audio'] = sum_audio
+
+
+ # if len(input_data['cond_audio'])==2:
+ # new_human_speech1, new_human_speech2, sum_human_speechs = audio_prepare_multi(input_data['cond_audio']['person1'], input_data['cond_audio']['person2'], input_data['audio_type'])
+ # audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ # audio_embedding_2 = get_embedding(new_human_speech2, wav2vec_feature_extractor, audio_encoder)
+ # emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ # emb2_path = os.path.join(args.audio_save_dir, '2.pt')
+ # sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
+ # sf.write(sum_audio, sum_human_speechs, 16000)
+ # torch.save(audio_embedding_1, emb1_path)
+ # torch.save(audio_embedding_2, emb2_path)
+ # input_data['cond_audio']['person1'] = emb1_path
+ # input_data['cond_audio']['person2'] = emb2_path
+ # input_data['video_audio'] = sum_audio
+ # elif len(input_data['cond_audio'])==1:
+ # human_speech = audio_prepare_single(input_data['cond_audio']['person1'])
+ # audio_embedding = get_embedding(human_speech, wav2vec_feature_extractor, audio_encoder)
+ # emb_path = os.path.join(args.audio_save_dir, '1.pt')
+ # sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
+ # sf.write(sum_audio, human_speech, 16000)
+ # torch.save(audio_embedding, emb_path)
+ # input_data['cond_audio']['person1'] = emb_path
+ # input_data['video_audio'] = sum_audio
+
+ logging.info("Generating video ...")
+ video = wan_i2v.generate(
+ input_data,
+ size_buckget=resolution_select,
+ motion_frame=args.motion_frame,
+ frame_num=args.frame_num,
+ shift=args.sample_shift,
+ sampling_steps=sd_steps,
+ text_guide_scale=text_guide_scale,
+ audio_guide_scale=audio_guide_scale,
+ seed=seed,
+ n_prompt=n_prompt,
+ offload_model=args.offload_model,
+ max_frames_num=args.frame_num if args.mode == 'clip' else 1000,
+ color_correction_strength = args.color_correction_strength,
+ extra_args=args,
+ )
+
+
+ if args.save_file is None:
+ formatted_time = datetime.now().strftime("%Y%m%d_%H%M%S")
+ formatted_prompt = input_data['prompt'].replace(" ", "_").replace("/",
+ "_")[:50]
+ args.save_file = f"{args.task}_{args.size.replace('*','x') if sys.platform=='win32' else args.size}_{args.ulysses_size}_{args.ring_size}_{formatted_prompt}_{formatted_time}"
+
+ logging.info(f"Saving generated video to {args.save_file}.mp4")
+ save_video_ffmpeg(video, args.save_file, [input_data['video_audio']], high_quality_save=False)
+ logging.info("Finished.")
+
+ return args.save_file + '.mp4'
+
+ def toggle_audio_mode(mode):
+ if 'TTS' in mode:
+ return [
+ gr.Audio(visible=False, interactive=False),
+ gr.Audio(visible=False, interactive=False),
+ gr.Textbox(visible=True, interactive=True)
+ ]
+ elif 'Single' in mode:
+ return [
+ gr.Audio(visible=True, interactive=True),
+ gr.Audio(visible=False, interactive=False),
+ gr.Textbox(visible=False, interactive=False)
+ ]
+ else:
+ return [
+ gr.Audio(visible=True, interactive=True),
+ gr.Audio(visible=True, interactive=True),
+ gr.Textbox(visible=False, interactive=False)
+ ]
+
+
+ with gr.Blocks() as demo:
+
+ gr.Markdown("""
+
+ MeiGen-MultiTalk
+
+
+ Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation.
+
+
+
+
+ """)
+
+ with gr.Row():
+ with gr.Column(scale=1):
+ img2vid_image = gr.Image(
+ type="filepath",
+ label="Upload Input Image",
+ elem_id="image_upload",
+ )
+ img2vid_prompt = gr.Textbox(
+ label="Prompt",
+ placeholder="Describe the video you want to generate",
+ )
+
+
+ with gr.Accordion("Audio Options", open=True):
+ mode_selector = gr.Radio(
+ choices=["Single Person(Local File)", "Single Person(TTS)", "Multi Person(Local File, audio add)", "Multi Person(Local File, audio parallel)", "Multi Person(TTS)"],
+ label="Select person and audio mode.",
+ value="Single Person(Local File)"
+ )
+ resolution_select = gr.Radio(
+ choices=["multitalk-480", "multitalk-720"],
+ label="Select resolution.",
+ value="multitalk-480"
+ )
+ img2vid_audio_1 = gr.Audio(label="Conditioning Audio for speaker 1", type="filepath", visible=True)
+ img2vid_audio_2 = gr.Audio(label="Conditioning Audio for speaker 2", type="filepath", visible=False)
+ tts_text = gr.Textbox(
+ label="Text for TTS",
+ placeholder="Refer to the format in the examples",
+ visible=False,
+ interactive=False
+ )
+ mode_selector.change(
+ fn=toggle_audio_mode,
+ inputs=mode_selector,
+ outputs=[img2vid_audio_1, img2vid_audio_2, tts_text]
+ )
+
+ with gr.Accordion("Advanced Options", open=False):
+ with gr.Row():
+ sd_steps = gr.Slider(
+ label="Diffusion steps",
+ minimum=1,
+ maximum=1000,
+ value=8,
+ step=1)
+ seed = gr.Slider(
+ label="Seed",
+ minimum=-1,
+ maximum=2147483647,
+ step=1,
+ value=42)
+ with gr.Row():
+ text_guide_scale = gr.Slider(
+ label="Text Guide scale",
+ minimum=0,
+ maximum=20,
+ value=1.0,
+ step=1)
+ audio_guide_scale = gr.Slider(
+ label="Audio Guide scale",
+ minimum=0,
+ maximum=20,
+ value=2.0,
+ step=1)
+ with gr.Row():
+ human1_voice = gr.Textbox(
+ label="Voice for the left person",
+ value="weights/Kokoro-82M/voices/am_adam.pt",
+ )
+ human2_voice = gr.Textbox(
+ label="Voice for right person",
+ value="weights/Kokoro-82M/voices/af_heart.pt"
+ )
+ # with gr.Row():
+ n_prompt = gr.Textbox(
+ label="Negative Prompt",
+ placeholder="Describe the negative prompt you want to add",
+ value="bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+ )
+
+ run_i2v_button = gr.Button("Generate Video")
+
+ with gr.Column(scale=2):
+ result_gallery = gr.Video(
+ label='Generated Video', interactive=False, height=600, )
+
+ gr.Examples(
+ examples = [
+ ["examples/single/single1.png", "A woman is passionately singing into a professional microphone in a recording studio. She wears large black headphones and a dark cardigan over a gray top. Her long, wavy brown hair frames her face as she looks slightly upwards, her mouth open mid-song. The studio is equipped with various audio equipment, including a mixing console and a keyboard, with soundproofing panels on the walls. The lighting is warm and focused on her, creating a professional and intimate atmosphere. A close-up shot captures her expressive performance.", "Single Person(Local File)", "examples/single/1.wav", None, None],
+ ["examples/single/single1.png", "A woman is passionately singing into a professional microphone in a recording studio. She wears large black headphones and a dark cardigan over a gray top. Her long, wavy brown hair frames her face as she looks slightly upwards, her mouth open mid-song. The studio is equipped with various audio equipment, including a mixing console and a keyboard, with soundproofing panels on the walls. The lighting is warm and focused on her, creating a professional and intimate atmosphere. A close-up shot captures her expressive performance.", "Single Person(TTS)", None, None, "Welcome to multi-talk, this is an audio-driven video generation model for multiple person."],
+ ["examples/multi/1/multi1.png", "In a casual, intimate setting, a man and a woman are engaged in a heartfelt conversation inside a car. The man, sporting a denim jacket over a blue shirt, sits attentively with a seatbelt fastened, his gaze fixed on the woman beside him. The woman, wearing a black tank top and a denim jacket draped over her shoulders, smiles warmly, her eyes reflecting genuine interest and connection. The car's interior, with its beige seats and simple design, provides a backdrop that emphasizes their interaction. The scene captures a moment of shared understanding and connection, set against the soft, diffused light of an overcast day. A medium shot from a slightly angled perspective, focusing on their expressions and body language.", "Multi Person(Local File, audio add)", "examples/multi/1/1.WAV", "examples/multi/1/2.WAV", None],
+ ["examples/multi/3/multi3.png", "In a cozy recording studio, a man and a woman are singing together. The man, with tousled brown hair, stands to the left, wearing a light green button-down shirt. His gaze is directed towards the woman, who is smiling warmly. She, with wavy dark hair, is dressed in a black floral dress and stands to the right, her eyes closed in enjoyment. Between them is a professional microphone, capturing their harmonious voices. The background features wooden panels and various audio equipment, creating an intimate and focused atmosphere. The lighting is soft and warm, highlighting their expressions and the intimate setting. A medium shot captures their interaction closely.", "Multi Person(Local File, audio parallel)", "examples/multi/3/1-man.WAV", "examples/multi/3/1-woman.WAV", None],
+ ["examples/multi/1/multi1.png", "In a casual, intimate setting, a man and a woman are engaged in a heartfelt conversation inside a car. The man, sporting a denim jacket over a blue shirt, sits attentively with a seatbelt fastened, his gaze fixed on the woman beside him. The woman, wearing a black tank top and a denim jacket draped over her shoulders, smiles warmly, her eyes reflecting genuine interest and connection. The car's interior, with its beige seats and simple design, provides a backdrop that emphasizes their interaction. The scene captures a moment of shared understanding and connection, set against the soft, diffused light of an overcast day. A medium shot from a slightly angled perspective, focusing on their expressions and body language.", "Multi Person(TTS)", None, None, "(s1) do you know multi-talk? (s2) yes, I know it, that's amazing! (s1) Me too."],
+ ],
+ inputs = [img2vid_image, img2vid_prompt, mode_selector, img2vid_audio_1, img2vid_audio_2, tts_text],
+ )
+
+
+ run_i2v_button.click(
+ fn=generate_video,
+ inputs=[img2vid_image, img2vid_prompt, n_prompt, img2vid_audio_1, img2vid_audio_2,sd_steps, seed, text_guide_scale, audio_guide_scale, mode_selector, tts_text, resolution_select, human1_voice, human2_voice],
+ outputs=[result_gallery],
+ )
+ demo.launch(server_name="0.0.0.0", debug=True, server_port=8418)
+
+
+
+
+if __name__ == "__main__":
+ args = _parse_args()
+ run_graio_demo(args)
+
diff --git a/assets/efficiency.png b/assets/efficiency.png
index f974a3c..b844795 100644
Binary files a/assets/efficiency.png and b/assets/efficiency.png differ
diff --git a/assets/none_quant_efficiency.png b/assets/none_quant_efficiency.png
new file mode 100644
index 0000000..2ae490d
Binary files /dev/null and b/assets/none_quant_efficiency.png differ
diff --git a/examples/multitalk_example_tts_1.json b/examples/multitalk_example_tts_1.json
new file mode 100644
index 0000000..ab4ed7f
--- /dev/null
+++ b/examples/multitalk_example_tts_1.json
@@ -0,0 +1,11 @@
+{
+ "prompt": "In a cozy recording studio, a man and a woman are singing together with passion and emotion. The man, with short brown hair, wears a light gray button-up shirt, his expression filled with concentration and warmth. The woman, with long wavy brown hair, dons a sleeveless dress adorned with small polka dots, her eyes closed as she belts out a heartfelt melody. The studio is equipped with professional microphones, and the background features soundproofing panels, creating an intimate and focused atmosphere. A close-up shot captures their expressions and the intensity of their performance.",
+ "cond_image": "examples/multi/1/multi1.png",
+ "audio_type": "para",
+ "tts_audio": {
+ "text": "(s1) do you know multi-talk? (s2) yes, I know it, that's amazing! (s1) Me too.",
+ "human1_voice": "weights/Kokoro-82M/voices/af_heart.pt",
+ "human2_voice": "weights/Kokoro-82M/voices/am_adam.pt"
+ },
+ "cond_audio":{}
+}
diff --git a/examples/single_example_tts_1.json b/examples/single_example_tts_1.json
new file mode 100644
index 0000000..f42287c
--- /dev/null
+++ b/examples/single_example_tts_1.json
@@ -0,0 +1,9 @@
+{
+ "prompt": "A woman is passionately singing into a professional microphone in a recording studio. She wears large black headphones and a dark cardigan over a gray top. Her long, wavy brown hair frames her face as she looks slightly upwards, her mouth open mid-song. The studio is equipped with various audio equipment, including a mixing console and a keyboard, with soundproofing panels on the walls. The lighting is warm and focused on her, creating a professional and intimate atmosphere. A close-up shot captures her expressive performance.",
+ "cond_image": "examples/single/single1.png",
+ "tts_audio": {
+ "text": "Welcome to multi-talk, this is an audio-driven video generation model for multiple person.",
+ "human1_voice": "weights/Kokoro-82M/voices/af_heart.pt"
+ },
+ "cond_audio":{}
+}
diff --git a/generate_multitalk.py b/generate_multitalk.py
index aebcc30..9a3b525 100644
--- a/generate_multitalk.py
+++ b/generate_multitalk.py
@@ -20,7 +20,7 @@
from wan.configs import SIZE_CONFIGS, SUPPORTED_SIZES, WAN_CONFIGS
from wan.utils.utils import cache_image, cache_video, str2bool
from wan.utils.multitalk_utils import save_video_ffmpeg
-
+from kokoro import KPipeline
from transformers import Wav2Vec2FeatureExtractor
from src.audio_analysis.wav2vec2 import Wav2Vec2Model
@@ -29,6 +29,8 @@
import numpy as np
from einops import rearrange
import soundfile as sf
+import re
+
def _validate_args(args):
# Basic check
@@ -83,11 +85,30 @@ def _parse_args():
type=str,
default=None,
help="The path to the Wan checkpoint directory.")
+ parser.add_argument(
+ "--quant_dir",
+ type=str,
+ default=None,
+ help="The path to the Wan quant checkpoint directory.")
parser.add_argument(
"--wav2vec_dir",
type=str,
default=None,
help="The path to the wav2vec checkpoint directory.")
+ parser.add_argument(
+ "--lora_dir",
+ type=str,
+ nargs='+',
+ default=None,
+ help="The paths to the LoRA checkpoint files."
+ )
+ parser.add_argument(
+ "--lora_scale",
+ type=float,
+ nargs='+',
+ default=[1.2],
+ help="Controls how much to influence the outputs with the LoRA parameters. Accepts multiple float values."
+ )
parser.add_argument(
"--offload_model",
type=str2bool,
@@ -174,6 +195,12 @@ def _parse_args():
required=False,
help="Maximum parameter quantity retained in video memory, small number to reduce VRAM required",
)
+ parser.add_argument(
+ "--audio_mode",
+ type=str,
+ default="localfile",
+ choices=['localfile', 'tts'],
+ help="localfile: audio from local wav file, tts: audio from TTS")
parser.add_argument(
"--use_teacache",
action="store_true",
@@ -204,7 +231,19 @@ def _parse_args():
default=55,
help="Norm threshold used in adaptive projected guidance (APG)."
)
+ parser.add_argument(
+ "--color_correction_strength",
+ type=float,
+ default=1.0,
+ help="strength for color correction [0.0 -- 1.0]."
+ )
+ parser.add_argument(
+ "--quant",
+ type=str,
+ default=None,
+ help="Quantization type, must be 'int8' or 'fp8'."
