{"id":"9d989119-3461-4e61-8913-409f7ad5b7d4","arxiv_id":"2608.13205","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HPSD improves a TI2V model's text-to-video quality by having the student continue denoising from teacher-generated anchor states and match the teacher's velocity under an image condition.","lead":"This paper proposes HPSD, a self-distillation method that makes a text-image-to-video (TI2V) model better at text-only video generation by learning from its own outputs when it is given a high-quality first frame and an enhanced prompt. On WAN-2.2 and LTX-2.3, the authors report large gains on reward-model metrics and VBench, but without released code or statistical error bars.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HPSD's Eq. 10 re-imposes the teacher's clean first frame onto student-evolved frames, but the paper never shows this hybrid state is in-distribution; if the student's roll-out drifts, the teacher's velocity is corrupted and the method's core mechanism is unsupported.","rationale":"The reader identifies the same load-bearing assumption: the teacher's velocity at the mixed state of Eq. 10 is assumed to be a valid correction target, yet the paper does not measure whether this state is in-distribution. I agree this is the most load-bearing concern because it attacks the very mechanism that distinguishes HPSD from on-policy distillation. The paper's own Figure 3 and Section 3.2 condemn the mixed state of Eq. 5 as an invalid input; Eq. 10 constructs a similar mixture, albeit one that starts from teacher anchors. Whether the K-step student roll-out preserves enough alignment with the teacher's first frame is an empirical question that the paper leaves unmeasured. This matters even though the reported gains are large and consistent, because if the teacher's velocity is corrupted, the method's stated rationale collapses and the improvement may be attributable to other factors (e.g., the privileged-condition anchor trajectory acting as a data augmentation, or the simple act of training on teacher roll-outs). The reader's conditional verdict is appropriate: the central claim and the mechanism are plausible but not fully established until this check is done. I therefore recommend keeping the verdict unchanged, pending the concrete test above.","tokens_in":22048,"tokens_out":8371,"duration_ms":75794,"concrete_test":"During HPSD training on WAN-2.2, after the final step, compute the average latent-space distance between the hybrid supervision states \\hat{x}^Hyb_{t_{i+K}} (Eq. 10) and the teacher's own TI2V states at the same timesteps, \\hat{x}^Tea_{t_{i+K}} = [c_img, x^{Tea,(2:F)}_{t_{i+K}}], where the latter is the teacher trajectory advanced K steps from the same anchor state. If the mean cosine or L2 distance is large (e.g., comparable to distances between unrelated TI2V inputs) or if the teacher's velocity at the hybrid state deviates substantially from its velocity at its own state, the re-imposed first frame is misaligned and the teacher's supervision is corrupted. A small distance with high velocity agreement would validate the in-distribution assumption and clear this concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of HPSD is that the student's K-step roll-out from teacher anchors yields states that are both policy-aligned and content-consistent with the teacher's privileged first frame. At the supervision point (Eq. 10), the paper discards the student's evolved first frame and re-imposes the teacher's clean c_img, pairing it with the student's frames 2..F. This is structurally identical to the condition-state mismatch in Eq. 5 that the paper uses to justify rejecting on-policy distillation: a clean first frame from one video coexisting with noised frames from another. The only difference is that the student's roll-out started from the teacher's anchor, so the divergence may be small for small K, but the paper provides no measurement of this divergence or its effect on the teacher's velocity. If the student's policy drifts during training, the teacher's velocity at these mixed states is corrupted, and the supervision is misleading, reintroducing the very failure the paper claims to solve. The ablation K=3 over K=0 shows the roll-out matters, but it does not establish that the supervision target is valid; the improvement could arise despite the re-imposition, not because of it, with the teacher's privileged content entering through the anchor trajectory rather than through the velocity match at the hybrid state.