Research map / Image generation / Image restoration
HVDualformer: Histogram-Vision Dual Transformer for White Balance
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 HVDualformer, a lightweight histogram-vision dual transformer that predicts corrected color-histogram features and uses histogram matching to adjust image features for end-to-end sRGB white balance correction.
Problem
Correcting color temperature shifts and color casts in sRGB images to achieve white balance, where prior methods often ignore global color temperature information.
Approach
Histoformer learns corrected RGB histogram features from the input color histogram; a Histogram-Specified Feature Transformation (HSFT) module uses histogram matching to align image features to these corrected histograms, and a Visformer then produces the white-balanced image from the concatenated original and transformed features.
Key results
- On Rendered WB Set1-Test, HVDualformer+ reaches mean MSE 24.09 and mean ΔE 2000 2.31, versus 77.49 MSE for KNN and 78.89 MSE for WBFlow.
- On the Rendered Cube+ Dataset, HVDualformer+ achieves the best mean MSE (68.63) and mean ΔE 2000 (4.09), while SWBNET has the best mean MAE (3.15 vs 3.22).
- HVDualformer has 2.92M parameters (11.0 MB) versus SWBNET at 258.8 MB, and ablation shows the full design (68.6 MSE) beats Histoformer-only (168.8) and Visformer-only (115.1) on Cube+.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HVDualformer+ instead of WBFlow | White balance correction · Rendered WB Dataset Set1-Test | MSE: 78.89 -> 24.09 | Helped |
| Use HVDualformer+ instead of WBFlow | White balance correction · Rendered WB Dataset Set2 | MSE: 117.60 -> 110.65 | Helped |
| Use HVDualformer+ instead of SWBNET | White balance correction · Rendered Cube+ Dataset | MSE: 74.35 -> 68.63 | Helped |
| Use Histoformer alone instead of the complete HVDualformer (ablation) | HVDualformer · Rendered Cube+ Dataset | MSE: 68.6 -> 168.8 | Hurt |
| Use Histoformer alone instead of the complete HVDualformer (ablation) | HVDualformer · Rendered Cube+ Dataset | MAE: 3.22° -> 3.84° | Hurt |
| Use Histoformer alone instead of the complete HVDualformer (ablation) | HVDualformer · Rendered Cube+ Dataset | ΔE 2000: 4.08 -> 5.76 | Hurt |
| Use Visformer alone instead of the complete HVDualformer (ablation) | HVDualformer · Rendered Cube+ Dataset | MSE: 68.6 -> 115.1 | Hurt |
| Use Visformer alone instead of the complete HVDualformer (ablation) | HVDualformer · Rendered Cube+ Dataset | MAE: 3.22° -> 4.08° | Hurt |
| Use Visformer alone instead of the complete HVDualformer (ablation) | HVDualformer · Rendered Cube+ Dataset | ΔE 2000: 4.08 -> 4.83 | Hurt |
| Remove HSFT (ablation) | HVDualformer · Rendered Cube+ Dataset | MSE: 68.6 -> 138.1 | Hurt |
| Remove HSFT (ablation) | HVDualformer · Rendered Cube+ Dataset | MAE: 3.22° -> 3.59° | Hurt |
| Remove HSFT (ablation) | HVDualformer · Rendered Cube+ Dataset | ΔE 2000: 4.08 -> 4.81 | Hurt |
Surprising findings
- Histoformer alone performs only average, and Visformer alone outperforms Histoformer, yet the cascade without HSFT (Case C) is worse than Visformer alone in MSE.
Limitations
- Mixed-WB and multi-illuminant handling are only mentioned as available on GitHub, not evaluated in the paper.
- SWBNET was retrained by the authors for Set1-Test and Set2, so those comparisons depend on their reimplementation.
- On Set1-Test and Set2 results are only generally favorable rather than uniformly best; HVDualformer does not win every metric on Cube+.
Experimental setup
Datasets: Rendered WB Dataset Set1 Rendered WB Dataset Set2 Rendered Cube+ Dataset
Benchmarks: Rendered WB Dataset Set1-Test Rendered WB Dataset Set2 Rendered Cube+ Dataset
Metrics: MSE MAE ΔE 2000 Bhattacharyya coefficient
Baselines: KNN DEEP-WB Mixed-WB WBFlow SWBNET
Builds on: Deep WB Restormer histogram matching
Closest papers by meaning
- ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance (CVPR 2025 · similarity 0.86)
- White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation (CVPR 2026 · similarity 0.83)
- EnColor: Improving Visual Accessibility with a Deep Encoder-Decoder Image Corrector for Color Vision Deficient Individuals (AAAI 2024 · similarity 0.83)
- Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing (CVPR 2024 · similarity 0.83)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.81)
- LuxDiT: Lighting Estimation with Video Diffusion Transformer (NeurIPS 2025 · similarity 0.81)
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.