From 1000-step DDPM to 1-step generation: a hands-on turbo diffusion walkthrough
RisingSayak · x · 2026-11-18
A step-by-step tutorial trains a 9.79M-param DiT on MNIST and walks the full acceleration ladder: 1000-step DDPM → DDIM/DPM-Solver/UniPC → rectified flow → 1-step methods (CT, CD, MeanFlow, DMD2).
Key observations:
- Going from 1000 to 5 steps costs almost nothing visible and needs no retraining; at 2-3 steps rectified flow already yields cleaner outputs while v-pred DDPM stays hazy.
- One-step generation needs a new objective: the denoiser predicts the average clean image, which from pure noise is just a blur of the mean '3'. Multi-step schedules sidestep this by narrowing gradually; CT, CD, MeanFlow and DMD2 change the training target instead.
A solid systematic intro to few-step sampling and distillation.
More from Multimodal
- LLaVA-OneVision-2.5 released: first clear signal that codec input beats frame sampling — liuziwei7 · 2026-11-18
- LMMS-Lab: codec sampling beats uniform frame sampling on general video tasks — liuziwei7 · 2026-11-18
- AI-generated Robert Pattinson photo set is nearly indistinguishable from real photography — justin_hart · 2026-11-18
- Clueso v2: AI video editor lets teammates pin comments to the second and the AI makes the edit — aakashgupta · 2026-11-18
- Users seek a simple Director node for Minimax H3 workflows — Kind_Owl2245 · 2026-11-18
- Adobe Firefly ambassador shares full workflow for embroidery-style AI animated short — azed_ai · 2026-11-18