{"id":"8653bcc6-7784-459e-b437-bd20a095fa32","arxiv_id":"2608.04349","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"Poly-OPD distills two heterogeneous text-to-image teachers into one 2.5B student via a pixel bridge and DINOv2-space supervision, surpassing both teachers on GenEval.","lead":"Poly-OPD is a training method that lets a small text-to-image model learn from two large, internally incompatible models, combining their separate strengths into one switchable network. In tests, it raised GenEval from 67.3 to 73.3 and DrawBench HPSv3 from 9.34 to 11.35, surpassing both larger teachers on composition.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central correctness lemma linking Eq. 3 and Eq. 4 is unproven and untested: if the teacher-refined target and the student's own continuation are not content-consistent, the DINOv2 loss supervises mismatched images and the on-policy claim collapses.","rationale":"The reader's weakest assumption identifies exactly the load-bearing technical gap: the absent appendix proof that the teacher's re-noised entry point (Eq. 3) and the student's cached entry point (Eq. 4) are content-consistent. I agree with that diagnosis and find it the most consequential issue in the paper. The manuscript contains a strong ablation suite and a novel adapter-sharing diagnostic, but the method's central claim--that supervision is genuinely on-policy across heterogeneous latent spaces--depends on an unstated and unverified first-order equivalence. The 'w/o warm start' collapse is consistent with the lemma being fragile, but it is not a test of the lemma; a direct cosine diagnostic would settle whether the loss compares semantically matched images. I also note secondary concerns that reinforce the conditionality but do not change the verdict: all benchmark numbers are single-run without error bars, no code or data are released, and the appendix is absent. For these reasons I retain the reader's CONDITIONAL verdict: the paper should be accepted only after the missing appendix is supplied, the content-consistency diagnostic is reported, and ideally multi-seed statistics are provided. This is not a rejection of the method's plausibility; it is a request for the evidence that the central argument requires.","tokens_in":13010,"tokens_out":4431,"duration_ms":51156,"concrete_test":"Run a diagnostic on the trained model: for 1000 prompts, perform one student rollout, then compute x_ref from Eq. (3) and x_tilde from Eq. (4) using the same initial rollout. Report the DINOv2 CLS cosine between x_ref and x_tilde, stratified by refinement depth r and by gap-aware sampling weight. Include two controls: (i) cosine between two independent student rollouts for the same prompt, and (ii) cosine between x_tilde and x_ref where the teacher refinement is initialized from a different student image of the same prompt. If the Eq. (3)-(4) cosine is close to control (i) and clearly above control (ii), the lemma is empirically supported. If it approaches control (ii), the loss is comparing unrelated content and the reported gains should be re-attributed to the warm-start stage alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2's on-policy objective rests on a specific content-consistency claim. The teacher target x_ref in Eq. (3) is produced by decoding the student sample x_S, re-encoding it with the teacher autoencoder E_e, adding noise to magnitude sigma_k^T, and refining with the frozen teacher. The student comparison x_tilde in Eq. (4) is produced by resuming Euler integration from the student's cached latent z_j^S, matched only by the inequality sigma_j^S >= sigma_k^T > sigma_{j+1}^S. These are not the same point: they are different latent codes of the same image in two incompatible autoencoders at only approximately matched noise magnitudes. The DINOv2 cosine loss is a meaningful on-policy signal only if x_ref and x_tilde have near-identical semantic content before the loss is applied, so that the loss measures the teacher's correction rather than a difference between unrelated images. The manuscript states 'Appendix shows that the two entry points ... are content-consistent to first order once the student is near convergence,' but no appendix is included in the submitted text. The warm-start ablation (w/o warm start: GenEval 49.2 vs 73.3) confirms that the pipeline depends on the student distribution being close to the teacher distribution, but it does not prove the lemma; it is equally consistent with the on-policy objective being semantically meaningless away from warm start. Without the missing proof or an empirical check, the claim that the method performs on-policy correction rather than a warm-start-dependent imitation of teacher samples is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Poly-OPD, a multi-teacher distillation framework for flow-matching text-to-image