Ming-VideoMAR: an autoregressive video generation model based on continuous tokens
Original title:inclusionAI/Ming-VideoMAR
AI summary
Ming-VideoMAR is a decoder-only autoregressive image-to-video generation model that uses continuous tokens to unify visual representation. It is the first to treat temporal causality and spatial bidirectionality as core principles of video autoregression, and proposes a next-frame diffusion loss integrated with masked generation. The model achieves zero-shot resolution scaling for video generation for the first time, flexibly generating videos far beyond the training resolution. It performs strongly in training and inference efficiency; its parameter count, training data volume, and GPU consumption are only a tiny fraction of those of the previous best model Cosmos (9.3%, 0.5%, and 0.2%), while surpassing it in both quantitative and qualitative evaluations. The model code and checkpoints are open-sourced, and the paper has been accepted by NeurIPS 2025.
Key Takeaways
- Ming-VideoMAR is a decoder-only autoregressive image-to-video generation model that uses…
- It is the first to treat temporal causality and spatial bidirectionality as core principles of…
- The model achieves zero-shot resolution scaling for video generation for the first time,flexibly…
- It performs strongly in training and inference efficiency
- its parameter count,training data volume,and GPU consumption are only a tiny fraction of those of…
Quick Facts
- Published
- Updated
- Source
- 蚂蚁 inclusionAI:GitHub 新仓库
- Category
- Model Releases
- AI score
- 66
蚂蚁把自回归视频生成的训练成本砍到 Cosmos 的 0.2% 还能赢,这个效率信号比分数本身更值得关注,做视频生成的团队该认真看看它的课程学习和渐进分辨率策略。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Ming-VideoMAR: an autoregressive video generation model based on continuous tokens. October 10, 2025. https://new.nat6.net/en/news/cmorb7ik40074slhfvdmvwsw7/