{"id":"c4f94fc0-2da0-4768-bb22-60c2a3fa03d9","arxiv_id":"2506.19348","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MotionEcho adaptively re-injects teacher-model guidance into few-step distilled video generators so reference motion can be copied at test time without training.","lead":"Distilled video generators are fast because they collapse many denoising steps into a few, but that same speed breaks existing reference-motion control methods. MotionEcho restores motion control by letting a slower teacher model selectively correct the fast model's trajectory during inference, without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) interpolation is the load-bearing step, but its derivation is internally inconsistent and its validity rests on an unverified distributional assumption.","rationale":"The reader identified the key load-bearing assumption: that teacher and student clean-latent predictions can be safely blended via Eq. (5). I agree this is the central technical risk. I add a concrete internal inconsistency in the derivation: the gradient of the stated squared ℓ2 loss introduces a factor of 2, so the displayed convex-combination form is not the actual gradient update unless a half factor or reparameterization is silently assumed. This does not by itself falsify the method, since the implemented update may use a redefined λ, but it makes the paper's formal justification unreliable and hampers reproduction. Other issues, such as unreleased code, per-motion-type threshold tuning, and missing error bars, are real but secondary; they affect confidence in the numbers rather than the logical soundness of the mechanism. The proposed test directly checks whether the interpolation step is load-bearing by comparing endpoint-only teacher guidance, interpolation, and student-only guidance, and by measuring whether blending reduces the teacher-student latent gap in a way that predicts motion-fidelity gains. If the interpolation is robust across that sweep, the central claim holds; if not, the method's main contribution would need to be reformulated. This is consistent with the reader's CONDITIONAL verdict, so I leave the verdict unchanged.","tokens_in":17820,"tokens_out":5648,"duration_ms":66972,"concrete_test":"Run the released implementation on TurboBench with a full sweep of λ ∈ {0, 0.1, 0.2, 0.3, 0.4, 0.5, 1.0} under fixed thresholds, and additionally record the mean distance ||zhat^ψ - zhat^θ|| at each activated teacher-guidance step. If λ = 1 (teacher-only) is not near-optimal, or if the gain from λ = 0.3 over λ = 0.1 is within the noise of repeated runs, then the specific interpolation weight is not doing the claimed corrective work. If, instead, teacher-only blending is best and large teacher-student distances coincide with larger motion-fidelity gains, the interpolation concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of MotionEcho is the convex blending of student and teacher clean-latent predictions in Eq. (5) and Algorithm 1, line 30. Two issues make this the weakest link. First, the score-distillation derivation is arithmetically inconsistent: with ℓ_distill = ||zhat^ψ - zhat^θ||^2, the gradient with respect to zhat^ψ is 2(zhat^ψ - zhat^θ), so the stated update gives zhat_new = (1 - 2λ)zhat^ψ + 2λzhat^θ, not (1 - λ)zhat^ψ + λzhat^θ. The paper does not define a 1/2 factor or reparameterize λ, so the exact teacher weight used in the reported experiments is ambiguous. Second, even with a corrected λ, the method assumes that zhat^ψ and zhat^θ are interchangeable estimates of the same clean latent, so their linear combination remains a valid clean latent. The teacher endpoint is produced by a different model, on a different schedule, after motion-guided denoising starting from a renoised student prediction. No distributional argument or quantitative check is provided that the blended latent lies on the student's latent manifold. If teacher and student predictions exhibit systematic offsets, the interpolation in Eq. (5) injects bias rather than correcting it. Since the reported gains over directly applying MotionClone to the distilled model are attributed to this interpolation (Table 3, rows 3 vs 5), the central claim depends on this unverified step.