🌈 Introducing DiverseVAR, a training-free framework that unleashes the inherent generative diversity of Visual Autoregressive models while preserving fidelity and text–image alignment.

[wangtong627]

Thanks a lot to the inspiring progress in Visual Autoregressive (VAR) models!

We introduce DiverseVAR, a simple yet effective training-free framework that restores the lost generative diversity in VAR models.
By strategically manipulating the pivotal component at early scales, DiverseVAR significantly boosts diversity without harming fidelity or text-image alignment.
On both Infinity-2B and Infinity-8B, it consistently improves Recall, Coverage, and FID while keeping CLIP scores nearly unchanged.

Arxiv: https://arxiv.org/abs/2511.17074
Github: https://github.com/wangtong627/DiverseVAR
Huggingface Daily Paper: https://huggingface.co/papers/2511.17074