Visual autoregressive models have made progress in image generation but still struggle with compositional tasks—correctly assigning attributes to objects and preserving spatial relations. A new arXiv preprint introduces ATLAS (Attention-Guided Test-Time Latent Selection) to address this problem.
A straightforward fix is to search over multiple generation paths and pick the best result, but that requires completing every candidate, which is costly. ATLAS instead uses attention information at test time to select promising latents early, aiming to improve composition without the full computational overhead.
The abstract does not yet provide experimental details or quantitative results, so the claimed benefits are based on the authors' description. Still, the idea of using attention signals to guide latent selection is a plausible direction for reducing inference cost while improving compositional fidelity.