Diffusion models can already un-blur photos and fill in missing pixels. A new paper argues the math underneath those tricks has been shakier than it looks, and offers a fix.
The paper tackles inverse problems, reconstructing a clean image from a corrupted one, like a blurry photo or a low-resolution scan, using a pretrained diffusion model as a prior. The catch is a likelihood term inside the sampling process that is mathematically intractable, so most methods approximate it with heuristics. The authors instead build a family of sampling equations with a provable property: an exploration parameter can be tuned without changing what the model actually converges to. They pair that with a corrector step and a carefully split numerical solver, prove convergence and error bounds, then test it on FFHQ face images and ImageNet with 100 sampling steps for super-resolution and deblurring.
This matters because diffusion-based restoration already underpins consumer photo tools and scientific imaging pipelines, places where a confident-looking output is not the same as a correct one. Most existing samplers for this problem work well in practice but lean on approximations nobody has fully justified. Proofs like this one turn a black box into something closer to an engineering spec, which counts for more as these tools move beyond sharpening vacation photos.
Still, this is a 100-image ablation and two standard benchmarks, not a shipped product. Rigorous math is nice; it does not yet tell you how the thing behaves on the messy, real-world corrupted image that never showed up in FFHQ.