Research map / Image generation / Image restoration
GCC: Generative Color Constancy via Diffusing a Color Checker
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.
Fine-tunes Stable Diffusion 2 inpainting to paint a neutral color checker into an image in a single deterministic step, then reads the illuminant from the generated patches for cross-camera color constancy.
Problem
Learning-based color constancy models often fail to generalize across camera sensors because of differing spectral sensitivities, and diffusion-based generation is stochastic, which hurts consistency.
Approach
Inpaint a color checker reflecting scene illumination with a fine-tuned diffusion inpainting model run at a fixed timestep in one deterministic step, using Laplacian high-frequency latent input and masked color jittering to handle imprecise checker annotations.
Key results
- Cross-dataset (NUS-8 to Gehler) mean angular error 2.35 degrees, outperforming C4 (2.28 is slightly better; best among listed methods is C4 at 2.28), and on Gehler-to-NUS-8 mean 2.38.
- Leave-one-out on NUS-8 achieves mean 2.03 degrees, the best among compared methods (C5 2.54, SIIE 2.05).
- Single-step inference takes 0.18 s per 512x512 image on an RTX 4090, versus 17.98 s for 25-step SDXL inpainting with 10-ensemble.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GCC instead of C4 SqueezeNet-FC4 | Cross-camera color constancy · NUS-8 → Gehler | mean angular error (°): 2.73 -> 2.35 | Helped |
| Use GCC instead of C4 SqueezeNet-FC4 | Cross-camera color constancy · Gehler → NUS-8 | mean angular error (°): 2.28 -> 2.38 | Hurt |
| Use GCC instead of C5 | Color constancy · Gehler leave-one-out | mean angular error (°): 2.50 -> 2.80 | Hurt |
| Use GCC instead of CLCC | Color constancy · NUS-8 three-fold cross-validation | mean angular error (°): 1.84 -> 2.10 | Hurt |
| Use GCC instead of C4 SqueezeNet-FC4 | Color constancy · Gehler three-fold cross-validation | mean angular error (°): 1.35 -> 1.91 | Hurt |
| Use GCC instead of the best listed baseline | Spatially varying illumination estimation · LSMI Galaxy single | mean angular error (°): 2.85 -> 2.05 | Helped |
| Use GCC instead of the best listed baseline | Spatially varying illumination estimation · LSMI Galaxy multi | mean angular error (°): 3.13 -> 3.44 | Hurt |
| Use GCC instead of the best listed baseline | Spatially varying illumination estimation · LSMI Nikon multi | mean angular error (°): 2.36 -> 3.58 | Hurt |
| Use single-step GCC instead of SDEdit with SDXL inpainting | Color constancy · Gehler | mean angular error (°): 4.47 -> 2.35 | Helped |
| Use single-step inference instead of SDEdit | Color constancy · Gehler | inference time (s): 17.98 -> 0.18 | Helped |
| Add Laplacian decomposition (ablation) | GCC with masked color jittering and checker inpainting · Gehler | mean angular error (°): 3.71 -> 2.35 | Helped |
| Add masked color jittering (ablation) | GCC with Laplacian decomposition and checker inpainting · Gehler | mean angular error (°): 3.52 -> 2.35 | Helped |
Surprising findings
- Direct RGB prediction with the diffusion model underperforms color checker inpainting, indicating the checker reference matters.
- Noisy stochastic SDXL LoRA fine-tuning with ensembling underperforms the deterministic single-step model.
Limitations
- Worst-case cross-validation results on Gehler (3.46 worst-25%) and mean on three-fold Gehler (1.91) are weaker than C4 and CLCC.
- Multi-illuminant LSMI results are mixed, with Galaxy multi-illuminant error (3.44) worse than LSMI-H and LSMI-U.
Experimental setup
Datasets: NUS-8 dataset Gehler (re-processed Color Checker) dataset LSMI dataset
Benchmarks: NUS-8 Gehler LSMI
Metrics: mean angular error median angular error tri-mean best 25% worst 25%
Baselines: Gray-World Shades-of-Gray General Gray-World Gray Index FFCC FC4 C4 SIIE CLCC C5 SDXL inpainting (SDEdit) LSMI-H
Builds on: Stable Diffusion inpainting Garcia et al. deterministic fine-tuning DiffusionLight FC4
Closest papers by meaning
- White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation (CVPR 2026 · similarity 0.88)
- Color Alignment in Diffusion (CVPR 2025 · similarity 0.87)
- LuxDiT: Lighting Estimation with Video Diffusion Transformer (NeurIPS 2025 · similarity 0.85)
- Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing (CVPR 2024 · similarity 0.85)
- DiffusionLight: Light Probes for Free by Painting a Chrome Ball (CVPR 2024 · similarity 0.84)
- Efficient Diffusion as Low Light Enhancer (CVPR 2025 · 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.