Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior
Xiang Li, Dianbo Liu, Kenji Kawaguchi · 2026
Read Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior by Xiang Li, Dianbo Liu, Kenji Kawaguchi (2026) 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-2606-02453.txt.
Description
Li and co-authors attack mode collapse in image generation. Models often repeat a few favorite outputs. Past fixes intervene during the generation path itself. The team spots a blind spot at the start. Standard Gaussian noise ignores the guidance landscape. Trajectories then fall into dominant modes.
Their fix samples the start from a guidance potential posterior. The prior is reweighted toward diversity-rich zones. Sampling uses Langevin dynamics, in a method called DivIn. The noise is steered away from collapse zones. It stays anchored to the valid data manifold. DivIn runs at inference time only. It fits both diffusion and flow matching models. Tests cover class-to-image and text-to-image tasks. DivIn tops rivals on diversity. It also stacks with path-based fixes to widen the diversity-quality frontier.
Catalog record
| CREATOR | Xiang Li, Dianbo Liu, Kenji Kawaguchi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 77 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.02453