Mapping autoregressive vision until the field outgrew the map

What it does This repository is the companion to a TMLR 2025 survey paper, organizing hundreds of papers on autoregressive models for vision into a curated reading list. It catalogs research across image generation, video synthesis, 3D and point-cloud generation, multimodal models, and supporting infrastructure like tokenizers, acceleration techniques, and evaluation metrics. The maintainers frame it as a community resource for navigating a literature that has exploded across computer vision, embodied AI, and medical imaging.

The interesting bit The most telling entry is the update log: the authors explicitly paused proactive maintenance because “unified multimodal models” and “autoregressive diffusion-forcing video generation” have dissolved the boundaries of their own taxonomy. It is rare for a survey to admit, within a year of publication, that its categories no longer fit the landscape it set out to map.

Key highlights

• Peer-reviewed TMLR 2025 survey with authors from HKU, Tsinghua, Apple, and others
• Taxonomy spans pixel-wise and token-wise generation, safety, reasoning alignment, and scaling
• Tracks the field’s shift from pure autoregression toward hybrid diffusion-autoregressive methods
• Now accepts targeted pull requests but no longer proactively updates the paper list
• Includes a timeline visualization of methodological evolution

Caveats

• The repository is a curated bibliography, not a framework or codebase
• The maintainers note their categorical structure is increasingly outdated relative to current research directions

Verdict Researchers who need a structured bibliography of autoregressive vision generation will find the taxonomy useful; developers seeking runnable code or the latest unified multimodal architectures should look elsewhere.