+ )
args = parser.parse_args()
@@ -316,6 +355,78 @@ def audio_prepare_single(audio_path, sample_rate=16000):
human_speech_array = loudness_norm(human_speech_array, sr)
return human_speech_array
+def process_tts_single(text, save_dir, voice1):
+ s1_sentences = []
+
+ pipeline = KPipeline(lang_code='a', repo_id='weights/Kokoro-82M')
+
+ voice_tensor = torch.load(voice1, weights_only=True)
+ generator = pipeline(
+ text, voice=voice_tensor, # <= change voice here
+ speed=1, split_pattern=r'\n+'
+ )
+ audios = []
+ for i, (gs, ps, audio) in enumerate(generator):
+ audios.append(audio)
+ audios = torch.concat(audios, dim=0)
+ s1_sentences.append(audios)
+ s1_sentences = torch.concat(s1_sentences, dim=0)
+ save_path1 =f'{save_dir}/s1.wav'
+ sf.write(save_path1, s1_sentences, 24000) # save each audio file
+ s1, _ = librosa.load(save_path1, sr=16000)
+ return s1, save_path1
+
+
+
+def process_tts_multi(text, save_dir, voice1, voice2):
+ pattern = r'\(s(\d+)\)\s*(.*?)(?=\s*\(s\d+\)|$)'
+ matches = re.findall(pattern, text, re.DOTALL)
+
+ s1_sentences = []
+ s2_sentences = []
+
+ pipeline = KPipeline(lang_code='a', repo_id='weights/Kokoro-82M')
+ for idx, (speaker, content) in enumerate(matches):
+ if speaker == '1':
+ voice_tensor = torch.load(voice1, weights_only=True)
+ generator = pipeline(
+ content, voice=voice_tensor, # <= change voice here
+ speed=1, split_pattern=r'\n+'
+ )
+ audios = []
+ for i, (gs, ps, audio) in enumerate(generator):
+ audios.append(audio)
+ audios = torch.concat(audios, dim=0)
+ s1_sentences.append(audios)
+ s2_sentences.append(torch.zeros_like(audios))
+ elif speaker == '2':
+ voice_tensor = torch.load(voice2, weights_only=True)
+ generator = pipeline(
+ content, voice=voice_tensor, # <= change voice here
+ speed=1, split_pattern=r'\n+'
+ )
+ audios = []
+ for i, (gs, ps, audio) in enumerate(generator):
+ audios.append(audio)
+ audios = torch.concat(audios, dim=0)
+ s2_sentences.append(audios)
+ s1_sentences.append(torch.zeros_like(audios))
+
+ s1_sentences = torch.concat(s1_sentences, dim=0)
+ s2_sentences = torch.concat(s2_sentences, dim=0)
+ sum_sentences = s1_sentences + s2_sentences
+ save_path1 =f'{save_dir}/s1.wav'
+ save_path2 =f'{save_dir}/s2.wav'
+ save_path_sum = f'{save_dir}/sum.wav'
+ sf.write(save_path1, s1_sentences, 24000) # save each audio file
+ sf.write(save_path2, s2_sentences, 24000)
+ sf.write(save_path_sum, sum_sentences, 24000)
+
+ s1, _ = librosa.load(save_path1, sr=16000)
+ s2, _ = librosa.load(save_path2, sr=16000)
+ # sum, _ = librosa.load(save_path_sum, sr=16000)
+ return s1, s2, save_path_sum
+
def generate(args):
rank = int(os.getenv("RANK", 0))
world_size = int(os.getenv("WORLD_SIZE", 1))
@@ -423,41 +534,67 @@ def generate(args):
args.audio_save_dir = os.path.join(args.audio_save_dir, input_data['cond_image'].split('/')[-1].split('.')[0])
os.makedirs(args.audio_save_dir,exist_ok=True)
- if len(input_data['cond_audio'])==2:
- new_human_speech1, new_human_speech2, sum_human_speechs = audio_prepare_multi(input_data['cond_audio']['person1'], input_data['cond_audio']['person2'], input_data['audio_type'])
- audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
- audio_embedding_2 = get_embedding(new_human_speech2, wav2vec_feature_extractor, audio_encoder)
- emb1_path = os.path.join(args.audio_save_dir, '1.pt')
- emb2_path = os.path.join(args.audio_save_dir, '2.pt')
- sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
- sf.write(sum_audio, sum_human_speechs, 16000)
- torch.save(audio_embedding_1, emb1_path)
- torch.save(audio_embedding_2, emb2_path)
- input_data['cond_audio']['person1'] = emb1_path
- input_data['cond_audio']['person2'] = emb2_path
- input_data['video_audio'] = sum_audio
- elif len(input_data['cond_audio'])==1:
- human_speech = audio_prepare_single(input_data['cond_audio']['person1'])
- audio_embedding = get_embedding(human_speech, wav2vec_feature_extractor, audio_encoder)
- emb_path = os.path.join(args.audio_save_dir, '1.pt')
- sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
- sf.write(sum_audio, human_speech, 16000)
- torch.save(audio_embedding, emb_path)
- input_data['cond_audio']['person1'] = emb_path
- input_data['video_audio'] = sum_audio
+ if args.audio_mode=='localfile':
+ if len(input_data['cond_audio'])==2:
+ new_human_speech1, new_human_speech2, sum_human_speechs = audio_prepare_multi(input_data['cond_audio']['person1'], input_data['cond_audio']['person2'], input_data['audio_type'])
+ audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ audio_embedding_2 = get_embedding(new_human_speech2, wav2vec_feature_extractor, audio_encoder)
+ emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ emb2_path = os.path.join(args.audio_save_dir, '2.pt')
+ sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
+ sf.write(sum_audio, sum_human_speechs, 16000)
+ torch.save(audio_embedding_1, emb1_path)
+ torch.save(audio_embedding_2, emb2_path)
+ input_data['cond_audio']['person1'] = emb1_path
+ input_data['cond_audio']['person2'] = emb2_path
+ input_data['video_audio'] = sum_audio
+ elif len(input_data['cond_audio'])==1:
+ human_speech = audio_prepare_single(input_data['cond_audio']['person1'])
+ audio_embedding = get_embedding(human_speech, wav2vec_feature_extractor, audio_encoder)
+ emb_path = os.path.join(args.audio_save_dir, '1.pt')
+ sum_audio = os.path.join(args.audio_save_dir, 'sum.wav')
+ sf.write(sum_audio, human_speech, 16000)
+ torch.save(audio_embedding, emb_path)
+ input_data['cond_audio']['person1'] = emb_path
+ input_data['video_audio'] = sum_audio
+ elif args.audio_mode=='tts':
+ if 'human2_voice' not in input_data['tts_audio'].keys():
+ new_human_speech1, sum_audio = process_tts_single(input_data['tts_audio']['text'], args.audio_save_dir, input_data['tts_audio']['human1_voice'])
+ audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ torch.save(audio_embedding_1, emb1_path)
+ input_data['cond_audio']['person1'] = emb1_path
+ input_data['video_audio'] = sum_audio
+ else:
+ new_human_speech1, new_human_speech2, sum_audio = process_tts_multi(input_data['tts_audio']['text'], args.audio_save_dir, input_data['tts_audio']['human1_voice'], input_data['tts_audio']['human2_voice'])
+ audio_embedding_1 = get_embedding(new_human_speech1, wav2vec_feature_extractor, audio_encoder)
+ audio_embedding_2 = get_embedding(new_human_speech2, wav2vec_feature_extractor, audio_encoder)
+ emb1_path = os.path.join(args.audio_save_dir, '1.pt')
+ emb2_path = os.path.join(args.audio_save_dir, '2.pt')
+ torch.save(audio_embedding_1, emb1_path)
+ torch.save(audio_embedding_2, emb2_path)
+ input_data['cond_audio']['person1'] = emb1_path
+ input_data['cond_audio']['person2'] = emb2_path
+ input_data['video_audio'] = sum_audio
+
logging.info("Creating MultiTalk pipeline.")
wan_i2v = wan.MultiTalkPipeline(
config=cfg,
checkpoint_dir=args.ckpt_dir,
+ quant_dir=args.quant_dir,
device_id=device,
rank=rank,
t5_fsdp=args.t5_fsdp,
dit_fsdp=args.dit_fsdp,
use_usp=(args.ulysses_size > 1 or args.ring_size > 1),
- t5_cpu=args.t5_cpu
+ t5_cpu=args.t5_cpu,
+ lora_dir=args.lora_dir,
+ lora_scales=args.lora_scale,
+ quant=args.quant
)
+
if args.num_persistent_param_in_dit is not None:
wan_i2v.vram_management = True
wan_i2v.enable_vram_management(
@@ -477,6 +614,7 @@ def generate(args):
seed=args.base_seed,
offload_model=args.offload_model,
max_frames_num=args.frame_num if args.mode == 'clip' else 1000,
+ color_correction_strength = args.color_correction_strength,
extra_args=args,
)
@@ -490,7 +628,7 @@ def generate(args):
args.save_file = f"{args.task}_{args.size.replace('*','x') if sys.platform=='win32' else args.size}_{args.ulysses_size}_{args.ring_size}_{formatted_prompt}_{formatted_time}"
logging.info(f"Saving generated video to {args.save_file}.mp4")
- save_video_ffmpeg(video, args.save_file, [input_data['video_audio']])
+ save_video_ffmpeg(video, args.save_file, [input_data['video_audio']], high_quality_save=False)
logging.info("Finished.")
diff --git a/kokoro/__init__.py b/kokoro/__init__.py
new file mode 100644
index 0000000..9156e5c
--- /dev/null
+++ b/kokoro/__init__.py
@@ -0,0 +1,23 @@
+__version__ = '0.9.4'
+
+from loguru import logger
+import sys
+
+# Remove default handler
+logger.remove()
+
+# Add custom handler with clean format including module and line number
+logger.add(
+ sys.stderr,
+ format="{time:HH:mm:ss} | {module:>16}:{line} | {level: >8} | {message}",
+ colorize=True,
+ level="INFO" # "DEBUG" to enable logger.debug("message") and up prints
+ # "ERROR" to enable only logger.error("message") prints
+ # etc
+)
+
+# Disable before release or as needed
+logger.disable("kokoro")
+
+from .model import KModel
+from .pipeline import KPipeline
diff --git a/kokoro/__main__.py b/kokoro/__main__.py
new file mode 100644
index 0000000..34ee21a
--- /dev/null
+++ b/kokoro/__main__.py
@@ -0,0 +1,148 @@
+"""Kokoro TTS CLI
+Example usage:
+python3 -m kokoro --text "The sky above the port was the color of television, tuned to a dead channel." -o file.wav --debug
+
+echo "Bom dia mundo, como vão vocês" > text.txt
+python3 -m kokoro -i text.txt -l p --voice pm_alex > audio.wav
+
+Common issues:
+pip not installed: `uv pip install pip`
+(Temporary workaround while https://github.com/explosion/spaCy/issues/13747 is not fixed)
+
+espeak not installed: `apt-get install espeak-ng`
+"""
+
+import argparse
+import wave
+from pathlib import Path
+from typing import Generator, TYPE_CHECKING
+
+import numpy as np
+from loguru import logger
+
+languages = [
+ "a", # American English
+ "b", # British English
+ "h", # Hindi
+ "e", # Spanish
+ "f", # French
+ "i", # Italian
+ "p", # Brazilian Portuguese
+ "j", # Japanese
+ "z", # Mandarin Chinese
+]
+
+if TYPE_CHECKING:
+ from kokoro import KPipeline
+
+
+def generate_audio(
+ text: str, kokoro_language: str, voice: str, speed=1
+) -> Generator["KPipeline.Result", None, None]:
+ from kokoro import KPipeline
+
+ if not voice.startswith(kokoro_language):
+ logger.warning(f"Voice {voice} is not made for language {kokoro_language}")
+ pipeline = KPipeline(lang_code=kokoro_language)
+ yield from pipeline(text, voice=voice, speed=speed, split_pattern=r"\n+")
+
+
+def generate_and_save_audio(
+ output_file: Path, text: str, kokoro_language: str, voice: str, speed=1
+) -> None:
+ with wave.open(str(output_file.resolve()), "wb") as wav_file:
+ wav_file.setnchannels(1) # Mono audio
+ wav_file.setsampwidth(2) # 2 bytes per sample (16-bit audio)
+ wav_file.setframerate(24000) # Sample rate
+
+ for result in generate_audio(
+ text, kokoro_language=kokoro_language, voice=voice, speed=speed
+ ):
+ logger.debug(result.phonemes)
+ if result.audio is None:
+ continue
+ audio_bytes = (result.audio.numpy() * 32767).astype(np.int16).tobytes()
+ wav_file.writeframes(audio_bytes)
+
+
+def main() -> None:
+ parser = argparse.ArgumentParser()
+ parser.add_argument(
+ "-m",
+ "--voice",
+ default="af_heart",
+ help="Voice to use",
+ )
+ parser.add_argument(
+ "-l",
+ "--language",
+ help="Language to use (defaults to the one corresponding to the voice)",
+ choices=languages,
+ )
+ parser.add_argument(
+ "-o",
+ "--output-file",
+ "--output_file",
+ type=Path,
+ help="Path to output WAV file",
+ required=True,
+ )
+ parser.add_argument(
+ "-i",
+ "--input-file",
+ "--input_file",
+ type=Path,
+ help="Path to input text file (default: stdin)",
+ )
+ parser.add_argument(
+ "-t",
+ "--text",
+ help="Text to use instead of reading from stdin",
+ )
+ parser.add_argument(
+ "-s",
+ "--speed",
+ type=float,
+ default=1.0,
+ help="Speech speed",
+ )
+ parser.add_argument(
+ "--debug",
+ action="store_true",
+ help="Print DEBUG messages to console",
+ )
+ args = parser.parse_args()
+ if args.debug:
+ logger.level("DEBUG")
+ logger.debug(args)
+
+ lang = args.language or args.voice[0]
+
+ if args.text is not None and args.input_file is not None:
+ raise Exception("You cannot specify both 'text' and 'input_file'")
+ elif args.text:
+ text = args.text
+ elif args.input_file:
+ file: Path = args.input_file
+ text = file.read_text()
+ else:
+ import sys
+ print("Press Ctrl+D to stop reading input and start generating", flush=True)
+ text = '\n'.join(sys.stdin)
+
+ logger.debug(f"Input text: {text!r}")
+
+ out_file: Path = args.output_file
+ if not out_file.suffix == ".wav":
+ logger.warning("The output file name should end with .wav")
+ generate_and_save_audio(
+ output_file=out_file,
+ text=text,
+ kokoro_language=lang,
+ voice=args.voice,
+ speed=args.speed,
+ )
+
+
+if __name__ == "__main__":
+ main()
diff --git a/kokoro/custom_stft.py b/kokoro/custom_stft.py
new file mode 100644
index 0000000..c9cf0d2
--- /dev/null
+++ b/kokoro/custom_stft.py
@@ -0,0 +1,197 @@
+from attr import attr
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class CustomSTFT(nn.Module):
+ """
+ STFT/iSTFT without unfold/complex ops, using conv1d and conv_transpose1d.
+
+ - forward STFT => Real-part conv1d + Imag-part conv1d
+ - inverse STFT => Real-part conv_transpose1d + Imag-part conv_transpose1d + sum
+ - avoids F.unfold, so easier to export to ONNX
+ - uses replicate or constant padding for 'center=True' to approximate 'reflect'
+ (reflect is not supported for dynamic shapes in ONNX)
+ """
+
+ def __init__(
+ self,
+ filter_length=800,
+ hop_length=200,
+ win_length=800,
+ window="hann",
+ center=True,
+ pad_mode="replicate", # or 'constant'
+ ):
+ super().__init__()
+ self.filter_length = filter_length
+ self.hop_length = hop_length
+ self.win_length = win_length
+ self.n_fft = filter_length
+ self.center = center
+ self.pad_mode = pad_mode
+
+ # Number of frequency bins for real-valued STFT with onesided=True
+ self.freq_bins = self.n_fft // 2 + 1
+
+ # Build window
+ assert window == 'hann', window
+ window_tensor = torch.hann_window(win_length, periodic=True, dtype=torch.float32)
+ if self.win_length < self.n_fft:
+ # Zero-pad up to n_fft
+ extra = self.n_fft - self.win_length
+ window_tensor = F.pad(window_tensor, (0, extra))
+ elif self.win_length > self.n_fft:
+ window_tensor = window_tensor[: self.n_fft]
+ self.register_buffer("window", window_tensor)
+
+ # Precompute forward DFT (real, imag)
+ # PyTorch stft uses e^{-j 2 pi k n / N} => real=cos(...), imag=-sin(...)
+ n = np.arange(self.n_fft)
+ k = np.arange(self.freq_bins)
+ angle = 2 * np.pi * np.outer(k, n) / self.n_fft # shape (freq_bins, n_fft)
+ dft_real = np.cos(angle)
+ dft_imag = -np.sin(angle) # note negative sign
+
+ # Combine window and dft => shape (freq_bins, filter_length)
+ # We'll make 2 conv weight tensors of shape (freq_bins, 1, filter_length).
+ forward_window = window_tensor.numpy() # shape (n_fft,)
+ forward_real = dft_real * forward_window # (freq_bins, n_fft)
+ forward_imag = dft_imag * forward_window
+
+ # Convert to PyTorch
+ forward_real_torch = torch.from_numpy(forward_real).float()
+ forward_imag_torch = torch.from_numpy(forward_imag).float()
+
+ # Register as Conv1d weight => (out_channels, in_channels, kernel_size)
+ # out_channels = freq_bins, in_channels=1, kernel_size=n_fft
+ self.register_buffer(
+ "weight_forward_real", forward_real_torch.unsqueeze(1)
+ )
+ self.register_buffer(
+ "weight_forward_imag", forward_imag_torch.unsqueeze(1)
+ )
+
+ # Precompute inverse DFT
+ # Real iFFT formula => scale = 1/n_fft, doubling for bins 1..freq_bins-2 if n_fft even, etc.
+ # For simplicity, we won't do the "DC/nyquist not doubled" approach here.
+ # If you want perfect real iSTFT, you can add that logic.
+ # This version just yields good approximate reconstruction with Hann + typical overlap.
+ inv_scale = 1.0 / self.n_fft
+ n = np.arange(self.n_fft)
+ angle_t = 2 * np.pi * np.outer(n, k) / self.n_fft # shape (n_fft, freq_bins)
+ idft_cos = np.cos(angle_t).T # => (freq_bins, n_fft)
+ idft_sin = np.sin(angle_t).T # => (freq_bins, n_fft)
+
+ # Multiply by window again for typical overlap-add
+ # We also incorporate the scale factor 1/n_fft
+ inv_window = window_tensor.numpy() * inv_scale
+ backward_real = idft_cos * inv_window # (freq_bins, n_fft)
+ backward_imag = idft_sin * inv_window
+
+ # We'll implement iSTFT as real+imag conv_transpose with stride=hop.
+ self.register_buffer(
+ "weight_backward_real", torch.from_numpy(backward_real).float().unsqueeze(1)
+ )
+ self.register_buffer(
+ "weight_backward_imag", torch.from_numpy(backward_imag).float().unsqueeze(1)
+ )
+
+
+
+ def transform(self, waveform: torch.Tensor):
+ """
+ Forward STFT => returns magnitude, phase
+ Output shape => (batch, freq_bins, frames)
+ """
+ # waveform shape => (B, T). conv1d expects (B, 1, T).