models whose teachers use different autoencoders and noise schedules. The method bridges the student and teacher latents through pixel space, applies DINOv2 CLS supervision in a common semantic space, uses a gradient-compatibility diagnostic to share attention LoRA while isolating per-teacher FFN adapters, and schedules compositional prompts by the remaining teacher-student gap. The main empirical claim is that distilling FLUX.1-dev and Z-Image into a 2.5B SD3.5-Medium student raises GenEval from 67.3 to 73.3 and DrawBench HPSv3 from 9.34 to 11.35, while enabling inference-time capability switching via adapter swaps. The paper includes ablations of warm start, adapter sharing, gap-aware sampling, refinement range, and perceptual feature space.","tokens_in":13427,"tokens_out":4923,"duration_ms":49312,"significance":"If the results hold, Poly-OPD is a meaningful step for multi-capability consolidation: it relaxes the shared-latent assumption that limits existing distillation methods, proposes a concrete pixel-bridge plus semantic-space supervision scheme, and grounds the adapter sharing boundary in a direct measurement rather than convention. The empirical study is unusually thorough in its ablations: the warm-start dependence (Table 4), the refinement-noise sweep (Table 5), and the feature-space comparison (Table 6) isolate the effect of each design choice, and the gradient-compatibility result (Figure 3) is a useful diagnostic in itself. The significance is conditional, however, because the correctness of the on-policy objective rests on an unproven content-consistency lemma, the headline numbers are reported without variance estimates, and the relationship between the composition training pool and the GenEval evaluation set is not made explicit.","major_comments":[{"comment":"The claim that the teacher-refined target x_ref and the student re-rolled output x_tilde are 'content-consistent to first order once the student is near convergence' is load-bearing for the entire on-policy objective, because the DINOv2 loss only measures the teacher's correction if the two entry points represent the same semantic content. The manuscript refers to an Appendix for the proof, but no appendix is included in the submitted text. The warm-start ablation (Table 4, GenEval 49.2 vs 73.3) is consistent with the lemma but does not establish it; it is equally consistent with the loss being semantically meaningless when the student is far from the teacher. Please include the missing proof or, at minimum, an empirical check that reports the DINOv2 cosine similarity between x_ref and x_tilde at matched noise levels at the trained checkpoint and across refinement depths r.","section":"Sec. 3.2, Eqs. (3)-(4)"},{"comment":"No error bars, confidence intervals, or multiple-seed results are reported for any of the main comparisons. The headline claims are point differences of roughly 6 GenEval points over the base student and 4 points over the Z-Image teacher, as well as HPSv3 differences of about 2 points; without variance estimates the reader cannot judge whether these differences are beyond run-to-run noise. Please provide at least three independent training runs (or bootstrap intervals over generated samples) for the full configuration and for the most important ablations, especially Tables 1, 2, and 4.","section":"Sec. 4, Tables 1-6"},{"comment":"The composition-mode training pool is described as the 'GenEval-style split of Flow-GRPO' while evaluation is performed on GenEval itself. The probe set is stated to be disjoint from both the training pool and the evaluation set, but the manuscript does not state that the training prompts are disjoint from the 553 GenEval evaluation prompts or that they are not paraphrases of them. If the training pool overlaps with or is derived from the GenEval test prompts, the reported GenEval improvement would partly reflect memorization rather than compositional generalization. Please clarify the exact relationship between the Flow-GRPO split and the GenEval benchmark, and confirm that no GenEval evaluation prompt appears in the training pool or probe set.","section":"Sec. 4.1, Table 2"},{"comment":"The row 'Student latent MSE' reports 'Collapse' without a numeric value, and the text says the loss 'collapses outright.' This is a qualitative claim in a quantitative table, and it is used to justify the choice of DINOv2. Please quantify the collapse: report the metric values, training curves, or a clear failure mode (e.g., NaNs, divergence, or blank images) so the reader can verify that the comparison is not an artifact of a mis-specified MSE implementation.","section":"Sec. 4.3, Table 6"},{"comment":"The gradient-compatibility diagnostic is computed at the warm-start initialization and with the Eq. (1) losses, and the resulting sharing-isolation boundary is then applied throughout the