You Might Not Need 50 Diffusion Steps
summary
A video diffusion pipeline generates near real time on a single Blackwell B200 by combining dynamic quantization (Flux 2 work with Black Forest Labs), chunk-aware caching that skips latent regions that barely change, and step distillation that trains a student to match a teacher in four steps or one step in some cases. Results live in NVIDIA’s open-source FastGen repo and report additive gains up to 10–200×, trading compute and post‑training effort for real‑time throughput.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:14Diffusion models overview
- 3:23Optimization strategies
- 4:03Quantization techniques
- 6:52Caching strategies
- 9:39Distillation for faster steps
- 13:06FastGen framework
- 15:53Incremental optimization
- 16:38Q&A
related talks
12mEfficient Workflow Serving for Diffusion Models with Many Adapters
39mDemocratizing Diffusion Models with Diffusers
10mDiffusion models
25mtorch.compile and Diffusers: A Hands-On Guide to Peak Performance
10mLightning Talk: Diffusers: Bringing Cutting-Edge Diffusion Models to the Masses
45mDiffusion to Prod: vLLM-Omni Unifies Multimodal Inference
01h 16mOptimizing workloads with Transformers and Diffusers
20mRealtime Video AI with Diffusion Models