+ # Optional center pad
+ if self.center:
+ pad_len = self.n_fft // 2
+ waveform = F.pad(waveform, (pad_len, pad_len), mode=self.pad_mode)
+
+ x = waveform.unsqueeze(1) # => (B, 1, T)
+ # Convolution to get real part => shape (B, freq_bins, frames)
+ real_out = F.conv1d(
+ x,
+ self.weight_forward_real,
+ bias=None,
+ stride=self.hop_length,
+ padding=0,
+ )
+ # Imag part
+ imag_out = F.conv1d(
+ x,
+ self.weight_forward_imag,
+ bias=None,
+ stride=self.hop_length,
+ padding=0,
+ )
+
+ # magnitude, phase
+ magnitude = torch.sqrt(real_out**2 + imag_out**2 + 1e-14)
+ phase = torch.atan2(imag_out, real_out)
+ # Handle the case where imag_out is 0 and real_out is negative to correct ONNX atan2 to match PyTorch
+ # In this case, PyTorch returns pi, ONNX returns -pi
+ correction_mask = (imag_out == 0) & (real_out < 0)
+ phase[correction_mask] = torch.pi
+ return magnitude, phase
+
+
+ def inverse(self, magnitude: torch.Tensor, phase: torch.Tensor, length=None):
+ """
+ Inverse STFT => returns waveform shape (B, T).
+ """
+ # magnitude, phase => (B, freq_bins, frames)
+ # Re-create real/imag => shape (B, freq_bins, frames)
+ real_part = magnitude * torch.cos(phase)
+ imag_part = magnitude * torch.sin(phase)
+
+ # conv_transpose wants shape (B, freq_bins, frames). We'll treat "frames" as time dimension
+ # so we do (B, freq_bins, frames) => (B, freq_bins, frames)
+ # But PyTorch conv_transpose1d expects (B, in_channels, input_length)
+ real_part = real_part # (B, freq_bins, frames)
+ imag_part = imag_part
+
+ # real iSTFT => convolve with "backward_real", "backward_imag", and sum
+ # We'll do 2 conv_transpose calls, each giving (B, 1, time),
+ # then add them => (B, 1, time).
+ real_rec = F.conv_transpose1d(
+ real_part,
+ self.weight_backward_real, # shape (freq_bins, 1, filter_length)
+ bias=None,
+ stride=self.hop_length,
+ padding=0,
+ )
+ imag_rec = F.conv_transpose1d(
+ imag_part,
+ self.weight_backward_imag,
+ bias=None,
+ stride=self.hop_length,
+ padding=0,
+ )
+ # sum => (B, 1, time)
+ waveform = real_rec - imag_rec # typical real iFFT has minus for imaginary part
+
+ # If we used "center=True" in forward, we should remove pad
+ if self.center:
+ pad_len = self.n_fft // 2
+ # Because of transposed convolution, total length might have extra samples
+ # We remove `pad_len` from start & end if possible
+ waveform = waveform[..., pad_len:-pad_len]
+
+ # If a specific length is desired, clamp
+ if length is not None:
+ waveform = waveform[..., :length]
+
+ # shape => (B, T)
+ return waveform
+
+ def forward(self, x: torch.Tensor):
+ """
+ Full STFT -> iSTFT pass: returns time-domain reconstruction.
+ Same interface as your original code.
+ """
+ mag, phase = self.transform(x)
+ return self.inverse(mag, phase, length=x.shape[-1])
diff --git a/kokoro/istftnet.py b/kokoro/istftnet.py
new file mode 100644
index 0000000..cb22279
--- /dev/null
+++ b/kokoro/istftnet.py
@@ -0,0 +1,421 @@
+# ADAPTED from https://github.com/yl4579/StyleTTS2/blob/main/Modules/istftnet.py
+from kokoro.custom_stft import CustomSTFT
+from torch.nn.utils import weight_norm
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+# https://github.com/yl4579/StyleTTS2/blob/main/Modules/utils.py
+def init_weights(m, mean=0.0, std=0.01):
+ classname = m.__class__.__name__
+ if classname.find("Conv") != -1:
+ m.weight.data.normal_(mean, std)
+
+def get_padding(kernel_size, dilation=1):
+ return int((kernel_size*dilation - dilation)/2)
+
+
+class AdaIN1d(nn.Module):
+ def __init__(self, style_dim, num_features):
+ super().__init__()
+ # affine should be False, however there's a bug in the old torch.onnx.export (not newer dynamo) that causes the channel dimension to be lost if affine=False. When affine is true, there's additional learnably parameters. This shouldn't really matter setting it to True, since we're in inference mode
+ self.norm = nn.InstanceNorm1d(num_features, affine=True)
+ self.fc = nn.Linear(style_dim, num_features*2)
+
+ def forward(self, x, s):
+ h = self.fc(s)
+ h = h.view(h.size(0), h.size(1), 1)
+ gamma, beta = torch.chunk(h, chunks=2, dim=1)
+ return (1 + gamma) * self.norm(x) + beta
+
+
+class AdaINResBlock1(nn.Module):
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5), style_dim=64):
+ super(AdaINResBlock1, self).__init__()
+ self.convs1 = nn.ModuleList([
+ weight_norm(nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+ padding=get_padding(kernel_size, dilation[0]))),
+ weight_norm(nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+ padding=get_padding(kernel_size, dilation[1]))),
+ weight_norm(nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
+ padding=get_padding(kernel_size, dilation[2])))
+ ])
+ self.convs1.apply(init_weights)
+ self.convs2 = nn.ModuleList([
+ weight_norm(nn.Conv1d(channels, channels, kernel_size, 1, dilation=1,
+ padding=get_padding(kernel_size, 1))),
+ weight_norm(nn.Conv1d(channels, channels, kernel_size, 1, dilation=1,
+ padding=get_padding(kernel_size, 1))),
+ weight_norm(nn.Conv1d(channels, channels, kernel_size, 1, dilation=1,
+ padding=get_padding(kernel_size, 1)))
+ ])
+ self.convs2.apply(init_weights)
+ self.adain1 = nn.ModuleList([
+ AdaIN1d(style_dim, channels),
+ AdaIN1d(style_dim, channels),
+ AdaIN1d(style_dim, channels),
+ ])
+ self.adain2 = nn.ModuleList([
+ AdaIN1d(style_dim, channels),
+ AdaIN1d(style_dim, channels),
+ AdaIN1d(style_dim, channels),
+ ])
+ self.alpha1 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs1))])
+ self.alpha2 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs2))])
+
+ def forward(self, x, s):
+ for c1, c2, n1, n2, a1, a2 in zip(self.convs1, self.convs2, self.adain1, self.adain2, self.alpha1, self.alpha2):
+ xt = n1(x, s)
+ xt = xt + (1 / a1) * (torch.sin(a1 * xt) ** 2) # Snake1D
+ xt = c1(xt)
+ xt = n2(xt, s)
+ xt = xt + (1 / a2) * (torch.sin(a2 * xt) ** 2) # Snake1D
+ xt = c2(xt)
+ x = xt + x
+ return x
+
+
+class TorchSTFT(nn.Module):
+ def __init__(self, filter_length=800, hop_length=200, win_length=800, window='hann'):
+ super().__init__()
+ self.filter_length = filter_length
+ self.hop_length = hop_length
+ self.win_length = win_length
+ assert window == 'hann', window
+ self.window = torch.hann_window(win_length, periodic=True, dtype=torch.float32)
+
+ def transform(self, input_data):
+ forward_transform = torch.stft(
+ input_data,
+ self.filter_length, self.hop_length, self.win_length, window=self.window.to(input_data.device),
+ return_complex=True)
+ return torch.abs(forward_transform), torch.angle(forward_transform)
+
+ def inverse(self, magnitude, phase):
+ inverse_transform = torch.istft(
+ magnitude * torch.exp(phase * 1j),
+ self.filter_length, self.hop_length, self.win_length, window=self.window.to(magnitude.device))
+ return inverse_transform.unsqueeze(-2) # unsqueeze to stay consistent with conv_transpose1d implementation
+
+ def forward(self, input_data):
+ self.magnitude, self.phase = self.transform(input_data)
+ reconstruction = self.inverse(self.magnitude, self.phase)
+ return reconstruction
+
+
+class SineGen(nn.Module):
+ """ Definition of sine generator
+ SineGen(samp_rate, harmonic_num = 0,
+ sine_amp = 0.1, noise_std = 0.003,
+ voiced_threshold = 0,
+ flag_for_pulse=False)
+ samp_rate: sampling rate in Hz
+ harmonic_num: number of harmonic overtones (default 0)
+ sine_amp: amplitude of sine-wavefrom (default 0.1)
+ noise_std: std of Gaussian noise (default 0.003)
+ voiced_thoreshold: F0 threshold for U/V classification (default 0)
+ flag_for_pulse: this SinGen is used inside PulseGen (default False)
+ Note: when flag_for_pulse is True, the first time step of a voiced
+ segment is always sin(torch.pi) or cos(0)
+ """
+ def __init__(self, samp_rate, upsample_scale, harmonic_num=0,
+ sine_amp=0.1, noise_std=0.003,
+ voiced_threshold=0,
+ flag_for_pulse=False):
+ super(SineGen, self).__init__()
+ self.sine_amp = sine_amp
+ self.noise_std = noise_std
+ self.harmonic_num = harmonic_num
+ self.dim = self.harmonic_num + 1
+ self.sampling_rate = samp_rate
+ self.voiced_threshold = voiced_threshold
+ self.flag_for_pulse = flag_for_pulse
+ self.upsample_scale = upsample_scale
+
+ def _f02uv(self, f0):
+ # generate uv signal
+ uv = (f0 > self.voiced_threshold).type(torch.float32)
+ return uv
+
+ def _f02sine(self, f0_values):
+ """ f0_values: (batchsize, length, dim)
+ where dim indicates fundamental tone and overtones
+ """
+ # convert to F0 in rad. The interger part n can be ignored
+ # because 2 * torch.pi * n doesn't affect phase
+ rad_values = (f0_values / self.sampling_rate) % 1
+ # initial phase noise (no noise for fundamental component)
+ rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], device=f0_values.device)
+ rand_ini[:, 0] = 0
+ rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+ # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
+ if not self.flag_for_pulse:
+ rad_values = F.interpolate(rad_values.transpose(1, 2), scale_factor=1/self.upsample_scale, mode="linear").transpose(1, 2)
+ phase = torch.cumsum(rad_values, dim=1) * 2 * torch.pi
+ phase = F.interpolate(phase.transpose(1, 2) * self.upsample_scale, scale_factor=self.upsample_scale, mode="linear").transpose(1, 2)
+ sines = torch.sin(phase)
+ else:
+ # If necessary, make sure that the first time step of every
+ # voiced segments is sin(pi) or cos(0)
+ # This is used for pulse-train generation
+ # identify the last time step in unvoiced segments
+ uv = self._f02uv(f0_values)
+ uv_1 = torch.roll(uv, shifts=-1, dims=1)
+ uv_1[:, -1, :] = 1
+ u_loc = (uv < 1) * (uv_1 > 0)
+ # get the instantanouse phase
+ tmp_cumsum = torch.cumsum(rad_values, dim=1)
+ # different batch needs to be processed differently
+ for idx in range(f0_values.shape[0]):
+ temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
+ temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
+ # stores the accumulation of i.phase within
+ # each voiced segments
+ tmp_cumsum[idx, :, :] = 0
+ tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum
+ # rad_values - tmp_cumsum: remove the accumulation of i.phase
+ # within the previous voiced segment.
+ i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)
+ # get the sines
+ sines = torch.cos(i_phase * 2 * torch.pi)
+ return sines
+
+ def forward(self, f0):
+ """ sine_tensor, uv = forward(f0)
+ input F0: tensor(batchsize=1, length, dim=1)
+ f0 for unvoiced steps should be 0
+ output sine_tensor: tensor(batchsize=1, length, dim)
+ output uv: tensor(batchsize=1, length, 1)
+ """
+ f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+ # fundamental component
+ fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))
+ # generate sine waveforms
+ sine_waves = self._f02sine(fn) * self.sine_amp
+ # generate uv signal
+ # uv = torch.ones(f0.shape)
+ # uv = uv * (f0 > self.voiced_threshold)
+ uv = self._f02uv(f0)
+ # noise: for unvoiced should be similar to sine_amp
+ # std = self.sine_amp/3 -> max value ~ self.sine_amp
+ # for voiced regions is self.noise_std
+ noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+ noise = noise_amp * torch.randn_like(sine_waves)
+ # first: set the unvoiced part to 0 by uv
+ # then: additive noise
+ sine_waves = sine_waves * uv + noise
+ return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(nn.Module):
+ """ SourceModule for hn-nsf
+ SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+ add_noise_std=0.003, voiced_threshod=0)
+ sampling_rate: sampling_rate in Hz
+ harmonic_num: number of harmonic above F0 (default: 0)
+ sine_amp: amplitude of sine source signal (default: 0.1)
+ add_noise_std: std of additive Gaussian noise (default: 0.003)
+ note that amplitude of noise in unvoiced is decided
+ by sine_amp
+ voiced_threshold: threhold to set U/V given F0 (default: 0)
+ Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+ F0_sampled (batchsize, length, 1)
+ Sine_source (batchsize, length, 1)
+ noise_source (batchsize, length 1)
+ uv (batchsize, length, 1)
+ """
+ def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
+ add_noise_std=0.003, voiced_threshod=0):
+ super(SourceModuleHnNSF, self).__init__()
+ self.sine_amp = sine_amp
+ self.noise_std = add_noise_std
+ # to produce sine waveforms
+ self.l_sin_gen = SineGen(sampling_rate, upsample_scale, harmonic_num,
+ sine_amp, add_noise_std, voiced_threshod)
+ # to merge source harmonics into a single excitation
+ self.l_linear = nn.Linear(harmonic_num + 1, 1)
+ self.l_tanh = nn.Tanh()
+
+ def forward(self, x):
+ """
+ Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+ F0_sampled (batchsize, length, 1)
+ Sine_source (batchsize, length, 1)
+ noise_source (batchsize, length 1)
+ """
+ # source for harmonic branch
+ with torch.no_grad():
+ sine_wavs, uv, _ = self.l_sin_gen(x)
+ sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+ # source for noise branch, in the same shape as uv
+ noise = torch.randn_like(uv) * self.sine_amp / 3
+ return sine_merge, noise, uv
+
+
+class Generator(nn.Module):
+ def __init__(self, style_dim, resblock_kernel_sizes, upsample_rates, upsample_initial_channel, resblock_dilation_sizes, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, disable_complex=False):
+ super(Generator, self).__init__()
+ self.num_kernels = len(resblock_kernel_sizes)
+ self.num_upsamples = len(upsample_rates)
+ self.m_source = SourceModuleHnNSF(
+ sampling_rate=24000,
+ upsample_scale=math.prod(upsample_rates) * gen_istft_hop_size,
+ harmonic_num=8, voiced_threshod=10)
+ self.f0_upsamp = nn.Upsample(scale_factor=math.prod(upsample_rates) * gen_istft_hop_size)
+ self.noise_convs = nn.ModuleList()
+ self.noise_res = nn.ModuleList()
+ self.ups = nn.ModuleList()
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+ self.ups.append(weight_norm(
+ nn.ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
+ k, u, padding=(k-u)//2)))
+ self.resblocks = nn.ModuleList()
+ for i in range(len(self.ups)):
+ ch = upsample_initial_channel//(2**(i+1))
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes,resblock_dilation_sizes)):
+ self.resblocks.append(AdaINResBlock1(ch, k, d, style_dim))
+ c_cur = upsample_initial_channel // (2 ** (i + 1))
+ if i + 1 < len(upsample_rates):
+ stride_f0 = math.prod(upsample_rates[i + 1:])
+ self.noise_convs.append(nn.Conv1d(
+ gen_istft_n_fft + 2, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2))
+ self.noise_res.append(AdaINResBlock1(c_cur, 7, [1,3,5], style_dim))
+ else:
+ self.noise_convs.append(nn.Conv1d(gen_istft_n_fft + 2, c_cur, kernel_size=1))
+ self.noise_res.append(AdaINResBlock1(c_cur, 11, [1,3,5], style_dim))
+ self.post_n_fft = gen_istft_n_fft
+ self.conv_post = weight_norm(nn.Conv1d(ch, self.post_n_fft + 2, 7, 1, padding=3))
+ self.ups.apply(init_weights)
+ self.conv_post.apply(init_weights)
+ self.reflection_pad = nn.ReflectionPad1d((1, 0))
+ self.stft = (
+ CustomSTFT(filter_length=gen_istft_n_fft, hop_length=gen_istft_hop_size, win_length=gen_istft_n_fft)
+ if disable_complex
+ else TorchSTFT(filter_length=gen_istft_n_fft, hop_length=gen_istft_hop_size, win_length=gen_istft_n_fft)
+ )
+
+ def forward(self, x, s, f0):
+ with torch.no_grad():
+ f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2) # bs,n,t
+ har_source, noi_source, uv = self.m_source(f0)
+ har_source = har_source.transpose(1, 2).squeeze(1)
+ har_spec, har_phase = self.stft.transform(har_source)
+ har = torch.cat([har_spec, har_phase], dim=1)
+ for i in range(self.num_upsamples):
+ x = F.leaky_relu(x, negative_slope=0.1)
+ x_source = self.noise_convs[i](har)
+ x_source = self.noise_res[i](x_source, s)
+ x = self.ups[i](x)
+ if i == self.num_upsamples - 1:
+ x = self.reflection_pad(x)
+ x = x + x_source
+ xs = None
+ for j in range(self.num_kernels):
+ if xs is None:
+ xs = self.resblocks[i*self.num_kernels+j](x, s)
+ else:
+ xs += self.resblocks[i*self.num_kernels+j](x, s)
+ x = xs / self.num_kernels
+ x = F.leaky_relu(x)
+ x = self.conv_post(x)
+ spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
+ phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
+ return self.stft.inverse(spec, phase)
+
+
+class UpSample1d(nn.Module):
+ def __init__(self, layer_type):
+ super().__init__()
+ self.layer_type = layer_type
+
+ def forward(self, x):
+ if self.layer_type == 'none':
+ return x
+ else:
+ return F.interpolate(x, scale_factor=2, mode='nearest')
+
+
+class AdainResBlk1d(nn.Module):
+ def __init__(self, dim_in, dim_out, style_dim=64, actv=nn.LeakyReLU(0.2), upsample='none', dropout_p=0.0):
+ super().__init__()
+ self.actv = actv
+ self.upsample_type = upsample
+ self.upsample = UpSample1d(upsample)
+ self.learned_sc = dim_in != dim_out
+ self._build_weights(dim_in, dim_out, style_dim)
+ self.dropout = nn.Dropout(dropout_p)
+ if upsample == 'none':
+ self.pool = nn.Identity()
+ else:
+ self.pool = weight_norm(nn.ConvTranspose1d(dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1))
+
+ def _build_weights(self, dim_in, dim_out, style_dim):
+ self.conv1 = weight_norm(nn.Conv1d(dim_in, dim_out, 3, 1, 1))
+ self.conv2 = weight_norm(nn.Conv1d(dim_out, dim_out, 3, 1, 1))
+ self.norm1 = AdaIN1d(style_dim, dim_in)
+ self.norm2 = AdaIN1d(style_dim, dim_out)
+ if self.learned_sc:
+ self.conv1x1 = weight_norm(nn.Conv1d(dim_in, dim_out, 1, 1, 0, bias=False))
+
+ def _shortcut(self, x):
+ x = self.upsample(x)
+ if self.learned_sc:
+ x = self.conv1x1(x)
+ return x
+
+ def _residual(self, x, s):
+ x = self.norm1(x, s)
+ x = self.actv(x)
+ x = self.pool(x)
+ x = self.conv1(self.dropout(x))
+ x = self.norm2(x, s)
+ x = self.actv(x)
+ x = self.conv2(self.dropout(x))
+ return x
+
+ def forward(self, x, s):
+ out = self._residual(x, s)
+ out = (out + self._shortcut(x)) * torch.rsqrt(torch.tensor(2))
+ return out
+
+
+class Decoder(nn.Module):
+ def __init__(self, dim_in, style_dim, dim_out,
+ resblock_kernel_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ resblock_dilation_sizes,
+ upsample_kernel_sizes,
+ gen_istft_n_fft, gen_istft_hop_size,
+ disable_complex=False):
+ super().__init__()
+ self.encode = AdainResBlk1d(dim_in + 2, 1024, style_dim)
+ self.decode = nn.ModuleList()
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 512, style_dim, upsample=True))
+ self.F0_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
+ self.N_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
+ self.asr_res = nn.Sequential(weight_norm(nn.Conv1d(512, 64, kernel_size=1)))
+ self.generator = Generator(style_dim, resblock_kernel_sizes, upsample_rates,
+ upsample_initial_channel, resblock_dilation_sizes,
+ upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, disable_complex=disable_complex)
+
+ def forward(self, asr, F0_curve, N, s):
+ F0 = self.F0_conv(F0_curve.unsqueeze(1))
+ N = self.N_conv(N.unsqueeze(1))
+ x = torch.cat([asr, F0, N], axis=1)
+ x = self.encode(x, s)
+ asr_res = self.asr_res(asr)
+ res = True
+ for block in self.decode:
+ if res:
+ x = torch.cat([x, asr_res, F0, N], axis=1)
+ x = block(x, s)
+ if block.upsample_type != "none":
+ res = False
+ x = self.generator(x, s, F0_curve)
+ return x
diff --git a/kokoro/model.py b/kokoro/model.py
new file mode 100644
index 0000000..9d6554c
--- /dev/null
+++ b/kokoro/model.py
@@ -0,0 +1,155 @@
+from .istftnet import Decoder
+from .modules import CustomAlbert, ProsodyPredictor, TextEncoder
+from dataclasses import dataclass
+from huggingface_hub import hf_hub_download
+from loguru import logger
+from transformers import AlbertConfig
+from typing import Dict, Optional, Union
+import json
+import torch
+import os
+
+class KModel(torch.nn.Module):
+ '''
+ KModel is a torch.nn.Module with 2 main responsibilities:
+ 1. Init weights, downloading config.json + model.pth from HF if needed
+ 2. forward(phonemes: str, ref_s: FloatTensor) -> (audio: FloatTensor)
+
+ You likely only need one KModel instance, and it can be reused across
+ multiple KPipelines to avoid redundant memory allocation.
