Research map / Video / Video generation
CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CanvasMAR, a masked autoregressive video prediction model that uses a one-pass blurred next-frame "canvas" as a non-uniform mask, plus motion-aware sampling order and compositional CFG, to get high-fidelity frames with few sampling steps.
Problem
Video masked autoregressive models produce distorted frames when sampling individual frames with only a few spatial autoregressive steps, due to lacking a global structural prior.
Approach
A Canvas ViT predicts a blurred one-step expectation of the next frame that replaces the uniform learnable mask in MAR, supplying global structure early in sampling; a staticness head biases sampling toward low-motion regions first, and compositional CFG jointly up-weights canvas and temporal conditions.
Key results
- On BAIR, CanvasMAR (N_AR=4) reaches debiased FVD 26.6, the best among compared models, while standard FVD 66.0 trails only MAGVIT-L (62).
- On Kinetics-600 at 64x64, CanvasMAR (N_AR=10) reaches FVD 6.3, second only to MAGVIT-v2 (4.3) among autoregressive models and ahead of MAGI (11.5).
- CanvasMAR is about 5.7x faster than DFoT in latency at the largest batch size on Kinetics-600.
Surprising findings
- Under debiased FVD on BAIR, CanvasMAR surpasses MAGVIT despite generating frames sequentially, which the authors attribute to MAGVIT's non-causal design.
Limitations
- Evaluated primarily on video prediction with context frames; unconditional generation not evaluated.
- Produces severely distorted results for videos with substantial motion, since an overly blurred canvas can mislead the Spatial MAR.
- Largest Spatial MAR has only half the layers of NOVA, so scaling is untested.
Experimental setup
Datasets: BAIR UCF-101 Kinetics-600
Benchmarks: BAIR UCF-101 Kinetics-600
Metrics: FVD debiased FVD inference latency
Baselines: MAGI MAGVIT MAGVIT-v2 DFoT Rolling Diffusion VDM RaMViD MCVD Phenaki MaskViT NÜWA CCVS
Builds on: MAR NOVA MaskGiT MAGI compositional classifier-free guidance
Closest papers by meaning
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.90)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.88)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.88)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
- Taming Teacher Forcing for Masked Autoregressive Video Generation (CVPR 2025 · similarity 0.87)
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.