Poly-OPD: Heterogeneous Multi-Teacher On-Policy Distillation for Capability-Selectable Flow Models
Siming Fu, Haojun Xu, Ruizhe He, Zheming Fu, Hualiang Wang, Jie Huang, Xiaoxiao Ma, Mingchen Zhong, Weihu Huang, Xiaoxuan He, Linjiang Huang, Si Liu
cs.CV
2026-08-05
Poly-OPD distills FLUX and Z-Image into one 2.5B student despite incompatible latents, bridging them in pixel and DINOv2 space; the student reaches GenEval 73.3, beating both larger teachers.
Open text-to-image models specialize: FLUX.1-dev wins on aesthetics and preference alignment, Z-Image follows compositional instructions (count, position, color, attribute) more faithfully. The two strengths rarely coexist in one checkpoint, and the specialists are architecturally heterogeneous, with different VAEs, denoisers, and noise schedules. Standard distillation (score or trajectory matching) requires a shared VAE and noise schedule; a heterogeneous teacher shares neither, so its latents and denoising trajectories are not valid targets for the student. Falling back to training on teacher-generated images is off-policy, and the train-inference gap caps fidelity. On top of that, two capabilities distilled into one net interfere, and compositional sub-skills are unevenly developed.
The student is SD3.5-Medium (2.5B) flow-matching; teachers are FLUX.1-dev (12B, preference mode) and Z-Image (6B, composition mode), both frozen. Three mechanisms.
A pixel bridge for heterogeneous on-policy distillation. The student samples its own output without gradients, decodes to pixels (the only shared coordinate system), re-encodes with the active teacher's VAE, re-noises to a magnitude-matched level (not a timestep index, since schedules differ), and lets the frozen teacher refine the last r steps. The refined image is a teacher correction of the student's own output, an on-policy target from an otherwise unreadable teacher. Supervision is a cosine loss in frozen DINOv2 CLS feature space, not in any latent, because DINOv2 is invariant to model-specific pixel statistics and makes cross-model comparison well-posed. Two stages: an off-policy warm start on teacher samples moves the student near the teacher distribution, then on-policy distillation closes the gap.
A gradient-compatibility diagnostic for capability-selectable adapters. Rather than guess, the authors measure the mean cosine similarity between preference-mode and composition-mode LoRA gradients over 1,000 paired samples. Attention gradients stay aligned across modes (shareable); FFN gradients conflict and go negative in several blocks. So the design is dictated by the measurement: one attention LoRA shared across all teachers, per-capability FFN adapters kept separate. The backbone is frozen and only LoRA is trained; switching capability is an adapter swap, not a second model.
A gap-aware curriculum. Within composition mode, prompts are allocated by the remaining teacher-student gap rather than raw difficulty, and weights anneal to zero as a gap closes.
On GenEval the 2.5B student goes from 67.3 to 73.3, beating both larger teachers (Z-Image 69.4, FLUX 65.2); DrawBench HPSv3 goes from 9.34 to 11.35. By sub-category, two-object rises from 82.07 to 97.00 and attribute from 58.75 to 67.00 (above both teachers' 52.80 and 44.30), so the student is not capped by the teachers' per-category ceilings.
| Metric | SD3.5-M | FLUX | Z-Image | Poly-OPD |
| GenEval | 67.3 | 65.2 | 69.4 | 73.3 |
| DrawBench HPSv3 | 9.34 | 11.93 | 9.08 | 11.35 |
| ImageReward | 0.92 | 0.92 | 0.83 | 1.17 |
To be fair, on DrawBench HPSv3 the student at 11.35 still trails FLUX at 11.93, recovering about 78% of the gap, while on ImageReward and the alignment sub-score it surpasses FLUX. Ablations: removing the warm start collapses GenEval to 49.2 (-24.1), so warm start is mandatory; removing the capability-selectable adapters costs 11 GenEval points; swapping the perceptual space to SigLIP or ConvNeXt drops GenEval by about 7 points, and direct student-latent MSE collapses outright.
For T2I distillation, the inability to distill from heterogeneous teachers is a long-standing pain point, and the pixel bridge plus DINOv2 space is a clean fix; capability switching costs only an adapter swap with sublinear parameter growth, which is practical. The gradient diagnostic that decides what to share versus isolate is a generalizable, measurement-driven approach rather than guesswork.
There is no dedicated limitations section. Only two teachers are demonstrated, though the method claims to scale to several modes; only one student backbone (SD3.5-Medium) is tested. The pixel bridge costs one student rollout plus r teacher refinement steps plus a backward pass per step, but the paper reports no FLOPs or GPU hours, so whether it beats just running FLUX in practice is unclear. Warm start is mandatory, so this is not zero-shot. Evaluation is only at 512 by 512, with no higher resolution or aspect ratio. The counting sub-skill barely moves (GenEval count 59.06 to 59.90, below both teachers), so composition gains are uneven. Everything rests on automatic metrics with no human study. DPG-Bench average of 85.80 is still 0.28 below the Z-Image teacher's 86.08, so it does not beat the teacher on every benchmark. For the record, the authors list Joy Future Academy (with Zhejiang University and Beihang), with a jd.com contact email, so it is not straightforwardly a JD paper.