+
+ Unlike KPipeline, KModel is language-blind.
+
+ KModel stores self.vocab and thus knows how to map phonemes -> input_ids,
+ so there is no need to repeatedly download config.json outside of KModel.
+ '''
+
+ MODEL_NAMES = {
+ 'hexgrad/Kokoro-82M': 'kokoro-v1_0.pth',
+ 'hexgrad/Kokoro-82M-v1.1-zh': 'kokoro-v1_1-zh.pth',
+ }
+
+ def __init__(
+ self,
+ repo_id: Optional[str] = None,
+ config: Union[Dict, str, None] = None,
+ model: Optional[str] = None,
+ disable_complex: bool = False
+ ):
+ super().__init__()
+ if repo_id is None:
+ repo_id = 'hexgrad/Kokoro-82M'
+ print(f"WARNING: Defaulting repo_id to {repo_id}. Pass repo_id='{repo_id}' to suppress this warning.")
+ self.repo_id = repo_id
+ if not isinstance(config, dict):
+ if not config:
+ logger.debug("No config provided, downloading from HF")
+ config = hf_hub_download(repo_id=repo_id, filename='config.json')
+ with open(config, 'r', encoding='utf-8') as r:
+ config = json.load(r)
+ logger.debug(f"Loaded config: {config}")
+ self.vocab = config['vocab']
+ self.bert = CustomAlbert(AlbertConfig(vocab_size=config['n_token'], **config['plbert']))
+ self.bert_encoder = torch.nn.Linear(self.bert.config.hidden_size, config['hidden_dim'])
+ self.context_length = self.bert.config.max_position_embeddings
+ self.predictor = ProsodyPredictor(
+ style_dim=config['style_dim'], d_hid=config['hidden_dim'],
+ nlayers=config['n_layer'], max_dur=config['max_dur'], dropout=config['dropout']
+ )
+ self.text_encoder = TextEncoder(
+ channels=config['hidden_dim'], kernel_size=config['text_encoder_kernel_size'],
+ depth=config['n_layer'], n_symbols=config['n_token']
+ )
+ self.decoder = Decoder(
+ dim_in=config['hidden_dim'], style_dim=config['style_dim'],
+ dim_out=config['n_mels'], disable_complex=disable_complex, **config['istftnet']
+ )
+ if not model:
+ try:
+ model = hf_hub_download(repo_id=repo_id, filename=KModel.MODEL_NAMES[repo_id])
+ except:
+ model = os.path.join(repo_id, 'kokoro-v1_0.pth')
+ for key, state_dict in torch.load(model, map_location='cpu', weights_only=True).items():
+ assert hasattr(self, key), key
+ try:
+ getattr(self, key).load_state_dict(state_dict)
+ except:
+ logger.debug(f"Did not load {key} from state_dict")
+ state_dict = {k[7:]: v for k, v in state_dict.items()}
+ getattr(self, key).load_state_dict(state_dict, strict=False)
+
+ @property
+ def device(self):
+ return self.bert.device
+
+ @dataclass
+ class Output:
+ audio: torch.FloatTensor
+ pred_dur: Optional[torch.LongTensor] = None
+
+ @torch.no_grad()
+ def forward_with_tokens(
+ self,
+ input_ids: torch.LongTensor,
+ ref_s: torch.FloatTensor,
+ speed: float = 1
+ ) -> tuple[torch.FloatTensor, torch.LongTensor]:
+ input_lengths = torch.full(
+ (input_ids.shape[0],),
+ input_ids.shape[-1],
+ device=input_ids.device,
+ dtype=torch.long
+ )
+
+ text_mask = torch.arange(input_lengths.max()).unsqueeze(0).expand(input_lengths.shape[0], -1).type_as(input_lengths)
+ text_mask = torch.gt(text_mask+1, input_lengths.unsqueeze(1)).to(self.device)
+ bert_dur = self.bert(input_ids, attention_mask=(~text_mask).int())
+ d_en = self.bert_encoder(bert_dur).transpose(-1, -2)
+ s = ref_s[:, 128:]
+ d = self.predictor.text_encoder(d_en, s, input_lengths, text_mask)
+ x, _ = self.predictor.lstm(d)
+ duration = self.predictor.duration_proj(x)
+ duration = torch.sigmoid(duration).sum(axis=-1) / speed
+ pred_dur = torch.round(duration).clamp(min=1).long().squeeze()
+ indices = torch.repeat_interleave(torch.arange(input_ids.shape[1], device=self.device), pred_dur)
+ pred_aln_trg = torch.zeros((input_ids.shape[1], indices.shape[0]), device=self.device)
+ pred_aln_trg[indices, torch.arange(indices.shape[0])] = 1
+ pred_aln_trg = pred_aln_trg.unsqueeze(0).to(self.device)
+ en = d.transpose(-1, -2) @ pred_aln_trg
+ F0_pred, N_pred = self.predictor.F0Ntrain(en, s)
+ t_en = self.text_encoder(input_ids, input_lengths, text_mask)
+ asr = t_en @ pred_aln_trg
+ audio = self.decoder(asr, F0_pred, N_pred, ref_s[:, :128]).squeeze()
+ return audio, pred_dur
+
+ def forward(
+ self,
+ phonemes: str,
+ ref_s: torch.FloatTensor,
+ speed: float = 1,
+ return_output: bool = False
+ ) -> Union['KModel.Output', torch.FloatTensor]:
+ input_ids = list(filter(lambda i: i is not None, map(lambda p: self.vocab.get(p), phonemes)))
+ logger.debug(f"phonemes: {phonemes} -> input_ids: {input_ids}")
+ assert len(input_ids)+2 <= self.context_length, (len(input_ids)+2, self.context_length)
+ input_ids = torch.LongTensor([[0, *input_ids, 0]]).to(self.device)
+ ref_s = ref_s.to(self.device)
+ audio, pred_dur = self.forward_with_tokens(input_ids, ref_s, speed)
+ audio = audio.squeeze().cpu()
+ pred_dur = pred_dur.cpu() if pred_dur is not None else None
+ logger.debug(f"pred_dur: {pred_dur}")
+ return self.Output(audio=audio, pred_dur=pred_dur) if return_output else audio
+
+class KModelForONNX(torch.nn.Module):
+ def __init__(self, kmodel: KModel):
+ super().__init__()
+ self.kmodel = kmodel
+
+ def forward(
+ self,
+ input_ids: torch.LongTensor,
+ ref_s: torch.FloatTensor,
+ speed: float = 1
+ ) -> tuple[torch.FloatTensor, torch.LongTensor]:
+ waveform, duration = self.kmodel.forward_with_tokens(input_ids, ref_s, speed)
+ return waveform, duration
diff --git a/kokoro/modules.py b/kokoro/modules.py
new file mode 100644
index 0000000..05d1575
--- /dev/null
+++ b/kokoro/modules.py
@@ -0,0 +1,183 @@
+# https://github.com/yl4579/StyleTTS2/blob/main/models.py
+from .istftnet import AdainResBlk1d
+from torch.nn.utils import weight_norm
+from transformers import AlbertModel
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LinearNorm(nn.Module):
+ def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
+ super(LinearNorm, self).__init__()
+ self.linear_layer = nn.Linear(in_dim, out_dim, bias=bias)
+ nn.init.xavier_uniform_(self.linear_layer.weight, gain=nn.init.calculate_gain(w_init_gain))
+
+ def forward(self, x):
+ return self.linear_layer(x)
+
+
+class LayerNorm(nn.Module):
+ def __init__(self, channels, eps=1e-5):
+ super().__init__()
+ self.channels = channels
+ self.eps = eps
+ self.gamma = nn.Parameter(torch.ones(channels))
+ self.beta = nn.Parameter(torch.zeros(channels))
+
+ def forward(self, x):
+ x = x.transpose(1, -1)
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+ return x.transpose(1, -1)
+
+
+class TextEncoder(nn.Module):
+ def __init__(self, channels, kernel_size, depth, n_symbols, actv=nn.LeakyReLU(0.2)):
+ super().__init__()
+ self.embedding = nn.Embedding(n_symbols, channels)
+ padding = (kernel_size - 1) // 2
+ self.cnn = nn.ModuleList()
+ for _ in range(depth):
+ self.cnn.append(nn.Sequential(
+ weight_norm(nn.Conv1d(channels, channels, kernel_size=kernel_size, padding=padding)),
+ LayerNorm(channels),
+ actv,
+ nn.Dropout(0.2),
+ ))
+ self.lstm = nn.LSTM(channels, channels//2, 1, batch_first=True, bidirectional=True)
+
+ def forward(self, x, input_lengths, m):
+ x = self.embedding(x) # [B, T, emb]
+ x = x.transpose(1, 2) # [B, emb, T]
+ m = m.unsqueeze(1)
+ x.masked_fill_(m, 0.0)
+ for c in self.cnn:
+ x = c(x)
+ x.masked_fill_(m, 0.0)
+ x = x.transpose(1, 2) # [B, T, chn]
+ lengths = input_lengths if input_lengths.device == torch.device('cpu') else input_lengths.to('cpu')
+ x = nn.utils.rnn.pack_padded_sequence(x, lengths, batch_first=True, enforce_sorted=False)
+ self.lstm.flatten_parameters()
+ x, _ = self.lstm(x)
+ x, _ = nn.utils.rnn.pad_packed_sequence(x, batch_first=True)
+ x = x.transpose(-1, -2)
+ x_pad = torch.zeros([x.shape[0], x.shape[1], m.shape[-1]], device=x.device)
+ x_pad[:, :, :x.shape[-1]] = x
+ x = x_pad
+ x.masked_fill_(m, 0.0)
+ return x
+
+
+class AdaLayerNorm(nn.Module):
+ def __init__(self, style_dim, channels, eps=1e-5):
+ super().__init__()
+ self.channels = channels
+ self.eps = eps
+ self.fc = nn.Linear(style_dim, channels*2)
+
+ def forward(self, x, s):
+ x = x.transpose(-1, -2)
+ x = x.transpose(1, -1)
+ h = self.fc(s)
+ h = h.view(h.size(0), h.size(1), 1)
+ gamma, beta = torch.chunk(h, chunks=2, dim=1)
+ gamma, beta = gamma.transpose(1, -1), beta.transpose(1, -1)
+ x = F.layer_norm(x, (self.channels,), eps=self.eps)
+ x = (1 + gamma) * x + beta
+ return x.transpose(1, -1).transpose(-1, -2)
+
+
+class ProsodyPredictor(nn.Module):
+ def __init__(self, style_dim, d_hid, nlayers, max_dur=50, dropout=0.1):
+ super().__init__()
+ self.text_encoder = DurationEncoder(sty_dim=style_dim, d_model=d_hid,nlayers=nlayers, dropout=dropout)
+ self.lstm = nn.LSTM(d_hid + style_dim, d_hid // 2, 1, batch_first=True, bidirectional=True)
+ self.duration_proj = LinearNorm(d_hid, max_dur)
+ self.shared = nn.LSTM(d_hid + style_dim, d_hid // 2, 1, batch_first=True, bidirectional=True)
+ self.F0 = nn.ModuleList()
+ self.F0.append(AdainResBlk1d(d_hid, d_hid, style_dim, dropout_p=dropout))
+ self.F0.append(AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True, dropout_p=dropout))
+ self.F0.append(AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim, dropout_p=dropout))
+ self.N = nn.ModuleList()
+ self.N.append(AdainResBlk1d(d_hid, d_hid, style_dim, dropout_p=dropout))
+ self.N.append(AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True, dropout_p=dropout))
+ self.N.append(AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim, dropout_p=dropout))
+ self.F0_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
+ self.N_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
+
+ def forward(self, texts, style, text_lengths, alignment, m):
+ d = self.text_encoder(texts, style, text_lengths, m)
+ m = m.unsqueeze(1)
+ lengths = text_lengths if text_lengths.device == torch.device('cpu') else text_lengths.to('cpu')
+ x = nn.utils.rnn.pack_padded_sequence(d, lengths, batch_first=True, enforce_sorted=False)
+ self.lstm.flatten_parameters()
+ x, _ = self.lstm(x)
+ x, _ = nn.utils.rnn.pad_packed_sequence(x, batch_first=True)
+ x_pad = torch.zeros([x.shape[0], m.shape[-1], x.shape[-1]], device=x.device)
+ x_pad[:, :x.shape[1], :] = x
+ x = x_pad
+ duration = self.duration_proj(nn.functional.dropout(x, 0.5, training=False))
+ en = (d.transpose(-1, -2) @ alignment)
+ return duration.squeeze(-1), en
+
+ def F0Ntrain(self, x, s):
+ x, _ = self.shared(x.transpose(-1, -2))
+ F0 = x.transpose(-1, -2)
+ for block in self.F0:
+ F0 = block(F0, s)
+ F0 = self.F0_proj(F0)
+ N = x.transpose(-1, -2)
+ for block in self.N:
+ N = block(N, s)
+ N = self.N_proj(N)
+ return F0.squeeze(1), N.squeeze(1)
+
+
+class DurationEncoder(nn.Module):
+ def __init__(self, sty_dim, d_model, nlayers, dropout=0.1):
+ super().__init__()
+ self.lstms = nn.ModuleList()
+ for _ in range(nlayers):
+ self.lstms.append(nn.LSTM(d_model + sty_dim, d_model // 2, num_layers=1, batch_first=True, bidirectional=True, dropout=dropout))
+ self.lstms.append(AdaLayerNorm(sty_dim, d_model))
+ self.dropout = dropout
+ self.d_model = d_model
+ self.sty_dim = sty_dim
+
+ def forward(self, x, style, text_lengths, m):
+ masks = m
+ x = x.permute(2, 0, 1)
+ s = style.expand(x.shape[0], x.shape[1], -1)
+ x = torch.cat([x, s], axis=-1)
+ x.masked_fill_(masks.unsqueeze(-1).transpose(0, 1), 0.0)
+ x = x.transpose(0, 1)
+ x = x.transpose(-1, -2)
+ for block in self.lstms:
+ if isinstance(block, AdaLayerNorm):
+ x = block(x.transpose(-1, -2), style).transpose(-1, -2)
+ x = torch.cat([x, s.permute(1, 2, 0)], axis=1)
+ x.masked_fill_(masks.unsqueeze(-1).transpose(-1, -2), 0.0)
+ else:
+ lengths = text_lengths if text_lengths.device == torch.device('cpu') else text_lengths.to('cpu')
+ x = x.transpose(-1, -2)
+ x = nn.utils.rnn.pack_padded_sequence(
+ x, lengths, batch_first=True, enforce_sorted=False)
+ block.flatten_parameters()
+ x, _ = block(x)
+ x, _ = nn.utils.rnn.pad_packed_sequence(
+ x, batch_first=True)
+ x = F.dropout(x, p=self.dropout, training=False)
+ x = x.transpose(-1, -2)
+ x_pad = torch.zeros([x.shape[0], x.shape[1], m.shape[-1]], device=x.device)
+ x_pad[:, :, :x.shape[-1]] = x
+ x = x_pad
+
+ return x.transpose(-1, -2)
+
+
+# https://github.com/yl4579/StyleTTS2/blob/main/Utils/PLBERT/util.py
+class CustomAlbert(AlbertModel):
+ def forward(self, *args, **kwargs):
+ outputs = super().forward(*args, **kwargs)
+ return outputs.last_hidden_state
diff --git a/kokoro/pipeline.py b/kokoro/pipeline.py
new file mode 100644
index 0000000..098df8e
--- /dev/null
+++ b/kokoro/pipeline.py
@@ -0,0 +1,445 @@
+from .model import KModel
+from dataclasses import dataclass
+from huggingface_hub import hf_hub_download
+from loguru import logger
+from misaki import en, espeak
+from typing import Callable, Generator, List, Optional, Tuple, Union
+import re
+import torch
+import os
+
+ALIASES = {
+ 'en-us': 'a',
+ 'en-gb': 'b',
+ 'es': 'e',
+ 'fr-fr': 'f',
+ 'hi': 'h',
+ 'it': 'i',
+ 'pt-br': 'p',
+ 'ja': 'j',
+ 'zh': 'z',
+}
+
+LANG_CODES = dict(
+ # pip install misaki[en]
+ a='American English',
+ b='British English',
+
+ # espeak-ng
+ e='es',
+ f='fr-fr',
+ h='hi',
+ i='it',
+ p='pt-br',
+
+ # pip install misaki[ja]
+ j='Japanese',
+
+ # pip install misaki[zh]
+ z='Mandarin Chinese',
+)
+
+class KPipeline:
+ '''
+ KPipeline is a language-aware support class with 2 main responsibilities:
+ 1. Perform language-specific G2P, mapping (and chunking) text -> phonemes
+ 2. Manage and store voices, lazily downloaded from HF if needed
+
+ You are expected to have one KPipeline per language. If you have multiple
+ KPipelines, you should reuse one KModel instance across all of them.