OPD stage. The manuscript does not establish that the gradient-compatibility pattern is stable as training progresses, even though the OPD objective differs from the warm-start objective. If the pattern shifts during training, the measured boundary may not be the optimal one. Please report the diagnostic at a later training point (or at least argue why the initial measurement is representative), since the adapter design is one of the central contributions.","section":"Sec. 3.3, Eq. (6)"}],"minor_comments":[{"comment":"The symbol lambda_WS in Eq. (5) is used without a prior definition; please define it in the notation block or near the objective.","section":"Sec. 3.2, Eq. (5)"},{"comment":"The figure labels 'Student VAE Dec.' and 'Teacher VAE dec.' while the text in Eqs. (2)-(4) uses E_S, G_S, E_e, G_e; please align the figure notation with the equations for consistency.","section":"Fig. 2"},{"comment":"The GenEval 'Single' column for Poly-OPD is listed as 100.0 against a base of 99.06; please confirm this is not a typo and, if correct, mention it explicitly as a ceiling effect.","section":"Table 2"},{"comment":"The capitalization of 'Z-Image' is inconsistent across the text, tables, and references; please unify it.","section":"Throughout"},{"comment":"The evaluation protocol promises that 'Full sampler, guidance, and schedule settings are in the supplement,' but no supplement is included in the submitted text; if this is an artifact of the submission format, please ensure the supplement is supplied with the revision.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"This is a strong empirical paper with a clear and useful framework, but the missing proof of the content-consistency lemma and the absence of variance estimates are serious gaps that need to be addressed before publication. The relationship between the Flow-GRPO training prompts and the GenEval evaluation prompts is a potential contamination concern that reviewers will likely pursue; the authors should be asked to rule it out explicitly. If the authors can supply the missing proof or empirical validation and add error bars to the key tables, the paper would be a solid acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading. The paper takes on a real gap: existing on-policy distillation assumes shared latents and noise schedules, and Poly-OPD is the first thing I've seen that makes heterogeneous teachers work through a pixel bridge plus frozen DINOv2 supervision. The adapter design—shared attention LoRA, teacher-specific FFN adapters, boundary chosen from a gradient-compatibility measurement—is a genuinely useful idea. The gap-aware sampler is sensible. The experiments are well structured: separate ablations for each component, and the headline result (2.5B student beating both larger teachers on GenEval, approaching FLUX on preference metrics) is exactly what you'd want from the method.\n\nThe soft spots are real but not fatal, mostly because they are fixable. The content-consistency lemma connecting Eq. (3) and Eq. (4) is load-bearing and it is parked in an appendix that is not in the submitted text. Without it, the DINOv2 loss could be comparing semantically mismatched images, and the on-policy claim is not established. The warm-start ablation (49.2 vs 73.3) is consistent with the authors' story, but it doesn't prove the lemma; it is also consistent with the objective being meaningless away from warm start. I'd want either the proof or an empirical check of DINOv2 similarity between x_ref and x_tilde during training. Second, all main numbers appear to be single-run. For a method whose gains are several points on GenEval, one seed is thin. Third, the compositional training prompts come from a GenEval-style split, so evaluation on GenEval is in-family even if the probe set is held out. That is not disqualifying, but worth stating plainly. No code or data is released, which matters for a distillation recipe.\n\nThe citation pattern is honest, the prior work is positioned correctly, and the parameters are reported rather than hidden. The missing appendix is the main technical objection; single-run reporting is the second. I'd send this to peer review and let the reviewers push for the proof or an empirical substitute, multi-seed numbers, and ideally code. If those land, this becomes a useful reference for anyone doing multi-teacher distillation.","headline":"Solid new method for a real gap, with one load-bearing lemma parked in a missing appendix.","tokens_in":13925,"tokens_out":2121,"would_cite":true,"duration_ms":20801,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One small model can absorb two incompatible teachers and beat both on their own benchmarks.","keywords":["multi-teacher distillation","on-policy distillation","flow matching","text-to-image generation","latent-incompatible models","LoRA adapters","compositional