Research map / Image generation / Fast diffusion sampling
StreamDiffusion: A Pipeline-level Solution for Real-Time Interactive Generation
Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, Masayoshi Tomizuka, Kurt Keutzer
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 Stream Batch denoising (ablation) | StreamDiffusion · Img2img | inference time: reduction by half compared to a conventional U-Net loop | Improved |
| add Self-Negative R-CFG instead of conventional CFG (ablation) | StreamDiffusion · CFG comparison, 5 steps | inference time (ms): 64.64 -> 31.47 (2.05x) | Improved |
| add Onetime-Negative R-CFG instead of conventional CFG (ablation) | StreamDiffusion · CFG comparison, 5 steps | inference time (ms): 64.64 -> 36.04 (1.79x) | Improved |
| use Onetime-Negative R-CFG at 1 step vs CFG (ablation) | StreamDiffusion · CFG comparison, 1 step | inference time (ms): 16.74 -> 16.55 (1.01x) | No clear effect |
| add TensorRT (ablation) | StreamDiffusion · Img2img, 1 step | inference time (ms): 21.34 -> 10.65 | Improved |
| add TensorRT (ablation) | StreamDiffusion · Img2img, 10 steps | inference time (ms): 96.94 -> 62.00 | Improved |
| add stochastic similarity filter (ablation) | StreamDiffusion · Static scene, 20 frames, RTX 3060 | average power (W): 85.96w to 35.91w | Improved |
| add stochastic similarity filter (ablation) | StreamDiffusion · Static scene, RTX 4090 | average power (W): 238.68w to 119.77w | Improved |
| add stochastic similarity filter (ablation) | StreamDiffusion · Dynamic scene, 1000 frames, RTX 4090 | average power (W): 236.13w to 199.38w | Improved |
| add future-frame attention via Stream Batch (ablation) | StreamDiffusion · User study, temporal consistency | preference (%): 9.67 -> 90.33 | Improved |
| use probabilistic SSF instead of hard threshold (ablation) | StreamDiffusion · User study, streaming quality | preference (%): 24.44 -> 75.56 | Improved |
| add SSF (ablation) | StreamDiffusion · User study, SSF imperceptibility | preference (%): 48.61 -> 51.39 | No clear effect |
Closest papers by meaning
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.88)
- FIFO-Diffusion: Generating Infinite Videos from Text without Training (NeurIPS 2024 · similarity 0.87)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.87)
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 2025 · similarity 0.87)
- Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025 · similarity 0.87)
- Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion (CVPR 2025 · 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.
