worldbench/DanceOPD

What it solves

DanceOPD addresses the problem of "capability interference" in image generation models. When developers try to combine multiple capabilities—such as text-to-image generation, local editing, and global style transformations—into a single model, improving one often leads to the degradation of others.

How it works

DanceOPD treats different source capabilities as "velocity fields." Instead of simple data mixing, it uses an on-policy distillation process:

1. Hard-routed Sampling: For each training step, the system selects one specific capability (route) and its corresponding frozen teacher model.
2. Student Rollout: The current student model generates a trajectory (rollout).
3. On-Policy Query: The system queries the frozen teacher model using a state actually visited by the student, rather than a pre-defined offline state.
4. Velocity Matching: The student is updated using a simple velocity-MSE objective to match the teacher's output at that specific state.

Who it’s for

This project is for AI researchers and developers working with flow-matching image generators (like SD3.5 or Z-Image) who want to synthesize multiple specialized capabilities into one compact, high-performance model without needing complex reward models or adversarial critics.

Highlights

• On-policy field query: Teachers supervise states the student actually visits, improving alignment.
• Hard-routed capability matching: Prevents ambiguous averaging by assigning samples to a single semantically valid teacher.
• Backend-extensible: Supports multiple flow backbones including SD3.5 and Z-Image.
• Simple Objective: Uses direct velocity MSE, removing the need for reward models or adversarial training.
• Multiple Training Modes: Offers danceopd (single-query ODE), diffusionopd (dense ODE-KL), and flowopd (SDE + PPO clipping).