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MotionEcho is a training-free, inference-time framework for motion customization on distilled (few-step) text-to-video diffusion models. The paper argues that existing training-free motion guidance, such as MotionClone, fails on distilled models because large denoising steps discard the dense intermediate states that score guidance requires. The proposed method lets a slow, high-quality teacher diffusion model supervise the student's sampling trajectory: the student's one-step clean-latent prediction is re-noised onto the teacher's schedule, the teacher performs motion-guided denoising, and the resulting teacher clean-latent prediction is blended with the student's prediction via Eq. (5). An adaptive strategy decides when to invoke the teacher and how many inner steps to use, aiming to preserve the student's speed. Experiments are reported on T2V-Turbo-V2 (teacher VideoCrafter2) and AnimateDiff-Lightning (teacher AnimateDiff), with quantitative tables, ablations, qualitative comparisons, and a user study, claiming improved motion fidelity, text alignment, and FID at substantially lower inference time than applying MotionClone directly to the full teacher.","tokens_in":18163,"tokens_out":8471,"duration_ms":81633,"significance":"If validated, this is a practically relevant contribution: it targets a real gap, namely that training-free motion customization methods designed for full diffusion models fail on few-step distilled generators, and it reports large computational savings (e.g., 13 s versus 114 s for MotionClone in Table 1). The paper is also honest about limitations and provides a reasonable set of ablations, qualitative results across two student/teacher pairs, and a user study. The central claim is plausible. However, the core update rule in Eq. (5) has an arithmetically inconsistent derivation, the validity of linearly blending teacher and student clean-latent predictions is asserted rather than verified, and the empirical protocol is weakened by test-set hyperparameter tuning, single-run metrics without error bars, and no released code. The motion-fidelity metric is an external Co-Tracker-based score rather than the motion energy being optimized, so the method is not circular in the narrow sense, but the reported margins still need a stronger evaluation protocol to support the 'significantly improves' claim.","major_comments":[{"comment":"The stated update is inconsistent with the defined loss. With ℓ_distill = ||zhat^ψ_{0←t_{i+1}} − zhat^θ_{0←t_i}||², the gradient with respect to zhat^ψ is 2(zhat^ψ − zhat^θ), so zhat^ψ − λ∇ℓ gives (1−2λ)zhat^ψ + 2λzhat^θ, not the reported (1−λ)zhat^ψ + λzhat^θ. The manuscript does not introduce a ½ factor or reparameterize λ. Since λ is the central teacher-guidance strength and is ablated in Table 4, the exact teacher weight used in the experiments is ambiguous. Please correct the derivation, explicitly define the loss with the appropriate prefactor, and state the effective λ used for each reported configuration.","section":"Sec. 4.2, Eq. (5)"},{"comment":"The method assumes that after renoising the student endpoint onto the teacher's schedule and applying teacher motion guidance, the teacher's clean-latent prediction zhat^θ can be linearly blended with the student's zhat^ψ via Eq. (5) and remain a valid clean latent for the student's next sampling step. No distributional argument or empirical check is provided that the teacher and student estimates occupy compatible regions of latent space, nor that the blend does not inject systematic bias when the two models operate on different schedule granularities. Because Table 3 attributes a large part of the gain to this interpolation, please add quantitative evidence that the blended latent is plausible (e.g., per-step reconstruction error or latent-space FID/CLIP before and after blending, or a study of failure cases), or provide a theoretical justification for the linear combination.","section":"Sec. 4.2, Algorithm 1, lines 19-30"},{"comment":"The inner teacher denoising loop is not well-defined at its endpoint. If the loop reaches n = t − Δt_s without triggering the truncation condition on line 23, the else branch on line 26 updates z^θ_{n−Δt_t} = z^θ_{t−2Δt_s}, so no latent at the interval endpoint t − Δt_s is ever produced; line 29 then computes zhat^θ_{0←t−Δt_s} from a state that is either undefined or outside the intended interval. Only the truncation branch on line 24 sets the endpoint latent. Please correct the loop bounds or the final update so that the teacher always returns a prediction at the intended sub-interval endpoint, and clarify whether the loop is meant to be inclusive or exclusive of t − Δt_s.","section":"Algorithm 1, lines 20-28"},{"comment":"The empirical protocol leaves the reported numbers vulnerable to selection on the test set. Appendix B.3 states that the hyperparameters (including