Research map / Vision-language and multimodal / Multimodal LLMs

White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

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 VLM-CC, which estimates illuminant for cross-camera color constancy by iteratively white-balancing raw images, rendering pseudo-sRGB, and using a LoRA-tuned Qwen2.5-VL to predict the residual color cast (red/green/blue) that drives angular updates of the illuminant.

Problem

Learning-based color constancy models overfit to the training camera's color response, so accuracy degrades on images from other cameras.

Approach

Reframe illuminant estimation as iterative perceptual feedback: white-balance with the current estimate, then have a VLM name the residual color cast (red, green, or blue), and rotate the illuminant estimate toward that axis until convergence, instead of regressing RGB directly.

Key results

• Leave-one-out on Gehler-Shi: mean angular error 1.52 vs 2.23 for CCMNet (best prior learning-based baseline).
• NUS-8 cross-sensor: mean 1.49 vs 1.71 for CCMNet; worst-25% 2.98 vs 3.62.
• Camera-agnostic NUS-8 to Gehler-Shi: mean 2.03 vs 2.38 for CCMNet; Gehler-Shi to NUS-8 mean 2.07 vs 2.17.

Surprising findings

• Initialization barely matters: removing the Gray-World initialization gives mean error 1.61 vs 1.52 with it.
• Predicting discrete color-cast labels outperformed having the VLM output numerical RGB values iteratively (1.52 vs 1.70 mean error).

Limitations

• Relies on a 7B VLM at inference, with per-image multi-step VLM calls, which is costly relative to regression networks.
• Mapping to a discrete red/green/blue cast label quantizes the update direction and relies on hand-designed step-angle schedules and convergence rules.
• Gains over CCMNet are not uniform on some metrics (e.g., best-25% and median on some splits are worse or tied).

Experimental setup

Datasets: Gehler-Shi NUS-8 Intel-TAU Cube+

Benchmarks: Gehler-Shi NUS-8 Cube+ Intel-TAU

Metrics: mean angular error median trimean best 25% worst 25%

Baselines: CCMNet GCC C5 SIIE DMCC FFCC Gray-World Shades-of-Gray

Builds on: Gray-World CCMNet C5 LoRA

Closest papers by meaning

• When Color Constancy Goes Wrong: Correcting Improperly White-Balanced Images (CVPR 2019 · similarity 0.90)
• GCC: Generative Color Constancy via Diffusing a Color Checker (CVPR 2025 · similarity 0.88)
• Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing (CVPR 2024 · similarity 0.86)
• ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance (CVPR 2025 · similarity 0.85)
• MegaCoin: Enhancing Medium-Grained Color Perception for Vision-Language Models (AAAI 2026 · similarity 0.84)
• HVDualformer: Histogram-Vision Dual Transformer for White Balance (AAAI 2025 · similarity 0.83)