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HPSD, a self-distillation framework for Text-Image-to-Video (TI2V) diffusion models, with the goal of internalizing the quality gains obtained from privileged conditioning (an enhanced prompt and a high-quality first frame) into the model's base text-to-video (T2V) mode. The teacher is the same TI2V model operating in TI2V mode, while the student operates in T2V mode. During training, the teacher first generates an off-policy anchor trajectory under privileged conditions; the student then rolls out a short K-step sub-trajectory from each anchor state using its own velocity field, and is supervised by the teacher at the resulting hybrid state, where the clean first frame is re-imposed. Experiments on WAN-2.2 and LTX-2.3 report consistent improvements over vanilla T2V, supervised fine-tuning, D-OPSD, and an on-policy distillation baseline on reward-model metrics and VBench.","tokens_in":22284,"tokens_out":3732,"duration_ms":33050,"significance":"The problem addressed is timely and well-motivated: unified TI2V models exhibit a large gap between their conditioned and base generation abilities, and distilling that ability into the base mode is practically valuable. The hybrid-policy idea—anchoring the student on teacher states while letting it evolve under its own policy—is a plausible and novel contribution that generalizes the off-policy/on-policy dichotomy. The paper provides two backbones, ablations on the key hyperparameter K, ablations on the privileged conditions, and a transfer study across first-frame generators, which is a reasonable experimental campaign. The main strengths are the clear formulation of the condition-state mismatch problem, the explicit algorithmic procedure, and the breadth of automatic metrics. However, the evidence for the central mechanism is incomplete: the validity of the hybrid state in Eq. (10) is assumed rather than demonstrated, the on-policy baseline is under-specified, and all quantitative claims rest on single-seed reward-model scores without human evaluation. These issues need to be addressed before the results can be considered robust.","major_comments":[{"comment":"The core supervision target in Eq. (10) is the hybrid state [c_img, x^{Hyb,(2:F)}_{t_i+K}], which re-imposes the teacher's clean first frame onto the student's evolved frames. This is structurally identical to the invalid mixed state in Eq. (5) that the paper uses to reject on-policy distillation, except that the frames come from a roll-out anchored at the teacher's trajectory. The paper never measures whether this hybrid state is in-distribution for the teacher, nor whether the teacher's velocity at this state is a meaningful correction target. If the student's roll-out drifts from the teacher's content, the teacher's prediction is corrupted exactly as in Eq. (5), and the stated advantage of HPSD over OPD disappears. Please provide evidence that the hybrid state is plausible, for example by measuring the distribution gap (e.g., Fréchet distance of the mixed state versus the teacher's own TI2V states), by showing that the teacher's velocity at the hybrid state is consistent with the student's content (e.g., via denoising trajectory alignment), or by an ablation that replaces the re-imposed clean first frame with the student's own evolved first frame.","section":"§4.1, Baselines"},{"comment":"The on-policy distillation baseline is listed without specifying its conditioning inputs. Equation (4) defines OPD with a single condition c, but in this paper the teacher operates under privileged conditions while the student operates under the vanilla prompt. It is therefore unclear whether the OPD baseline (i) queries the teacher with c_img and the enhanced prompt at student T2V states (thereby inducing the condition-state mismatch), (ii) conditions the teacher on the enhanced prompt only, or (iii) uses some other configuration. This ambiguity matters because the paper attributes the collapse of OPD (VideoAlign 0.2613 on WAN-2.2 versus 0.5335 for vanilla T2V) to condition-state mismatch; under alternative (ii) the baseline would likely not collapse. Please state the exact conditions used for the OPD teacher and, if the mismatch is intended, show that this is the natural adaptation rather than a straw-man configuration.","section":"§4.2, Tab. 1"},{"comment":"All quantitative results are reported as single point estimates, with no standard deviations, no number of seeds, and no significance tests. The headline claim that HPSD 'significantly improves' T2V performance rests on differences in reward-model scores that may be well within run-to-run variance, especially since training uses LoRA and 500 steps on 8 GPUs. In addition, Tab. 3(a) is used to select K=3, and the same evaluation prompt set (VideoDPO/VideoFeedback) appears to be used for both the ablation and the final reported numbers, raising the risk of selection on the test set. Please report means and variances over at least three independent runs, and clarify whether K is chosen on a held-out validation split