Ambient diffusion was a really cool paper in this space. Given corrupted data, corrupt even further, such that the model can’t infer what wa…

Ambient diffusion was a really cool paper in this space. Given corrupted data, corrupt even further, such that the model can’t infer what was genuine corruption vs synthetically added. In expectation you recover distribution. And similar claimed benefits on avoiding memorization. Training a diffusion model has always been synonymous with one idea: add some noise to an image, then learn to remove it. Since we know where we started and where we ended up, the natural thing to do is to ask the model to recover the signal everywhere. But is it really needed?

Source: Linus Lee (X) | 2026-04-29