+
+ KPipeline is designed to work with a KModel, but this is not required.
+ There are 2 ways to pass an existing model into a pipeline:
+ 1. On init: us_pipeline = KPipeline(lang_code='a', model=model)
+ 2. On call: us_pipeline(text, voice, model=model)
+
+ By default, KPipeline will automatically initialize its own KModel. To
+ suppress this, construct a "quiet" KPipeline with model=False.
+
+ A "quiet" KPipeline yields (graphemes, phonemes, None) without generating
+ any audio. You can use this to phonemize and chunk your text in advance.
+
+ A "loud" KPipeline _with_ a model yields (graphemes, phonemes, audio).
+ '''
+ def __init__(
+ self,
+ lang_code: str,
+ repo_id: Optional[str] = None,
+ model: Union[KModel, bool] = True,
+ trf: bool = False,
+ en_callable: Optional[Callable[[str], str]] = None,
+ device: Optional[str] = None
+ ):
+ """Initialize a KPipeline.
+
+ Args:
+ lang_code: Language code for G2P processing
+ model: KModel instance, True to create new model, False for no model
+ trf: Whether to use transformer-based G2P
+ device: Override default device selection ('cuda' or 'cpu', or None for auto)
+ If None, will auto-select cuda if available
+ If 'cuda' and not available, will explicitly raise an error
+ """
+ if repo_id is None:
+ repo_id = 'hexgrad/Kokoro-82M'
+ print(f"WARNING: Defaulting repo_id to {repo_id}. Pass repo_id='{repo_id}' to suppress this warning.")
+ config=None
+ else:
+ config = os.path.join(repo_id, 'config.json')
+ self.repo_id = repo_id
+ lang_code = lang_code.lower()
+ lang_code = ALIASES.get(lang_code, lang_code)
+ assert lang_code in LANG_CODES, (lang_code, LANG_CODES)
+ self.lang_code = lang_code
+ self.model = None
+ if isinstance(model, KModel):
+ self.model = model
+ elif model:
+ if device == 'cuda' and not torch.cuda.is_available():
+ raise RuntimeError("CUDA requested but not available")
+ if device == 'mps' and not torch.backends.mps.is_available():
+ raise RuntimeError("MPS requested but not available")
+ if device == 'mps' and os.environ.get('PYTORCH_ENABLE_MPS_FALLBACK') != '1':
+ raise RuntimeError("MPS requested but fallback not enabled")
+ if device is None:
+ if torch.cuda.is_available():
+ device = 'cuda'
+ elif os.environ.get('PYTORCH_ENABLE_MPS_FALLBACK') == '1' and torch.backends.mps.is_available():
+ device = 'mps'
+ else:
+ device = 'cpu'
+ try:
+ self.model = KModel(repo_id=repo_id, config=config).to(device).eval()
+ except RuntimeError as e:
+ if device == 'cuda':
+ raise RuntimeError(f"""Failed to initialize model on CUDA: {e}.
+ Try setting device='cpu' or check CUDA installation.""")
+ raise
+ self.voices = {}
+ if lang_code in 'ab':
+ try:
+ fallback = espeak.EspeakFallback(british=lang_code=='b')
+ except Exception as e:
+ logger.warning("EspeakFallback not Enabled: OOD words will be skipped")
+ logger.warning({str(e)})
+ fallback = None
+ self.g2p = en.G2P(trf=trf, british=lang_code=='b', fallback=fallback, unk='')
+ elif lang_code == 'j':
+ try:
+ from misaki import ja
+ self.g2p = ja.JAG2P()
+ except ImportError:
+ logger.error("You need to `pip install misaki[ja]` to use lang_code='j'")
+ raise
+ elif lang_code == 'z':
+ try:
+ from misaki import zh
+ self.g2p = zh.ZHG2P(
+ version=None if repo_id.endswith('/Kokoro-82M') else '1.1',
+ en_callable=en_callable
+ )
+ except ImportError:
+ logger.error("You need to `pip install misaki[zh]` to use lang_code='z'")
+ raise
+ else:
+ language = LANG_CODES[lang_code]
+ logger.warning(f"Using EspeakG2P(language='{language}'). Chunking logic not yet implemented, so long texts may be truncated unless you split them with '\\n'.")
+ self.g2p = espeak.EspeakG2P(language=language)
+
+ def load_single_voice(self, voice: str):
+ if voice in self.voices:
+ return self.voices[voice]
+ if voice.endswith('.pt'):
+ f = voice
+ else:
+ f = hf_hub_download(repo_id=self.repo_id, filename=f'voices/{voice}.pt')
+ if not voice.startswith(self.lang_code):
+ v = LANG_CODES.get(voice, voice)
+ p = LANG_CODES.get(self.lang_code, self.lang_code)
+ logger.warning(f'Language mismatch, loading {v} voice into {p} pipeline.')
+ pack = torch.load(f, weights_only=True)
+ self.voices[voice] = pack
+ return pack
+
+ """
+ load_voice is a helper function that lazily downloads and loads a voice:
+ Single voice can be requested (e.g. 'af_bella') or multiple voices (e.g. 'af_bella,af_jessica').
+ If multiple voices are requested, they are averaged.
+ Delimiter is optional and defaults to ','.
+ """
+ def load_voice(self, voice: Union[str, torch.FloatTensor], delimiter: str = ",") -> torch.FloatTensor:
+ if isinstance(voice, torch.FloatTensor):
+ return voice
+ if voice in self.voices:
+ return self.voices[voice]
+ logger.debug(f"Loading voice: {voice}")
+ packs = [self.load_single_voice(v) for v in voice.split(delimiter)]
+ if len(packs) == 1:
+ return packs[0]
+ self.voices[voice] = torch.mean(torch.stack(packs), dim=0)
+ return self.voices[voice]
+
+ @staticmethod
+ def tokens_to_ps(tokens: List[en.MToken]) -> str:
+ return ''.join(t.phonemes + (' ' if t.whitespace else '') for t in tokens).strip()
+
+ @staticmethod
+ def waterfall_last(
+ tokens: List[en.MToken],
+ next_count: int,
+ waterfall: List[str] = ['!.?…', ':;', ',—'],
+ bumps: List[str] = [')', '”']
+ ) -> int:
+ for w in waterfall:
+ z = next((i for i, t in reversed(list(enumerate(tokens))) if t.phonemes in set(w)), None)
+ if z is None:
+ continue
+ z += 1
+ if z < len(tokens) and tokens[z].phonemes in bumps:
+ z += 1
+ if next_count - len(KPipeline.tokens_to_ps(tokens[:z])) <= 510:
+ return z
+ return len(tokens)
+
+ @staticmethod
+ def tokens_to_text(tokens: List[en.MToken]) -> str:
+ return ''.join(t.text + t.whitespace for t in tokens).strip()
+
+ def en_tokenize(
+ self,
+ tokens: List[en.MToken]
+ ) -> Generator[Tuple[str, str, List[en.MToken]], None, None]:
+ tks = []
+ pcount = 0
+ for t in tokens:
+ # American English: ɾ => T
+ t.phonemes = '' if t.phonemes is None else t.phonemes#.replace('ɾ', 'T')
+ next_ps = t.phonemes + (' ' if t.whitespace else '')
+ next_pcount = pcount + len(next_ps.rstrip())
+ if next_pcount > 510:
+ z = KPipeline.waterfall_last(tks, next_pcount)
+ text = KPipeline.tokens_to_text(tks[:z])
+ logger.debug(f"Chunking text at {z}: '{text[:30]}{'...' if len(text) > 30 else ''}'")
+ ps = KPipeline.tokens_to_ps(tks[:z])
+ yield text, ps, tks[:z]
+ tks = tks[z:]
+ pcount = len(KPipeline.tokens_to_ps(tks))
+ if not tks:
+ next_ps = next_ps.lstrip()
+ tks.append(t)
+ pcount += len(next_ps)
+ if tks:
+ text = KPipeline.tokens_to_text(tks)
+ ps = KPipeline.tokens_to_ps(tks)
+ yield ''.join(text).strip(), ''.join(ps).strip(), tks
+
+ @staticmethod
+ def infer(
+ model: KModel,
+ ps: str,
+ pack: torch.FloatTensor,
+ speed: Union[float, Callable[[int], float]] = 1
+ ) -> KModel.Output:
+ if callable(speed):
+ speed = speed(len(ps))
+ return model(ps, pack[len(ps)-1], speed, return_output=True)
+
+ def generate_from_tokens(
+ self,
+ tokens: Union[str, List[en.MToken]],
+ voice: str,
+ speed: float = 1,
+ model: Optional[KModel] = None
+ ) -> Generator['KPipeline.Result', None, None]:
+ """Generate audio from either raw phonemes or pre-processed tokens.
+
+ Args:
+ tokens: Either a phoneme string or list of pre-processed MTokens
+ voice: The voice to use for synthesis
+ speed: Speech speed modifier (default: 1)
+ model: Optional KModel instance (uses pipeline's model if not provided)
+
+ Yields:
+ KPipeline.Result containing the input tokens and generated audio
+
+ Raises:
+ ValueError: If no voice is provided or token sequence exceeds model limits
+ """
+ model = model or self.model
+ if model and voice is None:
+ raise ValueError('Specify a voice: pipeline.generate_from_tokens(..., voice="af_heart")')
+
+ pack = self.load_voice(voice).to(model.device) if model else None
+
+ # Handle raw phoneme string
+ if isinstance(tokens, str):
+ logger.debug("Processing phonemes from raw string")
+ if len(tokens) > 510:
+ raise ValueError(f'Phoneme string too long: {len(tokens)} > 510')
+ output = KPipeline.infer(model, tokens, pack, speed) if model else None
+ yield self.Result(graphemes='', phonemes=tokens, output=output)
+ return
+
+ logger.debug("Processing MTokens")
+ # Handle pre-processed tokens
+ for gs, ps, tks in self.en_tokenize(tokens):
+ if not ps:
+ continue
+ elif len(ps) > 510:
+ logger.warning(f"Unexpected len(ps) == {len(ps)} > 510 and ps == '{ps}'")
+ logger.warning("Truncating to 510 characters")
+ ps = ps[:510]
+ output = KPipeline.infer(model, ps, pack, speed) if model else None
+ if output is not None and output.pred_dur is not None:
+ KPipeline.join_timestamps(tks, output.pred_dur)
+ yield self.Result(graphemes=gs, phonemes=ps, tokens=tks, output=output)
+
+ @staticmethod
+ def join_timestamps(tokens: List[en.MToken], pred_dur: torch.LongTensor):
+ # Multiply by 600 to go from pred_dur frames to sample_rate 24000
+ # Equivalent to dividing pred_dur frames by 40 to get timestamp in seconds
+ # We will count nice round half-frames, so the divisor is 80
+ MAGIC_DIVISOR = 80
+ if not tokens or len(pred_dur) < 3:
+ # We expect at least 3: , token,
+ return
+ # We track 2 counts, measured in half-frames: (left, right)
+ # This way we can cut space characters in half
+ # TODO: Is -3 an appropriate offset?
