Yuhang He1, Kai Zhang1,†, Xiaoming Li1, Du Chen2, Jian Yang1
1Nanjing University, China
2VIVO BlueImage Lab, China
†Corresponding author
ECCV 2026
ColorFM image color transfer (left) and video color transfer (right).
🚀 Explore more application results: Image Color Transfer | Video Color Transfer
Table of Contents
ColorFM is an optimization-to-learning framework for accurate and semantically consistent color transfer. It connects instance-specific optimization with efficient feed-forward inference through two complementary variants: ColorFM-O and ColorFM-L.
| Method | Type | Demo |
|---|---|---|
| ColorFM-O | Optimization-based | Try online |
| ColorFM-L | Learning-based | Try online |
ColorFM formulates color transfer as transporting pixel distributions along velocity fields via Flow Matching. ColorFM-O optimizes an instance-specific velocity field with semantic guidance, while ColorFM-L learns from the generated pairs to provide efficient feed-forward inference.
Overview of the ColorFM-O and ColorFM-L frameworks.
The following table compares ColorFM with existing color transfer methods in terms of similarity, Lipschitz constant, and inference time. All results are evaluated at an image resolution of 512 x 512.
Create an environment and install the dependencies from the repository root:
conda create -n ColorFM python=3.10 -y conda activate ColorFM pip install torch torchvision pip install -r requirements.txt
Install the PyTorch build that matches your CUDA version when GPU acceleration is required. xFormers can optionally accelerate ColorFM-L on supported CUDA environments.
Download the pretrained ColorFM-L checkpoint from Hugging Face and place it under the repository-level checkpoints folder. ColorFM-O does not require a pretrained checkpoint.
[2026-10-12] Pre-generated Stage 1 training pairs are available on Hugging Face. They can be downloaded and used directly to train ColorFM-L in Stage 2; see the Training guide for download and configuration instructions !!!
| Guide | Description |
|---|---|
| Testing | Run the ColorFM-O and ColorFM-L image or video WebUIs. |
| Training | Use the pre-generated Stage 1 dataset or generate pairs yourself, train ColorFM-L, and run evaluation. |
This project builds upon the open-source implementations of DINOv2 by Meta AI.
The original images in dataset ColorFM were sourced from Unsplash and are used solely for non-commercial academic research. Copyright remains with the respective photographers; the images are not relicensed under this project's license. For rights-related inquiries or removal requests, please contact us.
If you find this work useful, please cite:
@inproceedings{he2026colorfm, title={ColorFM: An Optimization-to-Learning Framework for Color Transfer via Flow Matching}, author={He, Yuhang and Zhang, Kai and Li, Xiaoming and Chen, Du and Yang, Jian}, booktitle={European Conference on Computer Vision}, year={2026} }
This project is released under the Apache License 2.0.