Research map / Image generation / Editing and style transfer
ChromaFusionNet (CFNet): Natural Fusion of Fine-Grained Color Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CFNet, which reframes fine-grained color fusion of multiple local color edits as image color inpainting in CIELab ab-channels, using a ViT encoder, pixel-shuffle decoder, and RRDB refinement.
Problem
Direct composition of local color edits creates spatial color inconsistencies at mask boundaries; existing color style transfer, harmonization, and inpainting methods either alter non-boundary colors or texture.
Approach
Treats boundary color inconsistencies as holes in the ab color channels and fills them via a ViT-based encoder with pixel-shuffle decoder and an RRDB refinement module, trained self-supervised on ImageNet with Mask2Former-derived edge masks.
Key results
- On ImageNet, CFNet achieves W1 of 1.66 versus 3.15 for the best prior baseline CAP-VSTNet and 5.44 for Harmonizer (lower is better).
- On COCO, CFNet achieves W1 of 0.21 versus 1.67 for CAP-VSTNet and 1.98 for StyA2K.
- Refinement module improves COCO B-PSNR from 30.75 (coarse) to 32.05 and ImageNet PSNR from 34.40 to 34.96.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CFNet instead of CAP-VSTNet | Color fusion · ImageNet test set | W1: 3.15 -> 1.66 | Helped |
| use CFNet instead of CAP-VSTNet | Color fusion · COCO test set | W1: 1.67 -> 0.21 | Helped |
| use CFNet instead of Color Matcher | Color fusion · ImageNet test set | IP@0.9651 (%): 3.50 -> 4.14 | Hurt |
| use CFNet instead of Color Matcher | Color fusion · COCO test set | IP@0.9651 (%): 0.91 -> 1.70 | Hurt |
| add the refinement module (ablation) | CFNet · ImageNet test set | PSNR: 34.40 -> 34.96 | Helped |
| add the refinement module (ablation) | CFNet · ImageNet test set | B-PSNR: 25.61 -> 26.57 | Helped |
| add the refinement module (ablation) | CFNet · ImageNet test set | SSIM: 0.97 -> 0.97 | No clear effect |
| add the refinement module (ablation) | CFNet · ImageNet test set | ∆E: 0.68 -> 0.65 | Helped |
| add the refinement module (ablation) | CFNet · COCO test set | PSNR: 35.67 -> 36.12 | Helped |
| add the refinement module (ablation) | CFNet · COCO test set | B-PSNR: 30.75 -> 32.05 | Helped |
| add the refinement module (ablation) | CFNet · COCO test set | SSIM: 0.97 -> 0.97 | No clear effect |
| add the refinement module (ablation) | CFNet · COCO test set | ∆E: 0.74 -> 0.73 | Helped |
Surprising findings
- Direct composition has lower IP than several harmonization/style-transfer baselines, and CFNet's IP is only comparable to Color Matcher rather than best.
Limitations
- Struggles with transparent regions, showing color discrepancies.
- Performance degrades when the proportion of masked colors to predict is too high.
- Assumes the lightness (L) channel remains unchanged.
Experimental setup
Datasets: ImageNet COCO
Benchmarks: ImageNet test set COCO test set
Metrics: PSNR SSIM delta E B-PSNR W1 color deviation inharmonious region proportion (IP)
Baselines: Color Matcher Harmonizer S2CRNet CAP-VSTNet MCCNet StyA2K UniColor DeepLPF AdaInt
Builds on: Vision Transformer ESRGAN (RRDB) Mask2Former Canny edge detection
Closest papers by meaning
- Training-Free Text-Guided Color Editing with Multi-Modal Diffusion Transformer (ICLR 2026 · similarity 0.86)
- Free-Lunch Color-Texture Disentanglement for Stylized Image Generation (NeurIPS 2025 · similarity 0.85)
- Progressive Painterly Image Harmonization from Low-Level Styles to High-Level Styles (AAAI 2024 · similarity 0.84)
- HairDiffusion: Vivid Multi-Colored Hair Editing via Latent Diffusion (NeurIPS 2024 · similarity 0.84)
- Color Shift Estimation-and-Correction for Image Enhancement (CVPR 2024 · similarity 0.84)
- Automatic Controllable Colorization via Imagination (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.