HMAR: Efficient H ierarchical M asked A uto- R egressive Image Generation
Hermann Kumbong, Xian Liu, Tsung-Yi Lin, Ming-Yu Liu, Xihui Liu, Ziwei Liu, Daniel Y. Fu, Christopher Ré, David W. Romero · 2025
Read HMAR: Efficient H ierarchical M asked A uto- R egressive Image Generation by Hermann Kumbong, Xian Liu, Tsung-Yi Lin, Ming-Yu Liu, Xihui Liu, Ziwei Liu, Daniel Y. Fu, Christopher Ré, David W. Romero (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-2506-04421.txt.
Description
Hermann Kumbong and eight colleagues present a 2025 paper on generating images scale by scale, faster and with a freer hand. The earlier VAR approach predicts a whole resolution layer at once, conditioned on every earlier layer, which blurs quality, stretches sequences, and locks the sampling plan at training time. Their method, HMAR, changes the recipe.
Each scale now depends only on the one just before it, and masked prediction fills each scale in a few controllable steps. On standard ImageNet tests the models matched or beat their rivals, trained more than two and a half times faster, and used a third of the memory at generation time. The sampling plan can also change after training, and the model edits images without extra training.
Catalog record
| CREATOR | Hermann Kumbong, Xian Liu, Tsung-Yi Lin, Ming-Yu Liu, Xihui Liu, Ziwei Liu, Daniel Y. Fu, Christopher Ré, David W. Romero |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-04421