λ, η, k, δ1, δ2, and related budget parameters) are determined via grid search, and the search is carried out on the same TurboBench and AnimateBench benchmarks used for final evaluation. In addition, Tables 1-3 report single-run metrics without error bars or significance tests, and the benchmarks are small (66 video-edit pairs derived from 22 unique videos; 15 videos on AnimateBench). Please provide a clear validation/evaluation split, report variance across multiple runs or seeds, and state how the thresholds were selected without access to the test set.","section":"Sec. 5.1, Appendix B.3, Tables 1-3"}],"minor_comments":[{"comment":"The motion loss G_m is missing a closing parenthesis or bracket: the expression ||M_ref_{tα} · (A(z_ref_{tα}) − A(z_{t_i+1})||² is unbalanced. Please fix the notation so that the norm applies to the full difference term.","section":"Sec. 4.1, motion loss definition"},{"comment":"The notation 'Select inner teacher step s ∈ (t, t−Δt_s)' is confusing because the interval is written in decreasing order; it should be s ∈ (t−Δt_s, t) or stated as 's between t−Δt_s and t'.","section":"Algorithm 1, line 19"},{"comment":"Numerical precision is inconsistent (e.g., 0.9420 vs. 0.925 in Table 1), and several captions contain typos, such as 'ContolVideo' and 'Contol-A-Video' in Figure 16. Please proofread the final version.","section":"Tables 1-2 and figure captions"},{"comment":"The arXiv listing title is 'When Distillation Breaks Motion Control: Restoring Generative Trajectories for Fast Video Generators' while the manuscript body uses 'Training-Free Motion Customization for Distilled Video Generators with Adaptive Test-Time Distillation'; please ensure the published version uses one consistent title.","section":"Title consistency"}],"recommendation":"major_revision","confidential_remarks":"The paper is an arXiv preprint with no released code, and the central update rule contains a derivation ambiguity that affects the exact experimental configuration. I recommend requesting the code, the precise hyperparameter settings, and a validation/evaluation split before considering acceptance. The title inconsistency between the arXiv listing and the manuscript should also be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first paper I've seen that makes training-free motion customization work on few-step distilled video models, and the adaptive teacher scheduling is a clever way to keep cost down. The empirical story is mostly coherent. But the derivation of Eq. (5) has a real algebraic slip, and the evaluation is on small benchmarks with hyperparameters tuned on the same test set, so I'd want a corrected version and code before relying on the numbers.\n\nWhat's genuinely new: the combination of renoising the student's predicted clean latent onto the teacher's denser trajectory, running motion guidance there, and interpolating the teacher's clean prediction with the student's. The step-wise activation of guidance based on a moving average of motion loss is a sensible efficiency trick. Ablations in Table 3 and supplementary show each component adds something, and the user study is a plus.\n\nWhere it's soft: the score-distillation derivation in Sec. 4.2 doesn't hold. With ℓ_distill = ||zhatψ − zhatθ||², the gradient w.r.t. zhatψ is 2(zhatψ − zhatθ), so the gradient update gives zhat_new = (1−2λ)zhatψ + 2λ zhatθ, not (1−λ)zhatψ + λ zhatθ as written. That's not automatically fatal—you can reparameterize λ to absorb the 2—but as written it makes the reported λ in Table 4 ambiguous. The paper should either fix the formula or explicitly say the interpolation weight is a free hyperparameter. Relatedly, the method assumes student and teacher clean latents can be linearly blended, and no distributional check is offered. That's a common pragmatic assumption in this literature, but here it's the load-bearing step, so it deserves discussion.\n\nThe evaluation is the bigger practical issue. Benchmarks are 66 and 15 videos; no error bars; δ1, δ2, λ, η, k are tuned by grid search on the evaluation set, and τ (the teacher guidance schedule ratio) is never specified in the experimental section. The paper says code will be released upon acceptance, but for a training-free inference-time method I'd expect code or at least enough parameter detail to reproduce.