SAGE-Color
Semantic Appearance Grounding for Reference-Based Color Transfer
SAGE-Color is a reference-based color transfer model. Given a content image and a reference image, it transfers the reference image's palette, tone, contrast, and region-level color appearance while preserving the content image's geometry, identity, layout, and fine structure.
The central design goal is simple: the reference image provides chromatic evidence, not a spatial template. This makes SAGE-Color useful for color grading, appearance matching, and reference-guided recoloring where object boundaries and content identity should remain stable.
Highlights
- Reference-based color transfer without copying reference layout.
- Content image remains the spatial authority.
- Semantic Color Gallery organizes global, regional, and local chromatic cues.
- Intrinsic Preservation Field protects structure-sensitive content regions.
- Includes the final inference checkpoint and the first-stage grounding checkpoint for continued training.
Files
| File | Purpose |
|---|---|
checkpoints/sage-color-final.pt |
Final checkpoint for normal inference. |
checkpoints/sage-color-grounding.pt |
First-stage color-grounding checkpoint for continued final-stage training. |
Quick Start
Clone the code repository and install the environment:
git clone https://github.com/chenxib/sage-color.git
cd sage-color
conda env create -f environment.yml
conda activate zhuise-color-edit
bash scripts/bootstrap_external_diffusers.sh
pip install -r requirements.txt
Download the SAGE-Color checkpoints into the code repository:
pip install -U huggingface_hub
huggingface-cli download chenxib/sage-color \
checkpoints/sage-color-final.pt \
checkpoints/sage-color-grounding.pt \
--local-dir .
Download the required external backbones and feature extractors:
bash scripts/download_required_models.sh
Run inference:
CUDA_VISIBLE_DEVICES=0 \
CONTENT_IMAGE=path/to/content.png \
REFERENCE_IMAGE=path/to/reference.png \
OUTPUT_IMAGE=outputs/sage-color/sample.png \
bash scripts/final_model/bash/infer.sh
Code And Project Page
- Code: https://github.com/chenxib/sage-color
- Project page: https://chenxib.github.io/sage-color/
Intended Use
SAGE-Color is intended for research and creative workflows that need reference-guided color appearance transfer while preserving the source content structure. It is not intended to copy identities, layouts, or textures from the reference image.
License And Dependencies
This checkpoint release follows the repository license. The full system depends on third-party model licenses, including Stable Diffusion 3.5 Medium and the feature extractors documented in the GitHub repository. Users are responsible for complying with the licenses of those dependencies.
Citation
The arXiv identifier and final citation will be added after public release.
Model tree for chenxib/sage-color
Base model
stabilityai/stable-diffusion-3.5-medium