Sander Dieleman's 84-minute deep dive: flow maps learn the integral of diffusion models to speed up sampling
sedielem · x · 2026-10-25
DeepMind researcher Sander Dieleman published an 84-minute blog post, "Learning the integral of a diffusion model," a systematic deep dive into flow maps, a hot research topic.
Key points:
- Standard diffusion sampling iteratively predicts the tangent direction and integrates across noise levels in small steps; flow maps instead predict any point on a path from any other point — directly learning the integral — enabling much faster sampling.
- Beyond speed, flow maps enable more efficient reward-based learning and improved sampling steerability.
- The post surveys the many ways to build and train flow maps, notes the field's confusing formalisms and terminology, and connects to his two prior posts on flow maps and continuous diffusion for language.
More from Multimodal
- Stacking two Midjourney sref codes yields a consistent blue-ink pen sketch style — sergeantsref · 2026-10-25
- Creator builds year-long video pipeline pairing Hailuo H3 with Magnific MCP — TinfoilTricorn · 2026-10-25
- Kling 4 Noir Mystery Pilot Short Shows Off Stylized AI Video Storytelling — RealBrainRocket · 2026-10-25
- Experimenting with Suno: blending free jazz and noise rock — AndyMasley · 2026-10-25
- Same 2023 DALL-E 3 prompts rerun on Krea 2 show how far image models have come — Furacao__Boey · 2026-10-25
- SeeDance 2.5 at ~$7 for 30s of 720p video looks cheap next to human animators — AIandDesign · 2026-10-25