Autoregressive image models have a guidance problem — and a new framework called Information-Grounding Guidance aims to fix it.
Models that generate images by predicting progressively higher-resolution scales suffer from a subtle but damaging flaw: as resolution increases across timesteps, information between image patches falls out of sync. That inconsistency causes the model's guidance signals to drift away from the visually meaningful parts of the image, producing outputs that are blurry, incoherent, or semantically off. The researchers behind IGG address this by using an attention-based dynamic weighting system that anchors guidance to semantically important tokens at each step, keeping the model's attention where it should be. Tests on both class-conditioned and text-to-image tasks show sharper, more coherent results.
Autoregressive approaches to image generation have been gaining ground on diffusion models, in part because they share architecture with large language models and can be trained on the same infrastructure. A guidance failure at the sampling stage is therefore a meaningful obstacle to that momentum — fixing it without adding heavy compute overhead matters. IGG's attention-based approach sidesteps the need for retraining the underlying model, which makes it more practical to drop into existing pipelines.
The code is publicly available on GitHub, which means independent researchers can replicate and stress-test the claims — a useful check given that image quality benchmarks are notoriously easy to game.