Research map / Video / Video generation
Scaling Autoregressive Video Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Scale the spatiotemporal subscaling model to large size | Video Transformer · BAIR Robot Pushing | bits/dim: 1.53 -> 1.35 | Helped |
| Use spatiotemporal subscaling instead of spatial subscaling (ablation) | Video Transformer · Kinetics | FVD: 263±6 -> 195±7 | Helped |
| Use single-frame generation instead of spatiotemporal subscaling (ablation) | Video Transformer · Kinetics | bits/dim: 1.19 -> 1.14 | Helped |
| Use spatiotemporal subscaling instead of single-frame generation (ablation) | Video Transformer · Kinetics | FVD: 207±8 -> 170±5 | Helped |
| Use spatial subscaling rather than no subscaling (ablation) | Video Transformer · Moving MNIST | Nats/Frame: 86.2 -> 91.8 | Hurt |
| Increase model size from base to large | Video Transformer · Kinetics | FVD: 195±7 -> 170±5 | Helped |
| Increase hidden size (ablation) | Video Transformer · BAIR Robot Pushing validation | bits/dim: 1.65 -> 1.55 -> 1.47 -> 1.40 | Helped |
| Increase the number of layers (ablation) | Video Transformer · BAIR Robot Pushing validation | bits/dim: 1.63 -> 1.55 -> 1.48 -> 1.45 | Helped |
| Increase the number of attention heads (ablation) | Video Transformer · BAIR Robot Pushing validation | bits/dim: 1.59 -> 1.55 -> 1.51 -> 1.47 | Helped |
| Shuffle block-size ordering between layers (ablation) | Video Transformer · BAIR Robot Pushing validation | bits/dim: within 0.01 bits/dim | No clear effect |
| Lower sampling temperature from 1.0 to 0.9 (ablation) | Video Transformer · BAIR Robot Pushing | FVD: 122±4 | Helped |
| Replace discretized one-hot input channels with continuous representations (ablation) | Video Transformer · Video generation | performance and sample quality | Hurt |
Closest papers by meaning
- Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation (NeurIPS 2025 · similarity 0.90)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.89)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- Video In-context Learning: Autoregressive Transformers are Zero-Shot Video Imitators (ICLR 2025 · similarity 0.88)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- ScaMo: Exploring the Scaling Law in Autoregressive Motion Generation Model (CVPR 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.