Video Color Grading via Look-Up Table Generation

Seunghyun Shin¹, Dongmin Shin², Jisu Shin¹ , Hae-Gon Jeon^{2 †}, Joon-Young Lee^{3 †},

¹GIST ²Yonsei University ³Adobe Research

ICCV 2025

 

📝 Introduction

We present a reference-based video color grading framework. Our key idea is to generate a look-up table (LUT) for color attribute alignment between reference scenes and input video via a diffusion model.

If you find Video Color Grading useful, please help ⭐ this repo, which is important to Open-Source projects. Thanks!

• [26/06/2025] 🎉🎉🎉 Video Color Grading is accepted by ICCV 2025.

🚀 Quick Start

Installation

1. Clone the repository:

git clone https://github.com/seunghyuns98/VideoColorGrading.git

2. Install dependencies:

• Directly Generate conda with our bash script:

source fast_env.sh

• Or Install manually(please refer to fast_env.sh)

3. Download the pretrained model weights from Google Drive and place them in the 'pretrained/' directory.

Inference

Run inference code using the provided example reference image and input video.
If you placed pretrained models in a directory other than 'pretrained/', make sure to update its path in configs/prompts/video_demo.yaml.

python video_demo.py \
--ref_path examples/reference1.jpg \
--input_path examples/video1.mp4 \
--save_path output/example1.mp4 

🏋️‍♂️ Train Your Own Model

Training consists of two steps:

1. Training GS-Extractor
2. Training L-Diffuser

Before training, make sure to update the config files with your environment:

video_folder : PATH To Your Video Dataset
lut_folder: PATH To Your LUT Dataset
step1_checkpoint_path: PATH To Your Pretrained Step1 Model
etc.

You can see config files at confgis folder

Furthermore, change your lut path in your dataloader

📁 Dataset Preparation

We use the Condensed Movie Dataset which consists of over 33,000 clips from 3,000 movies covering the salient parts of the films and has two-minutes running time for each clip in average.
100 LUT bases which are selected as distinctive LUTs from the $400$ LUTs of the Video Harmonization Dataset.
You can download them through: Google Drive
They are originally from: Condensed Movie Dataset & Video Harmonization Dataset
We recommend dividing the video into frames using the script below.
Please make sure to change the video_path variable to the location where your dataset is stored.

bash video2frame.sh

🔧 Training Phase 1

torchrun --nnodes=1 --nproc_per_node=8 train_step1.py --config configs/training/train_stage_1.yaml

🔧 Training Phase 2

torchrun --nnodes=1 --nproc_per_node=8 train_step2.py --config configs/training/train_stage_2.yaml

📊 Evaluation

You can evaluate performance by running:

python eval.py \
    --video_path <PATH_TO_YOUR_VIDEO_DATASET> \
    --lut_path <PATH_TO_YOUR_LUTs> \
    --save_path <PATH_TO_SAVE_RESULTS>

🤝 Contributing

• Welcome to open issues and pull requests.
• Welcome to optimize the inference speed and memory usage, e.g., through model quantization, distillation, or other acceleration techniques.

❤️ Acknowledgement

We have used codes from other great research work, including Animate-Anyone, GeometryCrafter. We sincerely thank the authors for their awesome work!

📜 Citation

If you find this work helpful, please consider citing:

@article{shin2025video,
  title={Video Color Grading via Look-Up Table Generation},
  author={Shin, Seunghyun and Shin, Dongmin and Shin, Jisu and Jeon, Hae-Gon and Lee, Joon-Young},
  journal={arXiv preprint arXiv:2508.00548},
  year={2025}
}