Adaptive Destruction Processes for Diffusion Samplers
Timofei Gritsaev, Nikita Morozov, Kirill Tamogashev, Daniil Tiapkin, Sergey Samsonov, Alexey Naumov, Dmitry Vetrov, Nikolay Malkin · 2025
Read Adaptive Destruction Processes for Diffusion Samplers by Timofei Gritsaev, Nikita Morozov, Kirill Tamogashev, Daniil Tiapkin, Sergey Samsonov, Alexey Naumov, Dmitry Vetrov, Nikolay Malkin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-01541.txt.
Description
Timofei Gritsaev and seven co-authors argue in 2025 that diffusion samplers work better when the noise-adding process is learned too, not fixed by formula. Such samplers draw samples from a target distribution given only its unnormalized density. Most designs treat generation as an echo of a smooth continuous process.
The team instead treats the model as a discrete policy that must produce good samples in very few steps. Both the building and the destruction steps become free Gaussian transitions that training can shape. With few steps, the flexible version converged faster and sampled better across benchmarks. Careful ablations show which design choices keep training stable. The approach also scaled to guided image generation in a GAN latent space.
Catalog record
| CREATOR | Timofei Gritsaev, Nikita Morozov, Kirill Tamogashev, Daniil Tiapkin, Sergey Samsonov, Alexey Naumov, Dmitry Vetrov, Nikolay Malkin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 105 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-01541