A new AI method separates an object's true color from its lighting, using a technique borrowed from generative modeling to make the process cheaper and more reliable.
The paper tackles albedo estimation - figuring out a surface's base color without shadows, highlights, or other lighting artifacts - a task known as intrinsic image decomposition. The researchers apply latent bridge matching (LBM) to the problem, adding a pixel reconstruction loss to keep predictions physically plausible, relying on LBM's inherently cheap inference for speed, and conditioning the model on estimated shading so it generalizes across different image types. In an extended version of the work, they also fed the predicted albedo back into the shading estimate, which improved reconstruction fidelity further. The team then benchmarked its best model against current state-of-the-art intrinsic decomposition methods across five real and synthetic datasets.
This is unglamorous plumbing, but it matters: accurate albedo estimates feed into photo editing, AR object placement, and material recognition, and past generative approaches to the problem have tended to be slow, physically inconsistent, or bad at generalizing to new images. This paper claims to address all three problems in one architecture, which would be notable if it holds up.
That "if it holds up" is doing a lot of work - the improvements are measured on the authors' own choice of five datasets, and intrinsic image decomposition has a long history of methods that look great on curated benchmarks and less great on the messy photos people actually take.