Poly-OPD: Heterogeneous Multi-Teacher On-Policy Distillation for Capability-Selectable Flow Models
Poly-OPD distills two stronger image models into one smaller, switchable student without forcing their latent spaces to match.
The framework uses a pixel bridge so a student image can be re-encoded and refined by whichever teacher is selected. It supervises the result in frozen DINOv2 space, avoiding direct latent-space compatibility. The paper reports a 2.5B SD3.5-Medium student distilled from FLUX.1-dev and Z-Image, with GenEval rising from 67.3 to 73.3 and DrawBench HPSv3 from 9.34 to 11.35. HF Daily Papers' note
The framework uses a pixel bridge so a student image can be re-encoded and refined by whichever teacher is selected. It supervises the result in frozen DINOv2 space, avoiding direct latent-space compatibility. The paper reports a 2.5B SD3.5-Medium student distilled from FLUX.1-dev and Z-Image, with GenEval rising from 67.3 to 73.3 and DrawBench HPSv3 from 9.34 to 11.35. HF Daily Papers' note
score 4