Revisiting Image Fusion for Multi-Illuminant White-Balance Correction.
In ICCV, 2025.

David Serrano-Lozano^1,2, Aditya Arora^3,4,5, Luis Herranz⁶,
Konstantinos G. Derpanis^3,4, Michael S. Brown³ and Javier Vazquez-Corral<sup>1,2

1</sup>Computer Vision Center, ²Universitat Autònoma de Barcelona, ³York University,
⁴Vector Institute, ⁵TU Darmstadt and ⁶Universidad Politécnica de Madrid

---

TODOs (In Progress)

✅ Upload models for both splits of our dataset and RenderedWB.

✅ Upload the dataset.

Method

We propose a lightweight Transformer block to blend five white balance (WB) presets and produce a white-balanced image. Our model contains only 7.9K parameters.

Data

While the original LSMI dataset was designed for illumination estimation from RAW images, we repurpose it to compute ground-truth white-balanced images from multi-illuminant scenes. Please check the original dataset for data acquisition and other details.

To download the dataset, please check the following website:

Getting Started

Clone the repository and install the required dependencies.

Train and Inference

Pre-trained models are available in the weights folder. Each checkpoint is only ~38KB.

Both training and inference use the config.yaml file to specify all parameters and configurations. Please adapt it accordingly.

To run inference on our dataset:

python inference.py

To train a model:

python train.py