Research map / Video / Video generation
Lumos-1: On Autoregressive Video Generation with Discrete Diffusion from a Unified Model Perspective
Hangjie Yuan, Weihua Chen, Jun CEN, Hu Yu, Jingyun Liang, Shuning Chang, Zhihui Lin, Tao Feng, Pengwei Liu, Jiazheng Xing, Hao Luo, Jiasheng Tang and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use temporal tube masks (AR-DF) instead of global random masks (ablation) | Lumos-1 1B · VBench Imaging Quality | IQ: 0.424 -> 0.559 | Improved |
| use temporal tube masks instead of random masks (ablation) | Lumos-1 1B · GenEval | GenEval: 0.593 -> 0.591 | No clear effect |
| use loss reweighting (step/linear decay) instead of tube masks (ablation) | Lumos-1 1B · VBench IQ | IQ: 0.559 -> 0.464 / 0.410 | Hurt |
| add distributed frequency allocation to M-RoPE (ablation) | Lumos-1 1B · GenEval | GenEval: 0.414 -> 0.591 (with scaling) | Improved |
| add scaled 3D positions (ablation) | Lumos-1 1B · GenEval | GenEval: 0.566 -> 0.591 | Improved |
| scale factors (4,8,8) vs (1,1,1) (ablation) | MM-RoPE · GenEval | GenEval: 0.566 -> 0.591 | Improved |
| scale factors (8,16,16) vs (4,8,8) (ablation) | MM-RoPE · VBench IQ | IQ: 0.559 -> 0.553 | No clear effect |
| use MM-RoPE instead of HoPE (ablation) | Lumos-1 1B · GenEval | GenEval: 0.570 -> 0.591 | Improved |
| omit inference-time masks in AR-DF (ablation) | Lumos-1 · video quality | artifacts | Hurt |
| small-scale SFT on high quality data (ablation) | Lumos-1 · GenEval | overall: 0.664 -> 0.791 (3.6B, with resolution 352 -> 512) | Improved |
| use mask prediction with KV cache instead of next-token prediction (ablation) | Lumos-1 · video generation latency | seconds: 960.0 -> 77.8 | Improved |
| add 3D RoPE priors versus 1D RoPE (ablation) | Lumos-1 · inference latency | latency: 3.5%-4.1% overhead | Hurt |
Later papers that compared with this method
1 later accepted paper reports Lumos-1 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Uniform Discrete Diffusion with Metric Path for Video Generation (ICLR 2026 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.91)
- ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation (ICLR 2025 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.89)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