that is disjoint from the final evaluation set.","section":"§4.3, Tab. 3(a)"},{"comment":"The loss in Eq. (11) uses the teacher's velocity at the hybrid state as the supervision target, but the paper does not analyze how the teacher improves during training when it is updated by EMA of the student. Since the teacher and student share weights, the teacher's velocity at the hybrid state is partly a function of the student's own policy. This self-referential loop is not circular in itself because the privileged conditions remain fixed, but the interaction between EMA updates and the hybrid state's distribution is not addressed. If the student policy drifts, the EMA teacher may gradually adapt to the student's distribution, weakening the anchoring effect. The paper should either provide a theoretical argument or an empirical diagnostic (e.g., plotting the teacher-student velocity discrepancy over training) to show that the hybrid supervision remains informative.","section":"§3.3, Eq. (11)"}],"minor_comments":[{"comment":"The paper contains several typos and formatting issues, such as 'W AN-2.2' with a space in Table 1 and Figure captions, and 'UR-v2-A, UR-v2-P' expansions that are not consistently defined. Please proofread the manuscript carefully.","section":"§4.1"},{"comment":"The training curve in Figure 9 is described as converging around 500 steps, but the x-axis extends to 800 steps and the text in §B.2 says 'negligible fluctuation thereafter' without specifying the evaluation protocol. Please clarify how the VisionReward score is computed during training and whether the evaluation prompts overlap with the training prompts.","section":"§B.2"},{"comment":"The citation 'Daniel Verdú, 2024' in the related work appears as an author name; the reference entry is formatted as a paper title. Please correct the citation style to match the author-year format used elsewhere.","section":"Related Work"},{"comment":"The evaluation set of 500 prompts is sampled from VideoDPO and VideoFeedback, but the paper does not specify whether any of these prompts also appear in the training set (Pref-GRPO). If there is overlap, the reported numbers may be optimistic. Please state the overlap or ensure disjointness.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for a computer vision journal and the central idea is publishable if the experimental evidence is tightened. The main risk is that the hybrid state in Eq. (10) may be no more valid than the mismatched state in Eq. (5); this is a load-bearing correctness question that needs to be resolved with an explicit distributional check or a targeted ablation. The under-specification of the on-policy baseline is also a fairness issue: without knowing whether the baseline used the privileged image condition, the comparison is hard to interpret. I would encourage the authors to report multi-seed results and to be transparent about how K was selected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper proposes a training recipe that improves a TI2V model's text-to-video mode by distilling from its own image-conditioned mode, and the hybrid-policy mechanism—student starts from teacher anchor states, rolls out K steps, then receives teacher velocity supervision—is a real extension beyond existing on-policy self-distillation work. Second, the headline numbers are large, but every table reports a single run without variance or seeds, so treat the effect sizes as provisional until confirmed.\n\nThe paper does several things well. The condition-state mismatch failure in on-policy distillation is explained clearly and convincingly: querying a TI2V teacher at student T2V states forces a clean first frame to coexist with unrelated evolving frames, corrupting the supervision. The paper's solution—re-noising the teacher's clean first frame to make student-compatible anchors, then letting the student evolve under its own policy—is mechanically sensible. The ablations on K and on privileged conditions are informative and support the core design choices. Two backbones, several reward models, and VBench give a reasonable breadth of evidence.\n\nThe soft spots are real but addressable. The stress-test concern about Eq. 10 lands: the paper re-imposes the teacher's clean first frame onto student-evolved frames, which is structurally similar to the mismatch it criticizes in Eq. 5. The difference is that the anchor start should keep the divergence small for small K, but the paper never measures whether this hybrid state is actually in-distribution. If the student drifts, the teacher's velocity at that mixed state is corrupted. The K=3 vs K=0 ablation shows roll-out length matters, but it does not isolate the validity of the supervision target. I would ask for a direct measurement of frame-wise divergence between the student's evolved frames and the teacher's own frames at the supervision step, or an ablation that keeps the student's first frame instead of re-imposing c_img. Also, the on-policy baseline is under-specified: exactly what conditions it queries the teacher with is not stated, which makes its collapse hard to interpret. No human evaluation and no released code yet are additional gaps.