A new AI system can brighten a dark, noisy photo in a single pass, instead of the dozens of steps most restoration models need to get there.
Researchers built a system called Consist-Retinex that splits enhancement into two problems, following an old idea from color science called Retinex: separate a photo into "reflectance" (the actual color and texture of objects) and "illumination" (how much light is falling on them), then fix each separately. A neural network first decomposes the image into those two maps. Two other networks, called consistency models, then learn to restore each map in one step rather than the many steps typical of diffusion-style generative models. The trick is a training method that concentrates supervision on the noisiest, hardest-to-predict part of that one-step process, which is exactly where earlier one-step attempts fell apart. In tests, the method posted the best scores on the VE-LOL-L low-light benchmark among the one-step methods it was compared against, and held its own on the older LOL benchmark, while needing less training and sampling compute than the alternatives.
The interesting part isn't the low-light photos themselves, it's the training fix. Diffusion-based generative models are usually too slow for anything real-time, like a phone camera or a security feed, because they need many denoising passes. Squeezing that down to one step normally tanks quality, because standard training barely supervises the exact noise level a one-step model has to guess from. This paper's contribution is showing why that gap matters and how to patch it.
Worth noting: "best among compared methods" and "competitive" are doing a lot of work here. This is a benchmark paper, not a shipped product, and one-step models still trail full multi-step ones on at least one dataset.