+ left = right = 2 * max(0, pred_dur[0].item() - 3)
+ # Updates:
+ # left = right + (2 * token_dur) + space_dur
+ # right = left + space_dur
+ i = 1
+ for t in tokens:
+ if i >= len(pred_dur)-1:
+ break
+ if not t.phonemes:
+ if t.whitespace:
+ i += 1
+ left = right + pred_dur[i].item()
+ right = left + pred_dur[i].item()
+ i += 1
+ continue
+ j = i + len(t.phonemes)
+ if j >= len(pred_dur):
+ break
+ t.start_ts = left / MAGIC_DIVISOR
+ token_dur = pred_dur[i: j].sum().item()
+ space_dur = pred_dur[j].item() if t.whitespace else 0
+ left = right + (2 * token_dur) + space_dur
+ t.end_ts = left / MAGIC_DIVISOR
+ right = left + space_dur
+ i = j + (1 if t.whitespace else 0)
+
+ @dataclass
+ class Result:
+ graphemes: str
+ phonemes: str
+ tokens: Optional[List[en.MToken]] = None
+ output: Optional[KModel.Output] = None
+ text_index: Optional[int] = None
+
+ @property
+ def audio(self) -> Optional[torch.FloatTensor]:
+ return None if self.output is None else self.output.audio
+
+ @property
+ def pred_dur(self) -> Optional[torch.LongTensor]:
+ return None if self.output is None else self.output.pred_dur
+
+ ### MARK: BEGIN BACKWARD COMPAT ###
+ def __iter__(self):
+ yield self.graphemes
+ yield self.phonemes
+ yield self.audio
+
+ def __getitem__(self, index):
+ return [self.graphemes, self.phonemes, self.audio][index]
+
+ def __len__(self):
+ return 3
+ #### MARK: END BACKWARD COMPAT ####
+
+ def __call__(
+ self,
+ text: Union[str, List[str]],
+ voice: Optional[str] = None,
+ speed: Union[float, Callable[[int], float]] = 1,
+ split_pattern: Optional[str] = r'\n+',
+ model: Optional[KModel] = None
+ ) -> Generator['KPipeline.Result', None, None]:
+ model = model or self.model
+ if model and voice is None:
+ raise ValueError('Specify a voice: en_us_pipeline(text="Hello world!", voice="af_heart")')
+ pack = self.load_voice(voice).to(model.device) if model else None
+
+ # Convert input to list of segments
+ if isinstance(text, str):
+ text = re.split(split_pattern, text.strip()) if split_pattern else [text]
+
+ # Process each segment
+ for graphemes_index, graphemes in enumerate(text):
+ if not graphemes.strip(): # Skip empty segments
+ continue
+
+ # English processing (unchanged)
+ if self.lang_code in 'ab':
+ logger.debug(f"Processing English text: {graphemes[:50]}{'...' if len(graphemes) > 50 else ''}")
+ _, tokens = self.g2p(graphemes)
+ for gs, ps, tks in self.en_tokenize(tokens):
+ if not ps:
+ continue
+ elif len(ps) > 510:
+ logger.warning(f"Unexpected len(ps) == {len(ps)} > 510 and ps == '{ps}'")
+ ps = ps[:510]
+ output = KPipeline.infer(model, ps, pack, speed) if model else None
+ if output is not None and output.pred_dur is not None:
+ KPipeline.join_timestamps(tks, output.pred_dur)
+ yield self.Result(graphemes=gs, phonemes=ps, tokens=tks, output=output, text_index=graphemes_index)
+
+ # Non-English processing with chunking
+ else:
+ # Split long text into smaller chunks (roughly 400 characters each)
+ # Using sentence boundaries when possible
+ chunk_size = 400
+ chunks = []
+
+ # Try to split on sentence boundaries first
+ sentences = re.split(r'([.!?]+)', graphemes)
+ current_chunk = ""
+
+ for i in range(0, len(sentences), 2):
+ sentence = sentences[i]
+ # Add the punctuation back if it exists
+ if i + 1 < len(sentences):
+ sentence += sentences[i + 1]
+
+ if len(current_chunk) + len(sentence) <= chunk_size:
+ current_chunk += sentence
+ else:
+ if current_chunk:
+ chunks.append(current_chunk.strip())
+ current_chunk = sentence
+
+ if current_chunk:
+ chunks.append(current_chunk.strip())
+
+ # If no chunks were created (no sentence boundaries), fall back to character-based chunking
+ if not chunks:
+ chunks = [graphemes[i:i+chunk_size] for i in range(0, len(graphemes), chunk_size)]
+
+ # Process each chunk
+ for chunk in chunks:
+ if not chunk.strip():
+ continue
+
+ ps, _ = self.g2p(chunk)
+ if not ps:
+ continue
+ elif len(ps) > 510:
+ logger.warning(f'Truncating len(ps) == {len(ps)} > 510')
+ ps = ps[:510]
+
+ output = KPipeline.infer(model, ps, pack, speed) if model else None
+ yield self.Result(graphemes=chunk, phonemes=ps, output=output, text_index=graphemes_index)
diff --git a/requirements.txt b/requirements.txt
index 74ec9d1..6710040 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -9,7 +9,10 @@ easydict
ftfy
dashscope
imageio-ffmpeg
+scikit-image
+loguru
gradio>=5.0.0
numpy>=1.23.5,<2
xfuser>=0.4.1
-pyloudnorm
\ No newline at end of file
+pyloudnorm
+optimum-quanto==0.2.6
\ No newline at end of file
diff --git a/src/vram_management/layers.py b/src/vram_management/layers.py
index ba19056..cc85a6c 100644
--- a/src/vram_management/layers.py
+++ b/src/vram_management/layers.py
@@ -3,13 +3,20 @@
import torch
from src.utils import init_weights_on_device
-
+import optimum.quanto.nn.qlinear as qlinear
def cast_to(weight, dtype, device):
r = torch.empty_like(weight, dtype=dtype, device=device)
r.copy_(weight)
return r
+def cast_to_device(weight, device):
+ if hasattr(weight, '__class__') and 'optimum.quanto' in str(weight.__class__):
+ return weight.to(device)
+ else:
+ r = torch.empty_like(weight, device=device)
+ r.copy_(weight)
+ return r
class AutoWrappedModule(torch.nn.Module):
def __init__(
@@ -61,6 +68,63 @@ def forward(self, *args, **kwargs):
return module(*args, **kwargs)
+
+class AutoWrappedQLinear(qlinear.QLinear):
+ def __init__(
+ self,
+ module: qlinear.QLinear,
+ offload_dtype,
+ offload_device,
+ onload_dtype,
+ onload_device,
+ computation_dtype,
+ computation_device,
+ ):
+ with init_weights_on_device(device=torch.device("meta")):
+ super().__init__(
+ in_features=module.in_features,
+ out_features=module.out_features,
+ bias=module.bias is not None,
+ device=offload_device,
+ )
+ self.weight = module.weight
+ self.bias = module.bias
+ self.offload_device = offload_device
+
+ self.onload_device = onload_device
+ self.computation_device = computation_device
+ self.state = 0
+
+ def offload(self):
+ if self.state == 1 and (
+ self.offload_device != self.onload_device
+ ):
+ self.to(device=self.offload_device)
+ self.state = 0
+
+ def onload(self):
+ if self.state == 0 and (
+ self.offload_device != self.onload_device
+ ):
+ self.to(device=self.onload_device)
+ self.state = 1
+
+ def forward(self, x, *args, **kwargs):
+ if (
+ self.onload_device == self.computation_device
+ ):
+
+ return torch.nn.functional.linear(x, self.weight, bias=self.bias)
+ else:
+
+ qweight = cast_to_device(self.weight, self.computation_device)
+ bias = (
+ None
+ if self.bias is None
+ else cast_to_device(self.bias, self.computation_device)
+ )
+ return torch.nn.functional.linear(x, qweight, bias)
+
class AutoWrappedLinear(torch.nn.Linear):
def __init__(
self,
diff --git a/wan/modules/multitalk_model.py b/wan/modules/multitalk_model.py
index 25af83c..91b28ed 100644
--- a/wan/modules/multitalk_model.py
+++ b/wan/modules/multitalk_model.py
@@ -13,6 +13,13 @@
from .attention import flash_attention, SingleStreamMutiAttention
from ..utils.multitalk_utils import get_attn_map_with_target
+import logging
+try:
+ from sageattention import sageattn
+ USE_SAGEATTN = True
+ logging.info("Using sageattn")
+except:
+ USE_SAGEATTN = False
__all__ = ['WanModel']
@@ -144,14 +151,16 @@ def qkv_fn(x):
q = rope_apply(q, grid_sizes, freqs)
k = rope_apply(k, grid_sizes, freqs)
-
- x = flash_attention(
- q=q,
- k=k,
- v=v,
- k_lens=seq_lens,
- window_size=self.window_size
- ).type_as(x)
+ if USE_SAGEATTN:
+ x = sageattn(q.to(torch.bfloat16), k.to(torch.bfloat16), v, tensor_layout='NHD')
+ else:
+ x = flash_attention(
+ q=q,
+ k=k,
+ v=v,
+ k_lens=seq_lens,
+ window_size=self.window_size
+ ).type_as(x)
# output
x = x.flatten(2)
@@ -188,9 +197,13 @@ def forward(self, x, context, context_lens):
v = self.v(context).view(b, -1, n, d)
k_img = self.norm_k_img(self.k_img(context_img)).view(b, -1, n, d)
v_img = self.v_img(context_img).view(b, -1, n, d)
- img_x = flash_attention(q, k_img, v_img, k_lens=None)
- # compute attention
- x = flash_attention(q, k, v, k_lens=context_lens)
+ if USE_SAGEATTN:
+ img_x = sageattn(q, k_img, v_img, tensor_layout='NHD')
+ x = sageattn(q, k, v, tensor_layout='NHD')
+ else:
+ img_x = flash_attention(q, k_img, v_img, k_lens=None)
+ # compute attention
+ x = flash_attention(q, k, v, k_lens=context_lens)
# output
x = x.flatten(2)
@@ -450,7 +463,8 @@ def __init__(self,
vae_scale=4, # vae timedownsample scale
norm_input_visual=True,
- norm_output_audio=True):
+ norm_output_audio=True,
+ weight_init=True):
super().__init__()
assert model_type == 'i2v', 'MultiTalk model requires your model_type is i2v.'
@@ -527,7 +541,17 @@ def __init__(self,
# initialize weights
- self.init_weights()
+ if weight_init:
+ self.init_weights()
+
+ def init_freqs(self):
+ d = self.dim // self.num_heads
+ self.freqs = torch.cat([
+ rope_params(1024, d - 4 * (d // 6)),
+ rope_params(1024, 2 * (d // 6)),
+ rope_params(1024, 2 * (d // 6))
+ ],
+ dim=1)
def teacache_init(
self,
diff --git a/wan/modules/t5.py b/wan/modules/t5.py
index c841b04..eb08105 100644
--- a/wan/modules/t5.py
+++ b/wan/modules/t5.py
@@ -2,11 +2,16 @@
# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
import logging
import math
+import json
+import os
import torch
import torch.nn as nn
import torch.nn.functional as F
+from safetensors.torch import load_file
+from optimum.quanto import quantize, freeze, qint8,requantize
+
from .tokenizers import HuggingfaceTokenizer
__all__ = [
@@ -479,7 +484,10 @@ def __init__(
checkpoint_path=None,
tokenizer_path=None,
shard_fn=None,
+ quant=None,
+ quant_dir=None
):
+ assert quant is None or quant in ("int8", "fp8")
self.text_len = text_len
self.dtype = dtype
self.device = device
@@ -487,14 +495,28 @@ def __init__(
self.tokenizer_path = tokenizer_path
# init model
- model = umt5_xxl(
- encoder_only=True,
- return_tokenizer=False,
- dtype=dtype,
- device=device).eval().requires_grad_(False)
logging.info(f'loading {checkpoint_path}')
- model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
+ if quant is not None:
+ with torch.device('meta'):
+ model = umt5_xxl(
+ encoder_only=True,
+ return_tokenizer=False,
+ dtype=dtype,
+ device=torch.device('meta'))
+ logging.info(f'Loading quantized T5 from {os.path.join(quant_dir, "quant_models", f"t5_{quant}.safetensors")}')
+ model_state_dict = load_file(os.path.join(quant_dir, "quant_models", f"t5_{quant}.safetensors"))
+ with open(os.path.join(quant_dir, "quant_models", f"t5_map_{quant}.json"), "r") as f:
+ quantization_map = json.load(f)
+ requantize(model, model_state_dict, quantization_map, device='cpu')
+ else:
+ model = umt5_xxl(
+ encoder_only=True,
+ return_tokenizer=False,
+ dtype=dtype,
+ device=device).eval().requires_grad_(False)
+ model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
self.model = model
+ self.model.eval().requires_grad_(False)
if shard_fn is not None:
self.model = shard_fn(self.model, sync_module_states=False)
else:
diff --git a/wan/multitalk.py b/wan/multitalk.py
index bd3f81c..2685c42 100644
--- a/wan/multitalk.py
+++ b/wan/multitalk.py
@@ -1,6 +1,8 @@
# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
import gc
+from inspect import ArgSpec
import logging
+import json
import math
import importlib
import os
@@ -25,15 +27,27 @@
from .modules.multitalk_model import WanModel, WanLayerNorm, WanRMSNorm
from .modules.t5 import T5EncoderModel, T5LayerNorm, T5RelativeEmbedding
from .modules.vae import WanVAE, CausalConv3d, RMS_norm, Upsample
-from .utils.multitalk_utils import MomentumBuffer, adaptive_projected_guidance
-from src.vram_management import AutoWrappedLinear, AutoWrappedModule, enable_vram_management
+from .utils.multitalk_utils import MomentumBuffer, adaptive_projected_guidance, match_and_blend_colors
+from src.vram_management import AutoWrappedQLinear, AutoWrappedLinear, AutoWrappedModule, enable_vram_management
+from wan.utils.utils import load_torch_file, standardize_lora_key_format, load_lora_for_models, apply_lora
+from wan.wan_lora import WanLoraWrapper
+from safetensors.torch import load_file
+from optimum.quanto import quantize, freeze, qint8,requantize
+import optimum.quanto.nn.qlinear as qlinear
def torch_gc():
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
-
+def to_param_dtype_fp32only(model, param_dtype):
+ for module in model.modules():
+ for name, param in module.named_parameters(recurse=False):
+ if param.dtype == torch.float32:
+ param.data = param.data.to(param_dtype)
+ for name, buf in module.named_buffers(recurse=False):
+ if buf.dtype == torch.float32:
+ module._buffers[name] = buf.to(param_dtype)
def resize_and_centercrop(cond_image, target_size):
"""
Resize image or tensor to the target size without padding.
@@ -94,6 +108,7 @@ def __init__(
self,
config,
checkpoint_dir,
+ quant_dir=None,
device_id=0,
rank=0,
t5_fsdp=False,
@@ -102,7 +117,10 @@ def __init__(
t5_cpu=False,
init_on_cpu=True,
num_timesteps=1000,
- use_timestep_transform=True
+ use_timestep_transform=True,
+ lora_dir=None,
+ lora_scales=None,
+ quant = None
):
r"""
Initializes the image-to-video generation model components.
@@ -126,7 +144,11 @@ def __init__(
Whether to place T5 model on CPU. Only works without t5_fsdp.
init_on_cpu (`bool`, *optional*, defaults to True):
Enable initializing Transformer Model on CPU. Only works without FSDP or USP.
+ quant (`str`, *optional*, defaults to None):
+ Quantization type, must be 'int8' or 'fp8'.
"""
+ if quant is not None and quant not in ("int8", "fp8"):
+ raise ValueError("quant must be 'int8', 'fp8', or None(default fp32 model)")
self.device = torch.device(f"cuda:{device_id}")
self.config = config
self.rank = rank
@@ -144,6 +166,8 @@ def __init__(
checkpoint_path=os.path.join(checkpoint_dir, config.t5_checkpoint),
tokenizer_path=os.path.join(checkpoint_dir, config.t5_tokenizer),
shard_fn=shard_fn if t5_fsdp else None,
+ quant=quant,
+ quant_dir=quant_dir,
)
self.vae_stride = config.vae_stride
@@ -160,10 +184,38 @@ def __init__(
tokenizer_path=os.path.join(checkpoint_dir, config.clip_tokenizer))
logging.info(f"Creating WanModel from {checkpoint_dir}")
- self.model = WanModel.from_pretrained(checkpoint_dir)
+ if quant is not None:
+ logging.info(f"Loading Quantized MultiTalk from {os.path.join(quant_dir,'quant_models')}")
+ with torch.device('meta'):
+ wan_config = json.load(open(os.path.join(checkpoint_dir, "config.json")))
+ self.model = WanModel(weight_init=False,**wan_config)
+ torch_gc()
+ # load quantized model
+ if lora_dir is not None:
+ logging.info(f"Loading Quantized LoRA from {lora_dir[0]}")
+ model_state_dict = load_file(lora_dir[0])
+ map_json_path = os.path.join(os.path.dirname(lora_dir[0]),f"quantization_map_{quant}_FusionX.json")
+ else:
+ model_state_dict = load_file(os.path.join(quant_dir,"quant_models", f"dit_model_{quant}.safetensors"))
+ map_json_path = os.path.join(quant_dir,"quant_models", f"dit_model_map_{quant}.json")
+ self.model.init_freqs()
+ with open(map_json_path, "r") as f:
+ quantization_map = json.load(f)
+ requantize(self.model, model_state_dict, quantization_map, device='cpu')
+ else:
+ self.model = WanModel.from_pretrained(checkpoint_dir)
self.model.eval().requires_grad_(False)
+
+ to_param_dtype_fp32only(self.model, self.param_dtype)
+ if lora_dir is not None and quant is None :
+ lora_wrapper = WanLoraWrapper(self.model)
+ for lora_path, lora_scale in zip(lora_dir, lora_scales):
+ lora_name = lora_wrapper.load_lora(lora_path)
+ lora_wrapper.apply_lora(lora_name, lora_scale, param_dtype=self.param_dtype, device=self.device)
+
+
if t5_fsdp or dit_fsdp or use_usp:
init_on_cpu = False
if use_usp:
@@ -184,7 +236,7 @@ def __init__(
else:
self.sp_size = 1
- self.model.to(self.param_dtype)
+
if dist.is_initialized():
dist.barrier()
@@ -221,6 +273,7 @@ def enable_vram_management(self, num_persistent_param_in_dit=None):
enable_vram_management(
self.model,
module_map={
+ qlinear.QLinear: AutoWrappedQLinear,
torch.nn.Linear: AutoWrappedLinear,
torch.nn.Conv3d: AutoWrappedModule,
torch.nn.LayerNorm: AutoWrappedModule,
@@ -305,6 +358,7 @@ def generate(self,
max_frames_num=1000,
face_scale=0.05,
progress=True,
+ color_correction_strength=0.0,
extra_args=None):
r"""
Generates video frames from input image and text prompt using diffusion process.