\n\nWho is this for: researchers working on video diffusion post-training, especially unified TI2V architectures. It deserves a serious referee, with requests for multi-seed results, baseline specification, and an analysis of the Eq. 10 distribution assumption. I would engage with it.","headline":"HPSD is a plausible and genuinely interesting hybrid-policy distillation idea with strong but under-validated numbers; the main risk is an unmeasured distribution assumption at the supervision state.","tokens_in":22894,"tokens_out":2263,"would_cite":true,"duration_ms":21719,"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":"This paper claims that HPSD lets a text-image-to-video model internalize the quality its privileged image-and-prompt mode elicits, and reproduce it from text alone via hybrid-policy self-distillation.","keywords":["text-image-to-video","self-distillation","hybrid-policy distillation","flow-matching diffusion models","video generation","condition-elicited capability","on-policy vs off-policy","velocity-level supervision"],"falsifier":"Re-run HPSD with Equation 10 altered so the first frame fed to the teacher is the student's own denoised first frame rather than the privileged $c_{\\mathrm{img}}$, keeping everything else fixed; if the reported VideoAlign gains vanish, the clean-frame re-imposition is what carries the method.","tokens_in":1816,"feed_emoji":"🎬","tokens_out":4079,"duration_ms":83246,"temperature":0.7,"pith_summary":"This paper asks whether a text-image-to-video model can learn a capability that only shows up when it is given extra privileged inputs — a high-quality first frame and a rewritten detailed prompt — and keep that capability when generating from a plain text prompt alone. The authors propose a self-distillation scheme in which the same model plays both roles: a teacher running in the privileged image-and-enhanced-prompt mode, and a student running in the base text-only mode. The scheme avoids two known failure modes: supervised fine-tuning on teacher outputs is off-policy, while naive on-policy distillation creates a mismatched state by mixing a clean first frame with the student's own rollout. On two TI2V backbones, the method improves text-only video quality on all reported reward metrics and also improves the image-guided mode, suggesting the model's base generation ability is genuinely strengthened.","feed_headline":"A model's image-guided mode can teach its text-only mode","feed_subtitle":"Raising VideoAlign about 3.5x on WAN-2.2 without extra video data or an external video teacher.","key_machinery":"The load-bearing object is the hybrid-policy sub-trajectory. Starting from an anchor state $x^{\\mathrm{Tea}}_{t_i}$ on the teacher's privileged TI2V trajectory, with the first frame re-noised so all frames share noise level $t_i$, the student evolves the state for $K$ steps under its own T2V velocity $v_\\phi(\\cdot \\mid c_{\\mathrm{txt}})$, producing $x^{\\mathrm{Hyb}}_{t_{i+K}}$. The teacher is then queried at $\\hat{x}^{\\mathrm{Hyb}}_{t_{i+K}} = [c_{\\mathrm{img}}, x^{\\mathrm{Hyb},(2:F)}_{t_{i+K}}]$ in its own TI2V input format, and the student matches the teacher's velocity on frames $2{:}F$. The length $K$ is the interpolation dial: $K=0$ gives off-policy supervision on the teacher trajectory, large $K$ approaches on-policy supervision, and $K=3$ balances anchoring with policy alignment in the reported experiments.","core_discovery":"The central claim is that the quality gap between a TI2V model's privileged (first-frame + enhanced-prompt) mode and its text-only mode is not a fixed property of the architecture: by self-distilling with a hybrid-policy objective, the model can absorb the privileged mode's content and motion priors into its own T2V weights. Concretely, HPSD makes the same TI2V model serve as teacher under privileged conditions and as student under vanilla T2V conditions, supervises the student on states that begin on the teacher's denoising trajectory and then evolve under the student's own velocity, and matches the teacher's velocity field on those hybrid states. In experiments this raises WAN-2.2's T2V VideoAlign from 0.5335 to 1.8753, against 1.2046 for supervised fine-tuning and 0.2613 for on-policy distillation, and also lifts the model's TI2V VideoAlign from 0.7831 to 1.2139.","pith_inferences":["A testable extension the paper does not report: compare the teacher's velocity error on