SAGE-Color

Reference-Based Color Transfer

SAGE-Color

Semantic Appearance Grounding for Reference-Based Color Transfer

SAGE-Color transfers palette, tone, contrast, and region-level appearance from a reference image while preserving the content image's geometry, identity, layout, and fine structure.

GitHub Weights arXiv coming soon

Visual Showcase

Content structure stays fixed. Reference appearance moves.

Each example alternates from the original content image to the SAGE-Color result while the reference image jumps into the lower left corner as the appearance cue.

Street content image before color transfer. Street SAGE-Color result. Evening street reference image.
ContentSAGE-Color
Warm interior content image before color transfer. Warm interior SAGE-Color result. Window portrait reference image.
ContentSAGE-Color
Golden garden content image before color transfer. Golden garden SAGE-Color result. Bright garden reference image.
ContentSAGE-Color
Night lantern content image before color transfer. Night lantern SAGE-Color result. Lantern portrait reference image.
ContentSAGE-Color
Street portrait content image before color transfer. Street portrait SAGE-Color result. Green umbrella reference image.
ContentSAGE-Color
Village landscape content image before color transfer. Village landscape SAGE-Color result. Stormy mountain reference image.
ContentSAGE-Color
Waterfront content image before color transfer. Waterfront SAGE-Color result. Evening city reference image.
ContentSAGE-Color
Blue waterfront content image before color transfer. Blue waterfront SAGE-Color result. Clear sky reference image.
ContentSAGE-Color
Traditional portrait content image before color transfer. Traditional portrait SAGE-Color result. Blue umbrella reference image.
ContentSAGE-Color
Waterfront bicycle content image before color transfer. Waterfront bicycle SAGE-Color result. Bright scooter reference image.
ContentSAGE-Color

Abstract

Reference color as semantic evidence

SAGE-Color treats the reference image as chromatic evidence rather than a spatial template. The content image remains responsible for pose, object boundaries, and layout, while reference appearance is grounded to compatible content regions.

01

Semantic Color Gallery

Global, regional, and local reference cues are organized for semantic appearance grounding.

02

Dense Content Path

Content latents keep the generated image anchored to the original geometry and fine detail.

03

Intrinsic Preservation Field

Color-free structure cues reduce unsafe residual edits in identity-sensitive regions.

Sliding Gallery

Scene and style cases

Each slide keeps the content and reference inputs in fixed compact cells, then gives the SAGE-Color result the largest frame so the transferred scene style is easy to compare.

Qualitative Examples

Cross-scene color transfer

Night lantern content image before color transfer.
Content
Lantern portrait reference image.
Reference
Night lantern result with warm festival tone.
SAGE-Color result

Lantern night atmosphere

The lantern reference transfers warm night illumination while preserving the waterline, pavilion lights, figure scale, and composition.

Forest portrait content image before color transfer.
Content
Green forest reference image.
Reference
Forest portrait result with green reference style.
SAGE-Color result

Dense forest green transfer

The reference pulls the result toward layered forest greens while preserving the standing pose, tree placement, and depth cues.

Lake portrait content image before color transfer.
Content
Turquoise lake reference image.
Reference
Lake portrait result with turquoise reference color.
SAGE-Color result

Turquoise water transfer

The result borrows the clear lake palette while preserving the original pose, clothing silhouette, and shoreline placement.

Indoor portrait content image before color transfer.
Content
Violet sunset reference image.
Reference
Indoor portrait result with violet sunset tone.
SAGE-Color result

Violet sunset interior

Low-saturation violet light changes the room atmosphere while retaining the subject, hands, clothing, and wall structure.

Beach portrait content image before color transfer.
Content
Blue beach reference image.
Reference
Beach portrait result with deep blue reference style.
SAGE-Color result

Blue coastal stage lighting

A saturated beach reference cools the scene into a deep blue style while the body pose and open shoreline composition stay stable.

Stream content image before color transfer.
Content
Waterfall reference image.
Reference
Stream result with waterfall reference style.
SAGE-Color result

Green stream and water highlights

The water reference adds darker greens and crisp highlights while the content rocks, stream direction, and subject position stay stable.

Riverside portrait content image before color transfer.
Content
Shaded portrait reference image.
Reference
Riverside portrait result with shaded reference style.
SAGE-Color result

Soft riverside shade

The reference darkens local contrast and skin-side lighting while the stone steps, water edge, and seated pose remain content-driven.

Blossom portrait content image before color transfer.
Content
Sunset palm reference image.
Reference
Blossom portrait result with warm sunset tone.
SAGE-Color result

Warm blossom atmosphere

The sunset reference shifts the blossom portrait from bright spring pinks into a warmer brown tone while preserving the subject pose, branches, and path structure.

Architecture

Semantic appearance grounding

SAGE-Color separates dense content conditioning, reference appearance grounding, and content-only structure preservation inside an MMDiT flow editor.

SAGE-Color architecture overview showing the content path, Semantic Color Gallery, and Intrinsic Preservation Field.
Dense content conditioning anchors layout, the Semantic Color Gallery supplies reference appearance, and the Intrinsic Preservation Field protects structure-sensitive regions.

Evaluation

Colorist-Bench-1K results

SAGE-Color improves reference color fidelity while keeping content-compatible reconstruction.

6.47DeltaE00 down
23.99PSNR up
0.984DINO-T up
95.33FG-CES up
Qualitative comparison for reference-based color transfer.
Paper qualitative comparison. SAGE-Color follows reference appearance while preserving content structure.

Release

Run SAGE-Color

The GitHub repository contains inference, training, environment setup, and reproducibility scripts. The released checkpoints are hosted on Hugging Face so the code repository stays lightweight.

Open repository 
git clone https://github.com/chenxib/sage-color.git
cd sage-color
conda env create -f environment.yml
conda activate sage-color
bash scripts/bootstrap_external_diffusers.sh
pip install -r requirements.txt
bash scripts/download_weights.sh
bash scripts/download_required_models.sh

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