generation","gap-aware curriculum"],"falsifier":"Measure, at the end of training on a held-out prompt set, the DINOv2 cosine similarity between the teacher-refined target and the student's resumed rollout for the same prompt, seed, and noise magnitude; if the two are not consistently close, the aligned entry points that the distillation loss compares are semantically mismatched, and the pipeline's measured gains would not be attributable to on-policy correction.","tokens_in":12840,"feed_emoji":"🖼️","tokens_out":7157,"duration_ms":62245,"temperature":0.7,"pith_summary":"Poly-OPD claims that a single small text-to-image model can inherit the complementary skills of two large teachers that cannot directly communicate with each other, because the teachers use different autoencoders and noise schedules, and can switch between those skills at inference time by swapping a small adapter. The paper's route is on-policy distillation through a pixel bridge: the student's own image is decoded to pixels, re-encoded in the chosen teacher's latent space, refined by that teacher from a matched noise level, and then compared to the student's own resumed rollout in a frozen DINOv2 feature space. Two further mechanisms make the consolidation stick: a gradient-compatibility measurement decides which parts of the network to share across capabilities, and a gap-aware curriculum spends more training on compositional categories where the student still lags the teacher. Reported on benchmarks, the distilled 2.5B student raises GenEval from 67.3 to 73.3, above both larger teachers, and lifts DrawBench HPSv3 from 9.34 to 11.35.","feed_headline":"One small model absorbs two incompatible teachers and beats both","feed_subtitle":"Despite the teachers' incompatible internals, a 2.5B student wins on both benchmarks and switches skills with one small adapter.","key_machinery":"The load-bearing object is the pixel bridge with noise-magnitude alignment (Eqs. 2–4). Since the student and each teacher share no latent coordinate system, the student's own generated image is decoded from student latents to pixels, re-encoded with the active teacher's encoder, re-noised to a noise level matched in magnitude to a student-schedule level, and refined by the frozen teacher for r remaining steps; the refined image is compared, through a frozen DINOv2 CLS embedding, against the student's reroll of its cached trajectory from the same noise magnitude. The same DINOv2 embedding is what makes supervision canonical across models. The second mechanism is the gradient-compatibility diagnostic, computed on paired samples that share noise and differ only in teacher, which determines that attention LoRA gradients are aligned across modes while FFN gradients are not, dictating that attention LoRA be shared and FFN adapters be per-capability. The third is gap-aware sampling, where category weights follow the smoothed positive residual between cached teacher scores and rolling student probe scores, annealing back toward uniform as gaps close.","core_discovery":"The central claim is that heterogeneous on-policy distillation is possible: a flow-matching student can learn from teachers whose latent coordinates are mutually unintelligible, provided the supervision is routed through pixels and measured in a semantic space. In the paper's formulation, the student sample is decoded and re-encoded into the active teacher's autoencoder, re-noised to a level matched by magnitude (not by timestep index), and refined by the frozen teacher; the refined image is then a fixed target in DINOv2 CLS space, and the student is trained to make its own resumed rollout match that target. The paper further claims that the sharing–isolation boundary between capabilities should be set by direct gradient measurement, yielding shared attention LoRA modules and teacher-specific feed-forward adapters, and that training budget within the compositional mode should follow the remaining teacher–student gap rather than raw difficulty. With this machinery, a 2.5B SD3.5-Medium student surpasses both a 12B teacher and a 6B teacher on GenEval and improves preference metrics on DrawBench, which the authors read as evidence that on-policy training can exceed the teacher's own sample distribution.","pith_inferences":["The pixel-bridge scheme is not tied to diffusion or flow teachers: any generative model that can refine a decoded image, such as an autoregressive or GAN-based model, could in principle serve as an on-policy teacher for a flow-matching student as long as a semantic feature space like DINOv2 spans both.","The gradient-compatibility diagnostic could be reused as a general tool in multi-task and multi-teacher adaptation to decide, for any module type, whether to share parameters or isolate them, rather than fixing the boundary by architecture convention.","If the missing