hybrid states with its error on true TI2V rollouts at matched noise levels; if the error is much higher on hybrid states, the apparent gains may come from something other than trustworthy teacher correction.","The same anchoring idea could apply whenever one conditioning modality is privileged over another in a shared generative model, such as distilling reference-image or depth-conditioned quality into a text-only mode, provided the privileged condition can be converted into a student-compatible state.","The paper ablates K only on WAN-2.2; testing the K sweep on LTX-2.3 would show whether K=3 is a global optimum or a backbone-dependent choice.","Because the offline stage relies on auxiliary LLM and text-to-image generators, HPSD's ceiling should scale with those generators' quality; this is implied by the paper's additivity ablation but not directly measured across generator strengths."],"forward_implications":["Base text-to-video quality of a TI2V model can be improved substantially (VideoAlign 0.5335 to 1.8753 on WAN-2.2) without video-level human annotations or an external video teacher.","The same model's image-guided mode also improves after T2V-oriented distillation (VideoAlign 0.7831 to 1.2139), so the internalized capability is not purely a text-conditioning artifact.","K is a genuine interpolation knob: K=0 reduces to off-policy distillation and large K approaches on-policy supervision, with K=3 optimal in the reported experiments.","Privileged conditions act additively: distilling from an enhanced prompt plus a synthesized first frame outperforms first-frame-only guidance, so better auxiliary LLM and text-to-image tools should raise the achievable ceiling.","Naively applying on-policy distillation to TI2V models can collapse T2V quality (VideoAlign drops to 0.2613 on WAN-2.2), making the hybrid-policy correction necessary rather than optional."],"supporting_citations":[{"why":"Provides the WAN-2.2 TI2V backbone whose T2V and TI2V modes define the teacher and student roles in the main experiments.","marker":"Wan et al., 2025"},{"why":"Provides the LTX-2.3 backbone used to show the method transfers across model families.","marker":"HaCohen et al., 2026"},{"why":"Introduces D-OPSD, the on-policy self-distillation baseline HPSD is compared against, and motivates the EMA teacher update.","marker":"Jiang et al., 2026"},{"why":"Supplies the flow-matching velocity-field formulation and training objective that HPSD's velocity-level supervision builds on.","marker":"Lipman et al., 2022"},{"why":"DanceOPD, an on-policy distillation method whose teacher-query-on-student-state design HPSD diagnoses as condition-state mismatch.","marker":"Zhou et al., 2026"},{"why":"Provides the roughly 50K-prompt training set from Pref-GRPO used for HPSD and all compared baselines.","marker":"Wang et al., 2025b"},{"why":"VideoAlign, the primary reward-model metric on which HPSD reports its largest gains.","marker":"Liu et al., 2025"},{"why":"Z-Image-Turbo, the text-to-image model used offline to synthesize the privileged first frames.","marker":"Team, 2025"}],"fun_headline_variants":["Hybrid-policy distillation lets a video model teach itself","Self-distillation from image-guided mode boosts text-to-video quality","A TI2V model improves its own T2V via hybrid-policy self-distillation","Video model's privileged mode trains its text-only mode"],"cache_read_input_tokens":24960,"weakest_assumption_plain":"The method assumes that a video whose first frame is the teacher's clean reference image and whose later frames come from the student's own rollout is a valid input that yields trustworthy teacher guidance; if that mixed state is off-distribution for the model, the teacher's correction signal is corrupted.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid-policy distillation lets a video model teach itself","Self-distillation from image-guided mode boosts text-to-video quality","A TI2V model improves its own T2V via hybrid-policy self-distillation","Video model's privileged mode trains its text-only mode"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000793,"raw_usage":{"total_tokens":3567,"prompt_tokens":1093,"completion_tokens":2474,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":2399}},"tokens_in":709,"tokens_out":2474,"duration_ms":14534,"temperature":1.0,"reasoning_tokens":2399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:25:53.226028+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run HPSD with Equation 10 altered so the first frame fed to the teacher is the student's own denoised first frame rather than the privileged $c_{\\mathrm{img}}$, keeping everything else fixed; if the reported VideoAlign gains vanish, the clean-frame re-imposition is what carries the method.","supporting_citations":[],"review_version":1}