@@ -357,6 +411,11 @@ def generate(self,
cond_image = (cond_image - 0.5) * 2 # normalization
cond_image = cond_image.to(self.device) # 1 C 1 H W
+ # Store the original image for color reference if strength > 0
+ original_color_reference = None
+ if color_correction_strength > 0.0:
+ original_color_reference = cond_image.clone()
+
# read audio embeddings
audio_embedding_path_1 = input_data['cond_audio']['person1']
@@ -502,9 +561,8 @@ def generate(self,
background_mask = torch.zeros([src_h, src_w])
human_mask1 = torch.zeros([src_h, src_w])
human_mask2 = torch.zeros([src_h, src_w])
- src_w = src_w//2
- lefty_min, lefty_max = int(src_w * face_scale), int(src_w * (1 - face_scale))
- righty_min, righty_max = int(src_w * face_scale + src_w), int(src_w * (1 - face_scale) + src_w)
+ lefty_min, lefty_max = int((src_w//2) * face_scale), int((src_w//2) * (1 - face_scale))
+ righty_min, righty_max = int((src_w//2) * face_scale + (src_w//2)), int((src_w//2) * (1 - face_scale) + (src_w//2))
human_mask1[x_min:x_max, lefty_min:lefty_max] = 1
human_mask2[x_min:x_max, righty_min:righty_max] = 1
background_mask += human_mask1
@@ -563,6 +621,15 @@ def noop_no_sync():
'ref_target_masks': ref_target_masks
}
+ arg_null_audio = {
+ 'context': [context],
+ 'clip_fea': clip_context,
+ 'seq_len': max_seq_len,
+ 'y': y,
+ 'audio': torch.zeros_like(audio_embs)[-1:],
+ 'ref_target_masks': ref_target_masks
+ }
+
arg_null = {
'context': [context_null],
@@ -603,30 +670,46 @@ def noop_no_sync():
noise_pred_cond = self.model(
latent_model_input, t=timestep, **arg_c)[0]
torch_gc()
- noise_pred_drop_text = self.model(
- latent_model_input, t=timestep, **arg_null_text)[0]
- torch_gc()
- noise_pred_uncond = self.model(
- latent_model_input, t=timestep, **arg_null)[0]
- torch_gc()
+
+ if math.isclose(text_guide_scale, 1.0):
+ noise_pred_drop_audio = self.model(
+ latent_model_input, t=timestep, **arg_null_audio)[0]
+ torch_gc()
+ else:
+ noise_pred_drop_text = self.model(
+ latent_model_input, t=timestep, **arg_null_text)[0]
+ torch_gc()
+ noise_pred_uncond = self.model(
+ latent_model_input, t=timestep, **arg_null)[0]
+ torch_gc()
if extra_args.use_apg:
# correct update direction
- diff_uncond_text = noise_pred_cond - noise_pred_drop_text
- diff_uncond_audio = noise_pred_drop_text - noise_pred_uncond
- noise_pred = noise_pred_cond + (text_guide_scale - 1) * adaptive_projected_guidance(diff_uncond_text,
- noise_pred_cond,
- momentum_buffer=text_momentumbuffer,
- norm_threshold=extra_args.apg_norm_threshold) \
- + (audio_guide_scale - 1) * adaptive_projected_guidance(diff_uncond_audio,
- noise_pred_cond,
- momentum_buffer=audio_momentumbuffer,
- norm_threshold=extra_args.apg_norm_threshold)
+ if math.isclose(text_guide_scale, 1.0):
+ diff_uncond_audio = noise_pred_cond - noise_pred_drop_audio
+ noise_pred = noise_pred_cond + (audio_guide_scale - 1)* adaptive_projected_guidance(diff_uncond_audio,
+ noise_pred_cond,
+ momentum_buffer=audio_momentumbuffer,
+ norm_threshold=extra_args.apg_norm_threshold)
+ else:
+ diff_uncond_text = noise_pred_cond - noise_pred_drop_text
+ diff_uncond_audio = noise_pred_drop_text - noise_pred_uncond
+ noise_pred = noise_pred_cond + (text_guide_scale - 1) * adaptive_projected_guidance(diff_uncond_text,
+ noise_pred_cond,
+ momentum_buffer=text_momentumbuffer,
+ norm_threshold=extra_args.apg_norm_threshold) \
+ + (audio_guide_scale - 1) * adaptive_projected_guidance(diff_uncond_audio,
+ noise_pred_cond,
+ momentum_buffer=audio_momentumbuffer,
+ norm_threshold=extra_args.apg_norm_threshold)
else:
# vanilla CFG strategy
- noise_pred = noise_pred_uncond + text_guide_scale * (
- noise_pred_cond - noise_pred_drop_text) + \
- audio_guide_scale * (noise_pred_drop_text - noise_pred_uncond)
+ if math.isclose(text_guide_scale, 1.0):
+ noise_pred = noise_pred_drop_audio + audio_guide_scale* (noise_pred_cond - noise_pred_drop_audio)
+ else:
+ noise_pred = noise_pred_uncond + text_guide_scale * (
+ noise_pred_cond - noise_pred_drop_text) + \
+ audio_guide_scale * (noise_pred_drop_text - noise_pred_uncond)
noise_pred = -noise_pred
# update latent
@@ -654,6 +737,11 @@ def noop_no_sync():
# cache generated samples
videos = torch.stack(videos).cpu() # B C T H W
+ # >>> START OF COLOR CORRECTION STEP <<<
+ if color_correction_strength > 0.0 and original_color_reference is not None:
+ videos = match_and_blend_colors(videos, original_color_reference, color_correction_strength)
+ # >>> END OF COLOR CORRECTION STEP <<<
+
if is_first_clip:
gen_video_list.append(videos)
else:
@@ -706,4 +794,4 @@ def noop_no_sync():
del noise, latent
torch_gc()
- return gen_video_samples[0] if self.rank == 0 else None
\ No newline at end of file
+ return gen_video_samples[0] if self.rank == 0 else None
diff --git a/wan/utils/multitalk_utils.py b/wan/utils/multitalk_utils.py
index 1215961..d33be09 100644
--- a/wan/utils/multitalk_utils.py
+++ b/wan/utils/multitalk_utils.py
@@ -17,7 +17,10 @@
import numpy as np
import subprocess
import soundfile as sf
-
+import torchvision
+import binascii
+import os.path as osp
+from skimage import color
VID_EXTENSIONS = (".mp4", ".avi", ".mov", ".mkv")
ASPECT_RATIO_627 = {
@@ -188,7 +191,50 @@ def forward(self, x, pos_indices):
return x_.type_as(x)
-def save_video_ffmpeg(gen_video_samples, save_path, vocal_audio_list, fps=25, quality=5):
+
+def rand_name(length=8, suffix=''):
+ name = binascii.b2a_hex(os.urandom(length)).decode('utf-8')
+ if suffix:
+ if not suffix.startswith('.'):
+ suffix = '.' + suffix
+ name += suffix
+ return name
+
+def cache_video(tensor,
+ save_file=None,
+ fps=30,
+ suffix='.mp4',
+ nrow=8,
+ normalize=True,
+ value_range=(-1, 1),
+ retry=5):
+
+ # cache file
+ cache_file = osp.join('/tmp', rand_name(
+ suffix=suffix)) if save_file is None else save_file
+
+ # save to cache
+ error = None
+ for _ in range(retry):
+
+ # preprocess
+ tensor = tensor.clamp(min(value_range), max(value_range))
+ tensor = torch.stack([
+ torchvision.utils.make_grid(
+ u, nrow=nrow, normalize=normalize, value_range=value_range)
+ for u in tensor.unbind(2)
+ ],
+ dim=1).permute(1, 2, 3, 0)
+ tensor = (tensor * 255).type(torch.uint8).cpu()
+
+ # write video
+ writer = imageio.get_writer(cache_file, fps=fps, codec='libx264', quality=10, ffmpeg_params=["-crf", "10"])
+ for frame in tensor.numpy():
+ writer.append_data(frame)
+ writer.close()
+ return cache_file
+
+def save_video_ffmpeg(gen_video_samples, save_path, vocal_audio_list, fps=25, quality=5, high_quality_save=False):
def save_video(frames, save_path, fps, quality=9, ffmpeg_params=None):
writer = imageio.get_writer(
@@ -200,10 +246,20 @@ def save_video(frames, save_path, fps, quality=9, ffmpeg_params=None):
writer.close()
save_path_tmp = save_path + "-temp.mp4"
- video_audio = (gen_video_samples+1)/2 # C T H W
- video_audio = video_audio.permute(1, 2, 3, 0).cpu().numpy()
- video_audio = np.clip(video_audio * 255, 0, 255).astype(np.uint8)
- save_video(video_audio, save_path_tmp, fps=fps, quality=quality)
+ if high_quality_save:
+ cache_video(
+ tensor=gen_video_samples.unsqueeze(0),
+ save_file=save_path_tmp,
+ fps=fps,
+ nrow=1,
+ normalize=True,
+ value_range=(-1, 1)
+ )
+ else:
+ video_audio = (gen_video_samples+1)/2 # C T H W
+ video_audio = video_audio.permute(1, 2, 3, 0).cpu().numpy()
+ video_audio = np.clip(video_audio * 255, 0, 255).astype(np.uint8) # to [0, 255]
+ save_video(video_audio, save_path_tmp, fps=fps, quality=quality)
# crop audio according to video length
@@ -220,27 +276,41 @@ def save_video(frames, save_path, fps, quality=9, ffmpeg_params=None):
]
subprocess.run(final_command, check=True)
-
- # generate video with audio
save_path = save_path + ".mp4"
- final_command = [
- "ffmpeg",
- "-y",
- "-i",
- save_path_tmp,
- "-i",
- save_path_crop_audio,
- "-c:v",
- "libx264",
- "-c:a",
- "aac",
- "-shortest",
- save_path,
- ]
- subprocess.run(final_command, check=True)
- os.remove(save_path_tmp)
- os.remove(save_path_crop_audio)
-
+ if high_quality_save:
+ final_command = [
+ "ffmpeg",
+ "-y",
+ "-i", save_path_tmp,
+ "-i", save_path_crop_audio,
+ "-c:v", "libx264",
+ "-crf", "0",
+ "-preset", "veryslow",
+ "-c:a", "aac",
+ "-shortest",
+ save_path,
+ ]
+ subprocess.run(final_command, check=True)
+ os.remove(save_path_tmp)
+ os.remove(save_path_crop_audio)
+ else:
+ final_command = [
+ "ffmpeg",
+ "-y",
+ "-i",
+ save_path_tmp,
+ "-i",
+ save_path_crop_audio,
+ "-c:v",
+ "libx264",
+ "-c:a",
+ "aac",
+ "-shortest",
+ save_path,
+ ]
+ subprocess.run(final_command, check=True)
+ os.remove(save_path_tmp)
+ os.remove(save_path_crop_audio)
class MomentumBuffer:
@@ -284,4 +354,110 @@ def adaptive_projected_guidance(
diff = diff * scale_factor
diff_parallel, diff_orthogonal = project(diff, pred_cond)
normalized_update = diff_orthogonal + eta * diff_parallel
- return normalized_update
\ No newline at end of file
+ return normalized_update
+
+
+
+def match_and_blend_colors(source_chunk: torch.Tensor, reference_image: torch.Tensor, strength: float) -> torch.Tensor:
+ """
+ Matches the color of a source video chunk to a reference image and blends with the original.
+
+ Args:
+ source_chunk (torch.Tensor): The video chunk to be color-corrected (B, C, T, H, W) in range [-1, 1].
+ Assumes B=1 (batch size of 1).
+ reference_image (torch.Tensor): The reference image (B, C, 1, H, W) in range [-1, 1].
+ Assumes B=1 and T=1 (single reference frame).
+ strength (float): The strength of the color correction (0.0 to 1.0).
+ 0.0 means no correction, 1.0 means full correction.
+
+ Returns:
+ torch.Tensor: The color-corrected and blended video chunk.
+ """
+ # print(f"[match_and_blend_colors] Input source_chunk shape: {source_chunk.shape}, reference_image shape: {reference_image.shape}, strength: {strength}")
+
+ if strength == 0.0:
+ # print(f"[match_and_blend_colors] Strength is 0, returning original source_chunk.")
+ return source_chunk
+
+ if not 0.0 <= strength <= 1.0:
+ raise ValueError(f"Strength must be between 0.0 and 1.0, got {strength}")
+
+ device = source_chunk.device
+ dtype = source_chunk.dtype
+
+ # Squeeze batch dimension, permute to T, H, W, C for skimage
+ # Source: (1, C, T, H, W) -> (T, H, W, C)
+ source_np = source_chunk.squeeze(0).permute(1, 2, 3, 0).cpu().numpy()
+ # Reference: (1, C, 1, H, W) -> (H, W, C)
+ ref_np = reference_image.squeeze(0).squeeze(1).permute(1, 2, 0).cpu().numpy() # Squeeze T dimension as well
+
+ # Normalize from [-1, 1] to [0, 1] for skimage
+ source_np_01 = (source_np + 1.0) / 2.0
+ ref_np_01 = (ref_np + 1.0) / 2.0
+
+ # Clip to ensure values are strictly in [0, 1] after potential float precision issues
+ source_np_01 = np.clip(source_np_01, 0.0, 1.0)
+ ref_np_01 = np.clip(ref_np_01, 0.0, 1.0)
+
+ # Convert reference to Lab
+ try:
+ ref_lab = color.rgb2lab(ref_np_01)
+ except ValueError as e:
+ # Handle potential errors if image data is not valid for conversion
+ print(f"Warning: Could not convert reference image to Lab: {e}. Skipping color correction for this chunk.")
+ return source_chunk
+
+
+ corrected_frames_np_01 = []
+ for i in range(source_np_01.shape[0]): # Iterate over time (T)
+ source_frame_rgb_01 = source_np_01[i]
+
+ try:
+ source_lab = color.rgb2lab(source_frame_rgb_01)
+ except ValueError as e:
+ print(f"Warning: Could not convert source frame {i} to Lab: {e}. Using original frame.")
+ corrected_frames_np_01.append(source_frame_rgb_01)
+ continue
+
+ corrected_lab_frame = source_lab.copy()
+
+ # Perform color transfer for L, a, b channels
+ for j in range(3): # L, a, b
+ mean_src, std_src = source_lab[:, :, j].mean(), source_lab[:, :, j].std()
+ mean_ref, std_ref = ref_lab[:, :, j].mean(), ref_lab[:, :, j].std()
+
+ # Avoid division by zero if std_src is 0
+ if std_src == 0:
+ # If source channel has no variation, keep it as is, but shift by reference mean
+ # This case is debatable, could also just copy source or target mean.
+ # Shifting by target mean helps if source is flat but target isn't.
+ corrected_lab_frame[:, :, j] = mean_ref
+ else:
+ corrected_lab_frame[:, :, j] = (corrected_lab_frame[:, :, j] - mean_src) * (std_ref / std_src) + mean_ref
+
+ try:
+ fully_corrected_frame_rgb_01 = color.lab2rgb(corrected_lab_frame)
+ except ValueError as e:
+ print(f"Warning: Could not convert corrected frame {i} back to RGB: {e}. Using original frame.")
+ corrected_frames_np_01.append(source_frame_rgb_01)
+ continue
+
+ # Clip again after lab2rgb as it can go slightly out of [0,1]
+ fully_corrected_frame_rgb_01 = np.clip(fully_corrected_frame_rgb_01, 0.0, 1.0)
+
+ # Blend with original source frame (in [0,1] RGB)
+ blended_frame_rgb_01 = (1 - strength) * source_frame_rgb_01 + strength * fully_corrected_frame_rgb_01
+ corrected_frames_np_01.append(blended_frame_rgb_01)
+
+ corrected_chunk_np_01 = np.stack(corrected_frames_np_01, axis=0)
+
+ # Convert back to [-1, 1]
+ corrected_chunk_np_minus1_1 = (corrected_chunk_np_01 * 2.0) - 1.0
+
+ # Permute back to (C, T, H, W), add batch dim, and convert to original torch.Tensor type and device
+ # (T, H, W, C) -> (C, T, H, W)
+ corrected_chunk_tensor = torch.from_numpy(corrected_chunk_np_minus1_1).permute(3, 0, 1, 2).unsqueeze(0)
+ corrected_chunk_tensor = corrected_chunk_tensor.contiguous() # Ensure contiguous memory layout
+ output_tensor = corrected_chunk_tensor.to(device=device, dtype=dtype)
+ # print(f"[match_and_blend_colors] Output tensor shape: {output_tensor.shape}")
+ return output_tensor
diff --git a/wan/utils/utils.py b/wan/utils/utils.py
index d725999..29baa2c 100644
--- a/wan/utils/utils.py
+++ b/wan/utils/utils.py
@@ -116,3 +116,356 @@ def str2bool(v):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected (True/False)')
+
+import safetensors.torch
+import logging
+from accelerate.utils import set_module_tensor_to_device
+from tqdm import tqdm
+
+# ALWAYS_SAFE_LOAD = False
+# if hasattr(torch.serialization, "add_safe_globals"): # TODO: this was added in pytorch 2.4, the unsafe path should be removed once earlier versions are deprecated
+# class ModelCheckpoint:
+# pass
+# ModelCheckpoint.__module__ = "pytorch_lightning.callbacks.model_checkpoint"
+
+# from numpy.core.multiarray import scalar
+# from numpy import dtype
+# from numpy.dtypes import Float64DType
+# from _codecs import encode
+
+# torch.serialization.add_safe_globals([ModelCheckpoint, scalar, dtype, Float64DType, encode])
+# ALWAYS_SAFE_LOAD = True
+# logging.info("Checkpoint files will always be loaded safely.")