\n\nWho it's for: people working on video editing, distillation, or test-time guidance. If the derivation is fixed and code appears, it's a citable subfield contribution. I'd send it to review, but only with major revision and a request for artifacts and a cleaned-up derivation.","headline":"A useful test-time distillation trick for motion control on fast video generators, but the core interpolation derivation is sloppy and the evaluation is too self-contained to fully trust.","tokens_in":18668,"tokens_out":4580,"would_cite":false,"duration_ms":46762,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Distilled video generators can be given reference-motion control at inference time by re-noising their predictions onto a full-diffusion teacher's trajectory.","keywords":["test-time distillation","motion customization","distilled video diffusion models","teacher guidance","adaptive acceleration","training-free methods","temporal attention maps","video generation"],"falsifier":"Measure, at a fixed student step and across many prompts, the distribution of teacher-guided clean latents versus student-only clean latents (for example, per-channel mean and variance or a latent FID). If after re-noising the teacher predictions lie far from the student's latent manifold, the interpolation in Eq. (5) should systematically distort generations; additionally, an ablation that replaces the teacher's motion-guided prediction with the teacher's unguided prediction at the same re-noised point should erase the reported motion-fidelity gain if the claimed mechanism is what causes it.","tokens_in":17650,"feed_emoji":"🎬","tokens_out":8580,"duration_ms":84380,"temperature":0.7,"pith_summary":"The paper tackles a practical mismatch: motion customization, which copies a reference video's movement into a generated video, was developed for full diffusion models that denoise in many small steps, and it fails when applied to the new distilled few-step video generators. It argues the failure is structural, because distillation collapses many denoising steps into a few and changes the denoising behavior, so the dense intermediate states that motion-score guidance relies on no longer exist. To fix this, the paper proposes MotionEcho, a training-free inference framework in which a slow, high-quality diffusion teacher supervises the fast student: the student's predicted clean endpoint is re-noised onto the teacher's dense trajectory, the teacher applies motion-guided denoising there, and its clean prediction is blended with the student's. The claim is that this restores accurate motion control on distilled generators in a handful of student steps, with an adaptive schedule deciding when teacher help is worth the extra compute.","feed_headline":"Teacher echo restores motion control for distilled video models","feed_subtitle":"Re-noising a few-step output onto a full teacher trajectory restores reference motion in 13 seconds","key_machinery":"The load-bearing mechanism is teacher-forced endpoint blending: the student's coarse one-step prediction is re-noised onto the teacher's fine-grained denoising schedule, the teacher performs motion-guided refinement there, and the teacher's refined clean prediction is averaged into the student's endpoint through Equation (5). The second mechanism is the adaptive acceleration policy, which decides per student step whether teacher guidance is worth running and how many inner teacher denoising iterations to spend, based on motion-loss thresholds and a maximum iteration cap.","core_discovery":"The central claim is that existing training-free motion-customization methods, such as MotionClone, are incompatible with distilled video generators not because the motion loss is weak but because the generative trajectory has changed: the distilled model's large denoising intervals omit the dense intermediate states that temporal-attention motion guidance needs, and its denoising behavior no longer matches the teacher's. MotionEcho addresses this with test-time distillation. At each student step, the student predicts a clean latent $\\hat z^\\psi_{0\\leftarrow t_{i+1}}$; this latent is re-noised to an intermediate teacher timestep $t_s$, the teacher runs motion-guided denoising down to $t_i$, and the teacher's one-step clean prediction $\\hat z^\\theta_{0\\leftarrow t_i}$ is linearly blended with the student's prediction, $\\hat z^{\\mathrm{new}}_{0\\leftarrow t_{i+1}} = (1-\\lambda)\\hat z^\\psi_{0\\leftarrow t_{i+1}} + \\lambda \\hat z^\\theta_{0\\leftarrow t_i}$, before updating the student latent. An adaptive scheduler triggers teacher guidance only when a moving-average motion loss exceeds a threshold and truncates the teacher's internal loop when its motion loss drops below a second threshold. The paper reports that on TurboBench this yields the best motion fidelity