Awesome Autoregressive Models in Vision: a TMLR 2025 survey repo now in maintenance mode

What the repository actually contains, and what it does not

This is a curated list of papers, not a library. The README describes it as "a curated list of papers about the latest advancements in autoregressive models in vision", and the repository has no releases and no stated primary language or licence in the metadata supplied. There is nothing to install, no API to call, and no inference code. The asset is the taxonomy plus the bibliography behind the TMLR 2025 survey of the same name, whose authors span HKU, Tsinghua, Duke, Rochester, Ohio State, ByteDance, UNC, Apple, PolyU and Princeton. If you arrived expecting a model checkpoint or a training script, you are in the wrong repository, and that is worth stating plainly because the name reads like a code project. What you get instead is an organised reading path through pixel-wise generation, token-based approaches, video, 3D, multimodal and acceleration work, with venue tags attached to entries such as the ICML 2021 oral on distribution smoothing for autoregressive modeling.

The problem it solves: finding the right paper among hundreds

Autoregressive visual generation sits at an awkward intersection. Image generation papers borrow tokenizers from language modelling, video papers borrow from diffusion, and 3D papers borrow from both. Searching arXiv directly returns a flat, undated pile. The repository's answer is a fixed hierarchy. The README's table of contents splits Image Generation into unconditional and class-conditioned, text-to-image, image-to-image translation and image editing; Video Generation into unconditional, conditional and embodied AI; 3D Generation into motion, point cloud and 3D medical; then Multimodal Generation, Other Generation, Benchmark and Analysis, Reasoning Alignment, Safety, Accelerating, Stability and Scaling, Tutorial, and Evaluation Metrics. That last group is the part most survey repositories omit. Having separate sections for safety and for evaluation metrics means a reader who cares about failure modes rather than architectures has somewhere to start. The intended audience is a graduate student or applied researcher entering the area, or someone already inside it who wants to check whether a subfield exists that they have not heard of.

How the taxonomy is organised and where it breaks

The structure is a two-level markdown outline: a top-level capability, then a sub-capability, then tagged entries. Each entry follows a consistent shape, with the venue in bold, the paper name, and links to the paper and to code where the authors provide one. The venue tag is doing real work here. It lets you filter out unrefereed preprints without opening each link, and the contributing instructions explicitly ask that an accepted version replace the preprint entry. The weakness is that the top level is organised by output modality, while the active research direction is organised by architecture. The maintainers say as much in the 2025-11-01 notice, writing that the current categories "no longer capture this evolving landscape" and naming unified multimodal models and autoregressive diffusion-forcing video generation as the two trends that broke the scheme. A reader looking for unified understanding-and-generation models will find one section for them, but a reader looking for diffusion-forcing video generation has no dedicated home for it. Treat the outline as a snapshot of how the field looked when the survey was written, not as a live map.

Getting value out of it: the actual workflow

There is no build step. You clone or browse, read the table of contents, and follow links. The one concrete convention is the entry format, which the README specifies in a markdown block: a bullet with the venue and year in bold, the paper name, then a Paper link and a Code link. If you intend to contribute, that format is the contract. The README directs additions to a pull request against the repository or to jhuang90@ur.rochester.edu, and asks that you look for missed papers or update a preprint that has since been accepted to a venue. Note that the pull request URL printed in the README points at a repository path spelled autoregressive-vision-survey, which differs from the repository name Autoregressive-Models-in-Vision-Survey. If you are opening a PR, use the GitHub interface from the repository you are actually on rather than copying that link verbatim. For citation, the README supplies a BibTeX entry keyed xiong2024autoregressive with eprint 2411.05902 and primaryClass cs.CV.

The maintenance freeze is the decisive fact about this project

On 2025-11-01 the maintainers announced they were moving to maintenance mode and pausing proactive updates. The wording is specific: the repository remains available as a reference and targeted pull requests are welcome, covering additions, corrections, or reorganisations around new trends. That is a meaningful distinction from abandonment. The list will not silently rot, because corrections and new-trend reorganisations can still land, but nothing arrives on a schedule. For a survey companion this is a reasonable end state. For a reader in late 2025 or 2026, it means the newest sections are the least reliable. If your interest is unified multimodal models or diffusion-forcing video generation, the repository is at its weakest exactly where the field is at its most active. The prior update, on 2025-05-31, revised the arXiv paper to add continuous autoregressive methods, computational costs, metric details and an expanded application roadmap, so the paper is ahead of the README list in those areas. Read the paper for the argument and the repository for the links.

Where a papers-with-code index is the better tool

The obvious alternative for discovery is a living index such as Papers with Code, or a venue proceedings browser. The difference in approach matters. This repository is curated and opinionated: a small group of authors decided what belongs in each section, and the value comes from that judgement, plus the accompanying survey that explains why the categories exist. A living index is exhaustive and mechanical: it indexes whatever is submitted and tags it by task, without an argument about structure. If you want to know whether a specific method has an official implementation, a task-tagged index answers faster. If you want to understand why pixel-wise and token-based generation are treated as separate families, and how acceleration work relates to both, the curated list plus the survey is the better route. The two are complements, not substitutes, and the maintenance freeze means the curated side will increasingly lag on recency while keeping its structural advantage.

Licence, contribution cost and what to verify before citing

The repository metadata supplied does not state a licence, and the README does not mention one. That is a gap worth checking directly on GitHub before you reuse the list wholesale, redistribute it, or mirror it into an internal wiki. The BibTeX block in the README covers the paper, not the repository contents, and a citation to the survey does not settle the terms for the list itself. Contribution cost is low in effort and non-trivial in latency: you need to match the markdown format exactly, and after the 2025-11-01 notice, additions are described as targeted rather than routine, so expect a slower review than the pre-freeze period. Before citing the survey, confirm the venue status on the OpenReview page linked from the README and pull the latest arXiv version, since the 2025-05-31 revision changed content on continuous autoregressive methods and computational costs. The repository's own description marks it TMLR 2025, and the README's update news confirms acceptance on 2025-03-11.

Editorial conclusion

Adopt it if you need a structured entry point into autoregressive visual generation and you are willing to read the arXiv paper alongside the list. Do not adopt it if you need runnable code, a package to install, or a list that is actively updated: the maintainers moved the repository to maintenance mode on 2025-11-01 and paused proactive updates. Before relying on it, open the arXiv abstract page for 2411.05902, confirm which taxonomy sections match your subfield, and check the most recent entries under Accelerating and Evaluation Metrics, since the README states those categories no longer capture unified multimodal models and autoregressive diffusion-forcing video generation.

Official sources

1. ChaofanTao/Autoregressive-Models-in-Vision-Survey on GitHub
2. Issues
3. Project website
4. README