Question about uniworldv2

[wyhlovecpp]

Hi, I have a question about UniWorld‑V2.
I noticed the paper “Uniworld-V2: Reinforce Image Editing with Diffusion Negative-aware Finetuning and MLLM Implicit Feedback” mentions they have released models called UniWorld‑Qwen‑Image‑Edit‑2509 and UniWorld‑FLUX.1‑Kontext‑Dev.

My question is:

For the best results reported in the paper (UniWorld-V2), is the backbone or pre-training weight based on Qwen-Image-Edit (i.e., you used Qwen-Edit as the base model for UniWorld-V2)?

When will the full UniWorld-V2 model weights be released?

If I want to train such a large model (20B + 7B) on 1024×1024 image-editing data pairs, using a single node of NVIDIA A100 ( 80 GB variant) — would that GPU memory be sufficient, or would I need a distributed multi-node setup or special memory techniques (e.g., full-shard data parallelism, gradient checkpointing, etc.)?

Thanks in advance for your clarification.

[qqingzheng]

UniWorld-V2 is a commercial model and may not be released.

It’s sufficient to shard the text encoder, enable gradient checkpointing, and train only at 512×512 resolution. We have not conducted training at 1024×1024.

However, if you plan to train at 1024×1024, you might need to shard the DiT as well.