GenColor: Generative and Expressive Color Enhancement with Pixel-Perfect Texture Preservation

Official PyTorch Implementation of NeurIPS 2025 (Spotlight) Paper "GenColor: Generative and Expressive Color Enhancement with Pixel-Perfect Texture Preservation"

Commercial models like NanoBanana Pro also have texture distortion issues where text changes from "NO STANDING" to "ON STATING", while GenColor's fusion model achieves pixel-perfect preservation.

Environment

conda create -n gencolor python=3.10 -y
conda activate gencolor
pip install -r requirements.txt

Inference

Model Checkpoint

Please download the model checkpoint from Google Drive, then run the demo script.

Please note that due to commercialization and copyright considerations, only stage 2 texture preservation network is released. Please stay tuned for the release of the stage 1 color enhancement model.

Single image enhancement demo script

python demo.py --input example/tree.png --output example/tree_result.png

Fusion demo notebook

Run the fusion demo notebook to combine the stage 1 color enhancement with the stage 2 texture preservation network:

demo_fusion.ipynb

ComfyUI Integration

Setup

cp -r /path/to/GenColor/comfyui /path/to/ComfyUI/custom_nodes/ComfyUI-GenColor-Fusion
mkdir -p /path/to/ComfyUI/models/gencolor
cp /path/to/GenColor/ckpt/GenColor_fusion/base.pth /path/to/ComfyUI/models/gencolor/

Run

cd /path/to/ComfyUI
python main.py --listen 0.0.0.0 --port 8188

Usage

1. Right-click → Add Node → GenColor → GenColor Fusion
2. Add 2x LoadImage → connect to input_image and color_reference
3. Add PreviewImage → connect to output
4. Click Queue Prompt (or Ctrl+Enter / Cmd+Enter)

Training

ARTISAN-1.2M Dataset

Before training, you need to download the ARTISAN-1.2M dataset and generate captions for your images.

Note: Due to commercialization and copyright considerations, we release a 240P preview version of the dataset.

We support two models for caption generation: BLIP and Qwen3-VL.

Using BLIP

python training/stage1/image_caption.py \
--image_dir="data/ARTISAN-1.2M/images" \
--output_json="data/ARTISAN-1.2M/captions.json" \
--model_type blip \
--batch_size 8

Using Qwen3-VL

python training/stage1/image_caption.py \
--image_dir="data/ARTISAN-1.2M/images" \
--output_json="data/ARTISAN-1.2M/captions.json" \
--model_type qwen \
--batch_size 1 \
--device_map auto

Then train the stage1 model:

CUDA_VISIBLE_DEVICES=0 python training/stage1/train_controlnet.py \
--pretrained_model_name_or_path="stabilityai/stable-diffusion-2-1-base" \
--img_dir="data/ARTISAN-1.2M/images" \
--prompt_json_path="data/ARTISAN-1.2M/captions.json" \
--output_dir="ckpt/GenColor_SD21" \
--resolution=512 \
--learning_rate=1e-5 \
--train_batch_size=16 \
--num_train_epochs=100000 \
--tracker_project_name="GenColor_SD21-stage1" \
--checkpointing_steps=10000 \
--report_to wandb \
--validation_steps=500 \
--lr_warmup_steps=0 \
--mixed_precision="bf16" \
--dataloader_num_workers=8 \
--resume_from_checkpoint='latest'

Acknowledgments

This code repository is partially borrowed from Harmonizer and SwinIR.

Citation

You can cite it as follows:

@inproceedings{donggencolor,
  title={GenColor: Generative and Expressive Color Enhancement with Pixel-Perfect Texture Preservation},
  author={Dong, Yi and Wang, Yuxi and Lin, Xianhui and Ouyang, Wenqi and Shen, Zhiqi and Ren, Peiran and Fan, Ruoxi and Lau, Rynson WH},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
}