Research map / Video / Video generation
BAgger: Backwards Aggregation for Mitigating Drift in Autoregressive Video Diffusion Models
Ryan Po, Eric Ryan Chan, Changan Chen, Gordon Wetzstein
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 |
|---|---|---|---|
| add a second BAgger round (ablation) | BAgger · VBench 50s | Imaging quality: 53.67 -> 59.98 | Improved |
| add a third BAgger round (ablation) | BAgger · VBench 50s | Imaging quality: 59.98 -> 63.41 | Improved |
| add a second BAgger round (ablation) | BAgger · VBench 50s | Imaging drift (lower is better): 11.83 -> 5.92 | Improved |
| add a third BAgger round (ablation) | BAgger · VBench 50s | Imaging drift (lower is better): 5.92 -> 3.57 | Improved |
| add a third BAgger round (ablation) | BAgger · VBench 50s | Aesthetic drift (lower is better): 3.79 -> 3.29 | Improved |
| use a single BAgger round instead of seed-only training (ablation) | BAgger Round 1 · VBench 50s | Imaging quality: 59.97 -> 53.67 | Hurt |
| use a single BAgger round instead of seed-only training (ablation) | BAgger Round 1 · VBench 50s | Imaging drift (lower is better): 7.34 -> 11.83 | Hurt |
| add three BAgger rounds on top of Diffusion Forcing seed training (ablation) | Diffusion Forcing · VBench 50s | Aesthetic drift (lower is better): 5.12 -> 3.29 | Improved |
| add three BAgger rounds on top of Diffusion Forcing seed training (ablation) | Diffusion Forcing · VBench 50s | Dynamic degree: 81.25 -> 76.56 | Hurt |
| add three BAgger rounds on top of Diffusion Forcing seed training (ablation) | Diffusion Forcing · VBench 50s | Subject consistency: 80.70 -> 84.05 | Improved |
| add noise to context frames at test time (sigma_test 0.2) (ablation) | Diffusion Forcing · VBench 50s | Imaging drift (lower is better): 7.34 -> 12.27 | Hurt |
| use History Guidance instead of plain Diffusion Forcing (ablation) | Diffusion Forcing · VBench 50s | Dynamic degree: 81.25 -> 59.38 | Hurt |
Closest papers by meaning
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.89)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.89)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- Ca2-VDM: Efficient Autoregressive Video Diffusion Model with Causal Generation and Cache Sharing (ICML 2025 · similarity 0.88)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.88)
- TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models (CVPR 2024 · 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.
