Research map / Video / Video generation
PipeDiT: Accelerating Diffusion Transformers in Video Generation with Task Pipelining and Model Decoupling
Sijie Wang, Qiang Wang, Shaohuai Shi
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 PipeSP (pipelined Ulysses) at moderate resolution (ablation) | OpenSoraPlan · 640x352x129 per-timestep (A6000) | per-timestep latency (s): 2.10 -> 1.83 (1.15×) | Improved |
| add PipeSP at low resolution (ablation) | OpenSoraPlan · 640x352x65 per-timestep (A6000) | per-timestep latency (s): 1.21 -> 1.69 (0.72×) | Hurt |
| add PipeSP across configurations (ablation) | OpenSoraPlan · per-timestep latency, 8 configs (A6000/L40) | speedup: 0.97× to 1.15× (A6000); 0.96× to 1.06× (L40) | Mixed |
| decouple diffusion and VAE (DeDiVAE) at low resolution (ablation) | OpenSoraPlan · 480x352x65, 30 steps (A6000) | latency (s): 314 -> 217 (1.45×) | Improved |
| decouple diffusion and VAE (DeDiVAE) at highest resolution (ablation) | OpenSoraPlan · 1024x576x129, 30 steps (A6000) | latency (s): 1995 -> 2138 (0.93×) | Hurt |
| add Aco attention co-processing on top of DeDiVAE+PipeSP at high workload (ablation) | OpenSoraPlan · 1024x576x129, 30 steps (A6000) | latency (s): 1936 -> 1690 | Improved |
| add Aco on top of DeDiVAE+PipeSP at low workload (ablation) | OpenSoraPlan · 480x352x65, 30 steps (A6000) | latency (s): 200 -> 261 | Hurt |
| add Aco on top of DeDiVAE+PipeSP at high workload (ablation) | HunyuanVideo · 1024x576x129, 30 steps (A6000) | latency (s): 3675 -> 3090 | Improved |
| decouple diffusion and VAE (DeDiVAE) versus offloading | HunyuanVideo · 480x352x129 peak memory | peak GPU memory (GB): 29.37 -> 41.44 | Hurt |
| decouple diffusion and VAE (DeDiVAE) versus offloading | OpenSoraPlan · 480x352x129 peak memory | peak GPU memory (GB): 18.4 -> 18.0 | Improved |
| use full PipeDiT instead of offloading baseline for 10 videos | HunyuanVideo · 480x352x97, 10 steps (A6000) | latency (s): 540 -> 165 (3.27×) | Improved |
| use full PipeDiT instead of offloading baseline at longest timesteps | HunyuanVideo · 1024x576x97, 50 steps (L40) | latency (s): 6952 -> 6453 (1.08×) | Improved |
Closest papers by meaning
- PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference (NeurIPS 2025 · similarity 0.91)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.90)
- Fast and Memory-Efficient Video Diffusion Using Streamlined Inference (NeurIPS 2024 · similarity 0.88)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.88)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.88)
- Training-Free Efficient Video Generation via Dynamic Token Carving (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.
