Albedo Estimation via Latent Bridge Matching (CIC 2026)
Weights for Albedo Estimation via Latent Bridge Matching, by Carme Corbi, David Serrano-Lozano, Javier Vazquez-Corral and Maria Vanrell (Universitat Autònoma de Barcelona and Computer Vision Center), published at the Color and Imaging Conference 2026.
Code: github.com/CVC-Color/albedoLBM
Contents
| Sub-folder | Model | Task | Conditioning |
|---|---|---|---|
albedo/ |
LBM-AID | RGB → albedo | shading |
shading/ |
LBM-SID | RGB → shading | albedo |
Each sub-folder holds a config.yaml and a model.safetensors (5.0 GB,
bfloat16): a Stable Diffusion XL UNet (~2.5B parameters) used as the drift
network, plus the frozen SDXL VAE. The two models are used together as a
two-stage decomposition pipeline.
Usage
from PIL import Image
from albedo_lbm import IntrinsicDecomposer
pipeline = IntrinsicDecomposer.from_pretrained("davidserra9/albedo-lbm")
result = pipeline(Image.open("image.png"))
result.albedo.save("albedo.png")
result.shading.save("shading.png")
Install the code with pip install -e . from the
repository.
Training
Both models were trained for the paper on InteriorVerse (44K images) and
Hypersim (59K images), on 256×256 random crops, with AdamW at a learning rate of
4e-5. The drift network is initialized from Stable Diffusion XL; the VAE stays
frozen. The albedo model is trained with the pixel-space reconstruction loss that
enforces the image formation model I = A · S; the shading model is trained
without it, which we found gives better shading estimates.
Sampling uses four equally spaced timesteps at training time and, by default, a single bridge step at inference.
Limitations
- Trained on synthetic indoor data only, so real-world images (especially outdoor scenes, people and objects) can show color shifts.
- Transparent, metallic and strongly non-Lambertian surfaces remain a failure case.
- Quality degrades above roughly 2K resolution.
Citation
@inproceedings{corbi2026albedo,
title = {Albedo Estimation via Latent Bridge Matching},
author = {Corbi, Carme and Serrano-Lozano, David and Vazquez-Corral, Javier and Vanrell, Maria},
booktitle = {Color and Imaging Conference (CIC)},
year = {2026}
}
License
CC BY-NC 4.0, following the license of the upstream LBM code these models build on. Use is also subject to the terms of the InteriorVerse and Hypersim datasets and of the Stable Diffusion XL backbone.