Research map / Video / Video generation
ASTRAEA: A Token-wise Acceleration Framework for Video Diffusion Transformers
Haosong Liu, Yuge Cheng, Wenxuan Miao, Zihan Liu, Aiyue Chen, Jing Lin, Yiwu Yao, Chen Chen, Jingwen Leng, Minyi Guo, Yu Feng
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 |
|---|---|---|---|
| Astraea token selection at 50% budget vs ToCa on Wan 4s | Wan v2.1 1.3B · Wan (4s) | PSNR: 18.13 -> 28.12 dB | Improved |
| Astraea 40% vs original speedup | Wan v2.1 1.3B · Wan (4s) | VBench: 80.28 -> 79.78, 2.29x speedup | No clear effect |
| compute only selected queries, keeping all K and V (vs selecting Q and K) (ablation) | Astraea · Wan (4s) | VBench: 79.01 -> 79.96 | Improved |
| compute only selected queries, keeping all K and V (vs selecting Q and K) (ablation) | Astraea · Wan (4s) | PSNR: 18.10 -> 28.12 | Improved |
| block-level token budget search instead of timestep-level selection (ablation) | Astraea · Wan (4s) | PSNR: 22.71 -> 28.12 | Improved |
| block-level selection instead of timestep-level on speed (ablation) | Astraea · Wan (4s) | speedup: 1.97 -> 1.86 | Hurt |
| evolutionary search allocation instead of fixed token budget per timestep (ablation) | Astraea · Wan (4s) | VBench: 77.92 -> 79.96 | Improved |
| reduce token budget below 30% (ablation) | Astraea · Wan (2s) and OpenSora (2s) | VBench | Hurt |
| use more than 4 prompts in evolutionary search (ablation) | Astraea search · search | search outcome | No clear effect |
| scale Astraea to 8 GPUs (50% budget) | OpenSora · multi-GPU | speedup: 13.2x on 8 GPUs | Improved |
Closest papers by meaning
- AsymRnR: Video Diffusion Transformers Acceleration with Asymmetric Reduction and Restoration (ICML 2025 · similarity 0.89)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.89)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Sparse VideoGen2: Accelerate Video Generation with Sparse Attention via Semantic-Aware Permutation (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.
