Skip to content

README.md

Latest commit

 

History

History
119 lines (93 loc) · 3.12 KB

README.md

File metadata and controls

119 lines (93 loc) · 3.12 KB

JarvisArt Data Scripts

Image editing training data generation pipeline.

Dependencies

pip install transformers accelerate qwen-vl-utils vllm tqdm openai google-generativeai pillow

Modules

data_scripts/
├── image_pairs_xmp_generation/    # Generate image pairs and XMP configs
├── instructions_generation/       # Generate user instructions  
├── cot_generation/                # Generate expert reasoning
└── format_conversion/             # Convert to training format

Usage

1. Generate Image Pairs and XMP Configs

cd image_pairs_xmp_generation/
bash run_recommendation.sh

2. Generate User Instructions

cd instructions_generation/
python vllm_api_instructions_gen.py \
    --base_path /path/to/data \
    --start_id 0 \
    --end_id 100 \
    --key YOUR_GEMINI_API_KEY

Outputs: user_want.txt, user_want_2.txt, user_want_4.txt, user_want_7.txt

3. Generate Expert Chain-of-Thought

cd cot_generation/
python cot_initial_gen.py \
    --base_path /path/to/data \
    --start_id 0 \
    --end_id 100 \
    --key YOUR_GOOGLE_API_KEY

python cot_refined_gen.py \
    --base_path /path/to/data \
    --start_id 0 \
    --end_id 100 \
    --key YOUR_GOOGLE_API_KEY

Outputs: original_expert_cot.txt, revised160_expert_cot.txt

4. Convert to Training Format

cd format_conversion/
python format_converter.py \
    --workers 100 \
    --tasks global_cot local_cot \
    --global_data_path /path/to/global/data \
    --local_data_path /path/to/local/data \
    --output_dir ./output \
    --before_image before.jpg \
    --roc_file config.lua \
    --user_intent_file user_want.txt \
    --cot_file revised160_expert_cot.txt

Data Structure

Each image folder contains:

image_folder/
├── before.jpg                    # Original image
├── processed.jpg                 # Edited image
├── config.lua                    # Processing config
├── config.xmp                    # XMP metadata
├── user_want*.txt               # User instructions (4 variants)
├── original_expert_cot.txt      # Initial expert reasoning
└── revised160_expert_cot.txt    # Refined expert reasoning

Parameters

  • --base_path: Path to image folders
  • --start_id: Start processing from this ID (0-based)
  • --end_id: Stop at this ID (-1 for all remaining)
  • --key: API key for Gemini/Google services
  • --workers: Parallel threads for format conversion

Complete Workflow

# 1. Generate configs
cd image_pairs_xmp_generation/ && bash run_recommendation.sh

# 2. Generate instructions
cd ../instructions_generation/
python vllm_api_instructions_gen.py --base_path ./data --start_id 0 --end_id 1000 --key YOUR_API_KEY

# 3. Generate reasoning
cd ../cot_generation/
python cot_initial_gen.py --base_path ./data --start_id 0 --end_id 1000 --key YOUR_API_KEY
python cot_refined_gen.py --base_path ./data --start_id 0 --end_id 1000 --key YOUR_API_KEY

# 4. Convert format
cd ../format_conversion/
python format_converter.py --workers 50 --tasks global_cot local_cot \
    --global_data_path ./data --output_dir ./output