Fiberwise Optimal Transport Schedules Cut Flow Matching FID by 38.6%
MaxUnfried · x · 2026-09-15
A new preprint introduces model-aware diffusion/flow-matching schedules based on fiberwise optimal transport: compatible signal/noise decompositions form affine 'fibers,' and a fiberwise prediction risk combined with a kinetic action yields a closed-form optimal time allocation, estimable from an early checkpoint.
Evaluated across DDPMs and flow matching over prediction targets, datasets, and architectures, the schedules consistently beat strong baselines — including a 38.6% relative FID reduction for flow matching on CIFAR-10 at 16 function evaluations. Risk profiles from independently trained models align closely in kinetic reference coordinates.
More from Multimodal
- Dev cracks image conditioning on a DIY video model trained on one hour of footage — pixlpa · 2026-09-15
- Meta's Muse Voice Transcribe Goes Live in LiveKit Agents with 20+ Speaker Diarization — armand_ruiz · 2026-09-15
- 24-Second AI 'Fire God Origin' Short: One Continuous Shot, Speed Ramps Only — The_boneguy · 2026-09-15
- Creator puts himself on Survivor using Seedance 2.5 video generation — chrisfirst · 2026-09-15
- Image Tokenizers Define the Visual Language of Unified Multimodal Models — peterxichen · 2026-09-15
- Creator Combines GPT-6 Astra With Godot and Blender for Striking Results — TheMoonMidas · 2026-09-15