Research map / Video / Video generation
Recurrent Autoregressive Diffusion: Global Memory Meets Local Attention
Taiye Chen, Zihan Ding, Anjian Li, Christina Zhang, Zeqi Xiao, Yisen Wang, Chi Jin
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 clean instead of noised frames as RNN memory input (ablation) | RAD LSTM-f · LSTM-f memory-noise ablation | PSNR: 14.70 -> 15.30 | Improved |
| use clean instead of noised frames as RNN memory input (ablation) | RAD LSTM-f · LSTM-f memory-noise ablation | SSIM: 0.38 -> 0.41 | Improved |
| use frame-wise (stride 1) instead of chunk-wise autoregression | RAD TTT · Memory Maze | PSNR: 14.36 -> 15.50 | Improved |
| use frame-wise instead of chunk-wise autoregression | RAD Mamba2 · Memory Maze | PSNR: 13.80 -> 15.35 | Improved |
| use frame-wise instead of chunk-wise autoregression | RAD LSTM · Memory Maze | PSNR: 15.64 -> 15.50 | Mixed |
| use frame-wise instead of chunk-wise autoregression | RAD Mamba2 · Minecraft | PSNR: 12.72 -> 16.70 | Improved |
| use frame-wise instead of chunk-wise autoregression | RAD TTT · Minecraft | PSNR: 14.04 -> 16.72 | Improved |
| use frame-wise instead of chunk-wise autoregression | RAD LSTM · Minecraft | PSNR: 14.24 -> 16.59 | Improved |
| use LSTM instead of Mamba2 in chunk-wise RAD | RAD chunk-wise · Memory Maze | PSNR: 13.80 -> 15.64 | Improved |
| use LSTM instead of TTT in frame-wise RAD | RAD frame-wise · Memory Maze | PSNR: 15.50 -> 15.50 | No clear effect |
| use partial-overlap stride 10 instead of chunk-wise stride 20 for LSTM (ablation) | RAD LSTM · Memory Maze | PSNR: 15.64 -> 15.06 | Hurt |
| add frame-wise RAD with LSTM over Diffusion Forcing | RAD LSTM-f · Memory Maze | PSNR: 14.73 -> 15.50 | Improved |
Closest papers by meaning
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.91)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.91)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.90)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.89)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.89)
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.
