ByteDance Seed's DiffusionOPSD: full-trajectory rewards for diffusion RL, 60% less GPU time
andrew_n_carr · x · 2026-08-29
ByteDance's Seed team released DiffusionOPSD (On-Policy Self-Distillation in Diffusion Models), tackling a core pain point of RL for diffusion models: endpoint rewards only arrive after the full image is decoded, leaving intermediate denoising steps unsupervised.
Method:
- A frozen behavior policy generates trajectories supplying low-noise query states and clean-output anchors;
- Normalized reward ascent/descent is converted into bounded positive and negative targets within a controlled clean-output radius;
- The trainable policy fits detached targets under a finite budget, then EMA refreshes the behavior policy for the next iteration.
Results: Extensive ablations on SD3.5-M and step-distilled Z-Image-Turbo across ten evaluators; best final held-out scores in reward-matched settings with roughly 60% fewer training GPU-hours than DiffusionNFT. Paper and code are public.
More from Research
- Anthropic shows AI researchers autonomously improving alignment of other models — VraserX · 2026-08-30
- Learn Positional Encodings derivation from first principles — zainhas · 2026-08-30
- COLM Paper Traces Capability Provenance in LLMs via Gradient Attribution — ziv_ravid · 2026-08-30
- Toby Ord paper argues recursive self-improvement has physical limits — Exponential View (Azeem Azhar) · 2026-08-30
- AI Formalization Tools Fable and Sol Spot First Repairable Error in Published Literature — Sauers_ · 2026-08-30
- Mark Schmidt Posts ICML Tutorial Video: Is Numerical Optimization Theory Irrelevant to ML Practice in 2026? — MarkSchmidtUBC · 2026-08-30