and FID among compared methods with a 16-step T2V-Turbo-V2 student running in 13 seconds, versus 114 seconds for MotionClone.","pith_inferences":["If the bottleneck is schedule granularity rather than the specific motion loss, the same re-noise-and-blend correction could apply to other test-time controls on few-step generators, such as depth, pose, or identity guidance.","The method's reliance on hybrid reference-noise initialization suggests that motion control on fast generators may generally need some preservation of the reference's temporal dynamics, not just endpoint adjustment.","The hand-set thresholds $\\delta_1$ and $\\delta_2$ could be learned or made adaptive per step, which would remove the main manual tuning burden and potentially widen the operating range.","Since the teacher correction is applied at inference, one could attempt to distill these corrections back into the student's weights during use, converting test-time distillation into an online self-improvement loop."],"forward_implications":["Training-free motion customization becomes feasible on 4-16 step distilled video generators, not just full-diffusion models.","The teacher-student recipe transfers across base-model pairs, with T2V-Turbo-V2 guided by VideoCrafter2 and AnimateDiff-Lightning guided by AnimateDiff, so the approach is not tied to one architecture.","At 16 student steps the reported inference cost is 13 seconds versus 114 seconds for MotionClone on the same benchmark, while motion fidelity and FID improve; at 8 or 4 steps the method remains competitive.","Because guidance is activated only when the student's motion loss is high, most runs do not pay the full teacher cost, which is why distilled efficiency is mostly preserved."],"supporting_citations":[{"why":"Provides the T2V-Turbo-V2 distilled student model on which the main experiments run.","marker":"[1]"},{"why":"Provides the AnimateDiff-Lightning distilled backbone used to show generality.","marker":"[2]"},{"why":"Supplies the MotionClone motion loss and temporal-attention prior that both teacher and student use for motion guidance.","marker":"[14]"},{"why":"Is the VideoCrafter2 teacher model whose dense denoising trajectory supervises the student.","marker":"[42]"},{"why":"Is the DMT zero-shot motion-transfer baseline and the source of the motion-fidelity evaluation protocol.","marker":"[15]"},{"why":"Is the DDIM inversion used to obtain the reference latent and hybrid noise initialization.","marker":"[21]"},{"why":"Underlies the consistency-model distillation that the fast student generators are built on.","marker":"[19]"}],"fun_headline_variants":["MotionEcho re-teaches distilled video generators motion control","Teacher guidance corrects distilled denoising for motion fidelity","Distilled video motion fixed by lightweight adaptive teacher echo","Training-free motion customization now works on distilled models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a teacher's cleaned-up version of the student's output, obtained after re-noising, is a trustworthy correction that can simply be averaged with the student's prediction; if the two predictions live in incompatible parts of the latent space, the blend would inject systematic bias instead of fixing motion.","fun_headline_variants_meta":{"raw":{"variants":["MotionEcho re-teaches distilled video generators motion control","Teacher guidance corrects distilled denoising for motion fidelity","Distilled video motion fixed by lightweight adaptive teacher echo","Training-free motion customization now works on distilled models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1802,"prompt_tokens":1124,"completion_tokens":678,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":740,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":740,"tokens_out":678,"duration_ms":7179,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:06:56.495126+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, at a fixed student step and across many prompts, the distribution of teacher-guided clean latents versus student-only clean latents (for example, per-channel mean and variance or a latent FID). If after re-noising the teacher predictions lie far from the student's latent manifold, the interpolation in Eq. (5) should systematically distort generations; additionally, an ablation that replaces the teacher's motion-guided prediction with the teacher's unguided prediction at the same re-noised point should erase the reported motion-fidelity gain if the claimed mechanism is what causes it.","supporting_citations":[],"review_version":1}