content-consistency lemma is confirmed, a testable prediction follows: the benefit of the on-policy stage over warm-start-only training should grow as the refinement depth r increases up to the point where the teacher fully regenerates the image, and should vanish if the teacher refinement is applied to semantically unrelated images.","The same framework invites a direct comparison against routing: a unified student with adapter switches versus a system that keeps both full teachers and routes prompts, where the measurable trade-off is benchmark quality per unit of memory and latency."],"forward_implications":["A compact student can consolidate capabilities from multiple large, architecturally incompatible teachers, so serving several specialists no longer requires deploying several full models.","Capability selection at inference can cost an adapter swap rather than a second model's memory and latency.","The refinement-depth knob r interpolates from off-policy imitation (full teacher rerun) to local on-policy correction, giving a single continuous trade-off between how closely the student tracks the teacher and training cost.","Because the student can surpass its teachers on aggregated or category-level benchmarks, teacher distillation need not be bounded by the teacher's own sample distribution.","Gap-aware sampling, driven by the measured teacher–student residual, automatically anneals toward uniform sampling as capabilities saturate, removing the need for an external curriculum schedule."],"supporting_citations":[{"why":"Defines the flow-matching objective that the student's velocity field is trained to satisfy in the warm-start stage.","marker":"(Lipman et al. 2023)"},{"why":"Introduces on-policy distillation for diffusion models, which this paper extends to heterogeneous teachers through the pixel bridge.","marker":"(Li et al. 2026)"},{"why":"Flow-OPD supplies the dense velocity-supervision view of on-policy distillation for flow teachers.","marker":"(Fang et al. 2026)"},{"why":"The frozen DINOv2 CLS embedding is the common semantic space in which cross-model supervision is applied.","marker":"(Oquab et al. 2023)"},{"why":"LoRA is the parameterization of the shared attention and per-capability FFN adapters.","marker":"(Hu et al. 2022)"},{"why":"GenEval is the compositional benchmark used to measure the student's two-object, counting, color, and position capabilities.","marker":"(Ghosh, Hajishirzi, and Schmidt 2023)"},{"why":"HPSv3 is the human-preference metric on DrawBench used to measure the preference-mode improvements.","marker":"(Ma et al. 2025)"},{"why":"FLUX.1-dev, the 12B preference teacher that supplies the aesthetics capability.","marker":"(Black Forest Labs 2024)"},{"why":"Z-Image, the 6B compositional teacher that supplies the counting, position, and attribute-binding capability.","marker":"(Cai et al. 2025)"},{"why":"SD3.5-Medium, the 2.5B flow-matching student that receives both capabilities.","marker":"(Esser et al. 2024)"}],"fun_headline_variants":["2.5B student beats 12B and 6B teachers on both benchmarks","Poly-OPD distills incompatible giants into one switchable small model","Pixel bridge + DINOv2 align teachers; student surpasses them","Small model masters two rival teachers via on-policy distillation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The two-stage pipeline rests on the unstated lemma that once the student is near convergence, the teacher's re-noised, re-encoded version of the student's image and the student's own cached latent state at the matched noise level describe the same image content to first order; the proof is deferred to an appendix that is not present in the submitted text.","fun_headline_variants_meta":{"raw":{"variants":["2.5B student beats 12B and 6B teachers on both benchmarks","Poly-OPD distills incompatible giants into one switchable small model","Pixel bridge + DINOv2 align teachers; student surpasses them","Small model masters two rival teachers via on-policy distillation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000587,"raw_usage":{"total_tokens":2827,"prompt_tokens":1085,"completion_tokens":1742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":1663}},"tokens_in":701,"tokens_out":1742,"duration_ms":13041,"temperature":1.0,"reasoning_tokens":1663,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:15:20.727159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, at the end of training on a held-out prompt set, the DINOv2 cosine similarity between the teacher-refined target and the student's resumed rollout for the same prompt, seed, and noise magnitude; if the two are not consistently close, the aligned entry points that the distillation loss compares are semantically mismatched, and the pipeline's measured gains would not be attributable to on-policy correction.","supporting_citations":[],"review_version":1}