+# else:
+# logging.info("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.")
+
+
+def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
+ if device is None:
+ device = torch.device("cpu")
+ metadata = None
+ if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
+ try:
+ with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
+ sd = {}
+ for k in f.keys():
+ sd[k] = f.get_tensor(k)
+ if return_metadata:
+ metadata = f.metadata()
+ except Exception as e:
+ if len(e.args) > 0:
+ message = e.args[0]
+ if "HeaderTooLarge" in message:
+ raise ValueError("{}\n\nFile path: {}\n\nThe safetensors file is corrupt or invalid. Make sure this is actually a safetensors file and not a ckpt or pt or other filetype.".format(message, ckpt))
+ if "MetadataIncompleteBuffer" in message:
+ raise ValueError("{}\n\nFile path: {}\n\nThe safetensors file is corrupt/incomplete. Check the file size and make sure you have copied/downloaded it correctly.".format(message, ckpt))
+ raise e
+ else:
+ if safe_load or ALWAYS_SAFE_LOAD:
+ pl_sd = torch.load(ckpt, map_location=device, weights_only=True)
+ else:
+ pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
+ if "global_step" in pl_sd:
+ logging.debug(f"Global Step: {pl_sd['global_step']}")
+ if "state_dict" in pl_sd:
+ sd = pl_sd["state_dict"]
+ else:
+ if len(pl_sd) == 1:
+ key = list(pl_sd.keys())[0]
+ sd = pl_sd[key]
+ if not isinstance(sd, dict):
+ sd = pl_sd
+ else:
+ sd = pl_sd
+ return (sd, metadata) if return_metadata else sd
+
+def standardize_lora_key_format(lora_sd):
+ new_sd = {}
+ for k, v in lora_sd.items():
+ # Diffusers format
+ if k.startswith('transformer.'):
+ k = k.replace('transformer.', 'diffusion_model.')
+ if k.startswith('pipe.dit.'): #unianimate-dit/diffsynth
+ k = k.replace('pipe.dit.', 'diffusion_model.')
+
+ # Fun LoRA format
+ if k.startswith('lora_unet__'):
+ # Split into main path and weight type parts
+ parts = k.split('.')
+ main_part = parts[0] # e.g. lora_unet__blocks_0_cross_attn_k
+ weight_type = '.'.join(parts[1:]) if len(parts) > 1 else None # e.g. lora_down.weight
+
+ # Process the main part - convert from underscore to dot format
+ if 'blocks_' in main_part:
+ # Extract components
+ components = main_part[len('lora_unet__'):].split('_')
+
+ # Start with diffusion_model
+ new_key = "diffusion_model"
+
+ # Add blocks.N
+ if components[0] == 'blocks':
+ new_key += f".blocks.{components[1]}"
+
+ # Handle different module types
+ idx = 2
+ if idx < len(components):
+ if components[idx] == 'self' and idx+1 < len(components) and components[idx+1] == 'attn':
+ new_key += ".self_attn"
+ idx += 2
+ elif components[idx] == 'cross' and idx+1 < len(components) and components[idx+1] == 'attn':
+ new_key += ".cross_attn"
+ idx += 2
+ elif components[idx] == 'ffn':
+ new_key += ".ffn"
+ idx += 1
+
+ # Add the component (k, q, v, o) and handle img suffix
+ if idx < len(components):
+ component = components[idx]
+ idx += 1
+
+ # Check for img suffix
+ if idx < len(components) and components[idx] == 'img':
+ component += '_img'
+ idx += 1
+
+ new_key += f".{component}"
+
+ # Handle weight type - this is the critical fix
+ if weight_type:
+ if weight_type == 'alpha':
+ new_key += '.alpha'
+ elif weight_type == 'lora_down.weight' or weight_type == 'lora_down':
+ new_key += '.lora_A.weight'
+ elif weight_type == 'lora_up.weight' or weight_type == 'lora_up':
+ new_key += '.lora_B.weight'
+ else:
+ # Keep original weight type if not matching our patterns
+ new_key += f'.{weight_type}'
+ # Add .weight suffix if missing
+ if not new_key.endswith('.weight'):
+ new_key += '.weight'
+
+ k = new_key
+ else:
+ # For other lora_unet__ formats (head, embeddings, etc.)
+ new_key = main_part.replace('lora_unet__', 'diffusion_model.')
+
+ # Fix specific component naming patterns
+ new_key = new_key.replace('_self_attn', '.self_attn')
+ new_key = new_key.replace('_cross_attn', '.cross_attn')
+ new_key = new_key.replace('_ffn', '.ffn')
+ new_key = new_key.replace('blocks_', 'blocks.')
+ new_key = new_key.replace('head_head', 'head.head')
+ new_key = new_key.replace('img_emb', 'img_emb')
+ new_key = new_key.replace('text_embedding', 'text.embedding')
+ new_key = new_key.replace('time_embedding', 'time.embedding')
+ new_key = new_key.replace('time_projection', 'time.projection')
+
+ # Replace remaining underscores with dots, carefully
+ parts = new_key.split('.')
+ final_parts = []
+ for part in parts:
+ if part in ['img_emb', 'self_attn', 'cross_attn']:
+ final_parts.append(part) # Keep these intact
+ else:
+ final_parts.append(part.replace('_', '.'))
+ new_key = '.'.join(final_parts)
+
+ # Handle weight type
+ if weight_type:
+ if weight_type == 'alpha':
+ new_key += '.alpha'
+ elif weight_type == 'lora_down.weight' or weight_type == 'lora_down':
+ new_key += '.lora_A.weight'
+ elif weight_type == 'lora_up.weight' or weight_type == 'lora_up':
+ new_key += '.lora_B.weight'
+ else:
+ new_key += f'.{weight_type}'
+ if not new_key.endswith('.weight'):
+ new_key += '.weight'
+
+ k = new_key
+
+ # Handle special embedded components
+ special_components = {
+ 'time.projection': 'time_projection',
+ 'img.emb': 'img_emb',
+ 'text.emb': 'text_emb',
+ 'time.emb': 'time_emb',
+ }
+ for old, new in special_components.items():
+ if old in k:
+ k = k.replace(old, new)
+
+ # Fix diffusion.model -> diffusion_model
+ if k.startswith('diffusion.model.'):
+ k = k.replace('diffusion.model.', 'diffusion_model.')
+
+ # Finetrainer format
+ if '.attn1.' in k:
+ k = k.replace('.attn1.', '.cross_attn.')
+ k = k.replace('.to_k.', '.k.')
+ k = k.replace('.to_q.', '.q.')
+ k = k.replace('.to_v.', '.v.')
+ k = k.replace('.to_out.0.', '.o.')
+ elif '.attn2.' in k:
+ k = k.replace('.attn2.', '.cross_attn.')
+ k = k.replace('.to_k.', '.k.')
+ k = k.replace('.to_q.', '.q.')
+ k = k.replace('.to_v.', '.v.')
+ k = k.replace('.to_out.0.', '.o.')
+
+ if "img_attn.proj" in k:
+ k = k.replace("img_attn.proj", "img_attn_proj")
+ if "img_attn.qkv" in k:
+ k = k.replace("img_attn.qkv", "img_attn_qkv")
+ if "txt_attn.proj" in k:
+ k = k.replace("txt_attn.proj", "txt_attn_proj")
+ if "txt_attn.qkv" in k:
+ k = k.replace("txt_attn.qkv", "txt_attn_qkv")
+ new_sd[k] = v
+ return new_sd
+
+def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
+ key_map = {}
+ if model is not None:
+ key_map = comfy.lora.model_lora_keys_unet(model.model, key_map)
+ if clip is not None:
+ key_map = comfy.lora.model_lora_keys_clip(clip.cond_stage_model, key_map)
+
+ lora = comfy.lora_convert.convert_lora(lora)
+ loaded = comfy.lora.load_lora(lora, key_map)
+ if model is not None:
+ new_modelpatcher = model.clone()
+ k = new_modelpatcher.add_patches(loaded, strength_model)
+ else:
+ k = ()
+ new_modelpatcher = None
+
+ if clip is not None:
+ new_clip = clip.clone()
+ k1 = new_clip.add_patches(loaded, strength_clip)
+ else:
+ k1 = ()
+ new_clip = None
+ k = set(k)
+ k1 = set(k1)
+ for x in loaded:
+ if (x not in k) and (x not in k1):
+ logging.warning("NOT LOADED {}".format(x))
+
+ return (new_modelpatcher, new_clip)
+
+def apply_lora(model, device_to, transformer_load_device, params_to_keep=None, dtype=None, base_dtype=None, state_dict=None, low_mem_load=False):
+ to_load = []
+ for n, m in model.model.named_modules():
+ params = []
+ skip = False
+ for name, param in m.named_parameters(recurse=False):
+ params.append(name)
+ for name, param in m.named_parameters(recurse=True):
+ if name not in params:
+ skip = True # skip random weights in non leaf modules
+ break
+ if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
+ to_load.append((n, m, params))
+
+ to_load.sort(reverse=True)
+ for x in tqdm(to_load, desc="Loading model and applying LoRA weights:", leave=True):
+ name = x[0]
+ m = x[1]
+ params = x[2]
+ if hasattr(m, "comfy_patched_weights"):
+ if m.comfy_patched_weights == True:
+ continue
+ for param in params:
+ name = name.replace("._orig_mod.", ".") # torch compiled modules have this prefix
+ if low_mem_load:
+ dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
+ if "patch_embedding" in name:
+ dtype_to_use = torch.float32
+ if name.startswith("diffusion_model."):
+ name_no_prefix = name[len("diffusion_model."):]
+ key = "{}.{}".format(name_no_prefix, param)
+ try:
+ set_module_tensor_to_device(model.model.diffusion_model, key, device=transformer_load_device, dtype=dtype_to_use, value=state_dict[key])
+ except:
+ continue
+ model.patch_weight_to_device("{}.{}".format(name, param), device_to=device_to)
+ if low_mem_load:
+ try:
+ set_module_tensor_to_device(model.model.diffusion_model, key, device=transformer_load_device, dtype=dtype_to_use, value=model.model.diffusion_model.state_dict()[key])
+ except:
+ continue
+ m.comfy_patched_weights = True
+
+ model.current_weight_patches_uuid = model.patches_uuid
+ if low_mem_load:
+ for name, param in model.model.diffusion_model.named_parameters():
+ if param.device != transformer_load_device:
+ dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
+ if "patch_embedding" in name:
+ dtype_to_use = torch.float32
+ try:
+ set_module_tensor_to_device(model.model.diffusion_model, name, device=transformer_load_device, dtype=dtype_to_use, value=state_dict[name])
+ except:
+ continue
+ return model
+
+def apply_lora(model, device_to, transformer_load_device, params_to_keep=None, dtype=None, base_dtype=None, state_dict=None, low_mem_load=False):
+ to_load = []
+ for n, m in model.model.named_modules():
+ params = []
+ skip = False
+ for name, param in m.named_parameters(recurse=False):
+ params.append(name)
+ for name, param in m.named_parameters(recurse=True):
+ if name not in params:
+ skip = True # skip random weights in non leaf modules
+ break
+ if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
+ to_load.append((n, m, params))
+
+ to_load.sort(reverse=True)
+ for x in tqdm(to_load, desc="Loading model and applying LoRA weights:", leave=True):
+ name = x[0]
+ m = x[1]
+ params = x[2]
+ if hasattr(m, "comfy_patched_weights"):
+ if m.comfy_patched_weights == True:
+ continue
+ for param in params:
+ name = name.replace("._orig_mod.", ".") # torch compiled modules have this prefix
+ if low_mem_load:
+ dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
+ if "patch_embedding" in name:
+ dtype_to_use = torch.float32
+ if name.startswith("diffusion_model."):
+ name_no_prefix = name[len("diffusion_model."):]
+ key = "{}.{}".format(name_no_prefix, param)
+ try:
+ set_module_tensor_to_device(model.model.diffusion_model, key, device=transformer_load_device, dtype=dtype_to_use, value=state_dict[key])
+ except:
+ continue
+ model.patch_weight_to_device("{}.{}".format(name, param), device_to=device_to)
+ if low_mem_load:
+ try:
+ set_module_tensor_to_device(model.model.diffusion_model, key, device=transformer_load_device, dtype=dtype_to_use, value=model.model.diffusion_model.state_dict()[key])
+ except:
+ continue
+ m.comfy_patched_weights = True
+
+ model.current_weight_patches_uuid = model.patches_uuid
+ if low_mem_load:
+ for name, param in model.model.diffusion_model.named_parameters():
+ if param.device != transformer_load_device:
+ dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
+ if "patch_embedding" in name:
+ dtype_to_use = torch.float32
+ try:
+ set_module_tensor_to_device(model.model.diffusion_model, name, device=transformer_load_device, dtype=dtype_to_use, value=state_dict[name])
+ except:
+ continue
+ return model
\ No newline at end of file
diff --git a/wan/wan_lora.py b/wan/wan_lora.py
new file mode 100644
index 0000000..e453f2c
--- /dev/null
+++ b/wan/wan_lora.py
@@ -0,0 +1,113 @@
+import os
+import torch
+from safetensors import safe_open
+from loguru import logger
+import gc
+from functools import lru_cache
+from tqdm import tqdm
+
+@lru_cache(maxsize=None)
+def GET_DTYPE():
+ RUNNING_FLAG = os.getenv("DTYPE")
+ return RUNNING_FLAG
+
+class WanLoraWrapper:
+ def __init__(self, wan_model):
+ self.model = wan_model
+ self.lora_metadata = {}
+ # self.override_dict = {} # On CPU
+
+ def load_lora(self, lora_path, lora_name=None):
+ if lora_name is None:
+ lora_name = os.path.basename(lora_path).split(".")[0]
+
+ if lora_name in self.lora_metadata:
+ logger.info(f"LoRA {lora_name} already loaded, skipping...")
+ return lora_name
+
+ self.lora_metadata[lora_name] = {"path": lora_path}
+ logger.info(f"Registered LoRA metadata for: {lora_name} from {lora_path}")
+
+ return lora_name
+
+ def _load_lora_file(self, file_path, param_dtype):
+ with safe_open(file_path, framework="pt") as f:
+ tensor_dict = {key: f.get_tensor(key).to(param_dtype) for key in f.keys()}
+ return tensor_dict
+
+ def apply_lora(self, lora_name, alpha=1.0, param_dtype=torch.bfloat16, device='cpu'):
+ if lora_name not in self.lora_metadata:
+ logger.info(f"LoRA {lora_name} not found. Please load it first.")
+
+
+
+ lora_weights = self._load_lora_file(self.lora_metadata[lora_name]["path"], param_dtype)
+ # weight_dict = self.model.original_weight_dict
+ self._apply_lora_weights(lora_weights, alpha, device)
+ # self.model._init_weights(weight_dict)
+
+ logger.info(f"Applied LoRA: {lora_name} with alpha={alpha}")
+ return True
+
+ def get_parameter_by_name(self, model, param_name):
+ parts = param_name.split('.')
+ current = model
+ for part in parts:
+ if part.isdigit():
+ current = current[int(part)]
+ else:
+ current = getattr(current, part)
+ return current
+
+ @torch.no_grad()
+ def _apply_lora_weights(self, lora_weights, alpha, device):
+ lora_pairs = {}
+ prefix = "diffusion_model."
+
+ for key in lora_weights.keys():
+ if key.endswith("lora_down.weight") and key.startswith(prefix):
+ base_name = key[len(prefix) :].replace("lora_down.weight", "weight")
+ b_key = key.replace("lora_down.weight", "lora_up.weight")
+ if b_key in lora_weights:
+ lora_pairs[base_name] = (key, b_key)
+ elif key.endswith("diff_b") and key.startswith(prefix):
+ base_name = key[len(prefix) :].replace("diff_b", "bias")
+ lora_pairs[base_name] = (key)
+ elif key.endswith("diff") and key.startswith(prefix):
+ base_name = key[len(prefix) :].replace("diff", "weight")
+ lora_pairs[base_name] = (key)
+
+ applied_count = 0
+ for name in tqdm(lora_pairs.keys(), desc="Loading LoRA weights"):
+ param = self.get_parameter_by_name(self.model, name)
+ if device == 'cpu':
+ dtype = torch.float32
+ else:
+ dtype = param.dtype
+ if isinstance(lora_pairs[name], tuple):
+ name_lora_A, name_lora_B = lora_pairs[name]
+ lora_A = lora_weights[name_lora_A].to(device, dtype)
+ lora_B = lora_weights[name_lora_B].to(device, dtype)
+ delta = torch.matmul(lora_B, lora_A) * alpha
+ delta = delta.to(param.device, param.dtype)
+ param.add_(delta)
+ else:
+ name_lora = lora_pairs[name]
+ delta = lora_weights[name_lora].to(param.device, dtype)* alpha
+ delta = delta.to(param.device, param.dtype)
+ param.add_(delta)
+ applied_count += 1
+
+
+ logger.info(f"Applied {applied_count} LoRA weight adjustments")
+ if applied_count == 0:
+ logger.info(
+ "Warning: No LoRA weights were applied. Expected naming conventions: 'diffusion_model..lora_A.weight' and 'diffusion_model..lora_B.weight'. Please verify the LoRA weight file."
+ )
+
+
+ def list_loaded_loras(self):
+ return list(self.lora_metadata.keys())
+
+ def get_current_lora(self):
+ return self.model.current_lora
\ No newline at end of file
+ 
+ 
+