emerging
Any-OPD: Heterogeneous On-Policy Distillation for Flow-Matching Models via Representation-Space Bridging
· 2026-08 · arXiv:2608.03316
Why it matters
Distills a 12B model into a 2.5B one while treating the teacher as a black-box sampler, requiring no shared latent space between them — relevant to the atlas because it makes distillation independent of the representation choice.
What this paper does
read: full textBefore this
On-policy distillation lets a student diffusion or flow-matching model be corrected by a teacher using samples the student itself generated, avoiding the exposure-bias problem of purely offline distillation. Every existing recipe assumes teacher and student share a VAE, so teacher latents are valid regression targets in student coordinates, a matching architecture for internal feature alignment, and a common timestep grid, so that step index i means the same noise level for both models.
The problem
The paper asks what happens when the strongest available teacher and the student one wants to deploy come from unrelated model families, so none of those shared-coordinate assumptions hold. Standard recipes have no answer here, teacher latents are meaningless in a foreign VAE's coordinate system, pixel-wise regression against a teacher that stochastically re-draws local detail collapses training into blur, and step-index alignment pairs states drawn from incomparable noise regimes.
The idea
Connect teacher and student at exactly one point that neither architecture nor latent space determines, a frozen, model-agnostic vision representation, specifically a frozen DINOv2-Base, that produces comparable embeddings for images decoded independently by either model. Correspondence between the two trajectories is recovered by matching continuous noise levels rather than step indices, since noise level is the only quantity the two trajectories genuinely share.
How it works
Both teacher and student are latent flow-matching generators, each with its own frozen VAE encoder and decoder and its own noise schedule. The student's velocity field, which maps a latent and a time in [0,1] to a velocity in that same latent space, is trained to predict a velocity in its own VAE latent space, exactly as in ordinary flow matching, both during the warm-start phase and during the main on-policy phase. Training proceeds in two phases. An anchoring phase runs offline for 400 steps over 10 noise levels; the teacher generates a target image, which is re-encoded through the student's own VAE encoder into a pseudo-target latent, and the student is trained with an ordinary flow-matching loss against that re-encoded target. This anchors the student in its own coordinate system before any cross-model comparison begins. The main on-policy phase, run for 800 steps with both teacher and student sampled at 20 Euler steps each, has the student generate a sample at its own noise level; the teacher re-noises that image to a chosen noise level, re-denoises it along its own trajectory, and decodes it. Correspondence between the two trajectories is matched by noise level rather than step index, picking the least-noisy student state that is still at least as noisy as the point where the teacher's refinement began. The student's sample and the teacher's refined output are then compared not in either model's latent space but through the frozen DINOv2-Base CLS token embedding, using a cosine-distance loss equal to one minus the cosine similarity between the two embeddings. Fine-tuning uses LoRA with rank 32 and alpha 64, learning rate 1e-4, batch size 4 per GPU, at 512x512 resolution.
Evidence
Distilling the 12B FLUX.1-dev teacher into the 2.5B SD3.5-Medium student, at matched 20-step Euler sampling for both models with no guidance-method comparison reported, PickScore rises from the SD3.5-Medium baseline's 0.846 to 0.884, close to FLUX.1-dev's own 0.878. HPSv3 rises from 9.12 to 10.97 against the teacher's 11.19, and ImageReward rises to 1.116, exceeding the teacher's own 0.971. An ablation on the bridging loss shows a naive per-pixel latent MSE loss collapses within the first training steps and an LPIPS loss degrades late in training, while the DINOv2-CLS cosine loss stays stable; plain CLS embeddings match or exceed richer DINOv2 variants tried. Direct latent regression against teacher latents, the naive same-space baseline, is reported to fail to train at all. Swapping in a different 6B teacher, Z-Image, with nothing else changed improves every metric further, including GenEval by 0.80 and DPG-Bench by 0.95, suggesting the recipe is not overfit to the FLUX and SD3.5 pair specifically.
Limitations
The authors' own discussion of limitations is brief. They frame combining multiple complementary teachers as a natural but unexplored remedy rather than something the paper demonstrates, and note that nothing in the formulation is specific to images beyond the choice of feature extractor, positioning extension to other latent generative modalities as further work rather than a result already shown.
Why it matters
It decouples on-policy distillation from representation choice. Because the bridge is a frozen, off-the-shelf vision embedding rather than a shared latent space, distillation can run from whichever teacher is currently strongest into whatever student architecture is being deployed, without retraining a shared VAE or matching timestep grids. For an atlas organized around representation and objective choices, this is a case where distillation is made independent of both.
Abstract, in the authors' own words
On-policy distillation, in which a teacher corrects samples that the student itself generates, presupposes that the two models speak the same language: identical VAE latents, matching architectures, and a common timestep grid. We ask what happens when none of this holds, as when the strongest teacher available and the student one wishes to deploy come from different model families, and find that the standard recipes have no answer: teacher latents cannot serve as targets in a foreign coordinate system, per-pixel losses against a teacher that stochastically re-draws local detail degenerate into blur or divergence, and timestep indices lose their meaning across mismatched schedules. We present Any-OPD, to our knowledge the first framework for on-policy distillation between arbitrary pairs of latent flow-matching generators. Any-OPD treats the teacher purely as a black-box sampler and connects the two models at exactly one point: a frozen, model-agnostic vision representation in which their independently decoded outputs are compared, sidestepping every assumption about latents, features, or architecture. Trajectory correspondence is recovered by matching continuous noise levels instead of step indices, and a brief anchoring phase, in which teacher samples are re-encoded through the student's own VAE, ensures the on-policy gradient measures sample quality rather than domain mismatch. Distilling the 12B FLUX.1-dev into the 2.5B SD3.5-Medium, Any-OPD lifts the student's PickScore from 0.846 to 0.884 and HPSv3 from 9.12 to 10.97, rivaling the teacher at a fifth of its size, where direct latent regression fails to train at all.
Research line
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a emerging entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.