worldbench / DiffusionOPSD

What it does

• On-policy query collection. Supervision is built on states visited by the current behavior policy rather than offline or forward-noised substitutes.
• Explicit reward-guided targets. Normalized reward ascent and descent construct bounded positive and negative clean-output targets.
• Detached finite fitting. Reward/decoder graphs are discarded before policy fitting, separating target quality from model realization.
• Continually refreshed supervision. An EMA behavior policy regenerates trajectories, anchors, and targets after each outer update.
• Two distinct backbones. The release supports SD3.5-M at 512² and the native few-step Z-Image-Turbo regime at 1024².