{"id":"b8cfd5b9-3e3e-43c9-b958-d1c0ecb94384","arxiv_id":"2608.09637","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DUET relays between an sCM expert at the high-noise step and a DMD expert at the low-noise step, obtaining both diversity and quality in two-step video generation, and DUET+ adds RL-based adaptation.","lead":"DUET combines two video-distillation methods into a two-step pipeline, letting one expert preserve variety and the other add sharpness. The result reports much higher sample diversity than sharpness-first methods while keeping most of their visual quality, which matters because two-step video generation is fast enough for practical use.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DUET's diversity gain is confounded with the re-noising schedule; without a DMD-DMD relay control, the central expert-specialization claim is underdetermined.","rationale":"Good-faith reading: DUET is a simple, well-motivated design and the reported numbers are consistent with its stated mechanism. The toy experiment (Fig. 2), the PCA spread (Fig. 4), and the ablations (Tables 2-3) all add support. My concern is not that the method fails; it is that the central causal attribution—diversity comes from the sCM expert's high-noise role—is not isolated from the sampling schedule. The DUET sampler injects fresh noise at the relay (Eq. 6), which is a second stochastic input. DMD's native two-step sampler may not have this input. Therefore the 2.08x diversity gain over DMD could in principle be explained by the re-noising step alone. DP-DMD partially addresses this, but its first stage is ODE regression, so it does not control for the identity of the first-step expert. The proposed DMD-DMD control is cheap (inference-only) and decisive: if it reproduces DUET's diversity, the expert duet is not the cause; if it does not, the sCM expert is confirmed as the diversity source. This is complementary to the reader's weakest assumption about p_sCM_tau vs p_DMD_tau: the control removes the gap entirely, so it cleanly separates 'the DMD expert cannot handle relay latents' from 'the sCM expert is what preserves diversity.' I agree with the CONDITIONAL verdict: the paper should be accepted only if this control (or an equivalent attribution check) is provided, along with error bars. No change to the reader's verdict is needed, hence UNCHANGED.","tokens_in":17409,"tokens_out":13051,"duration_ms":125798,"concrete_test":"Run the DMD-DMD relay control on the same Wan2.1-T2V-1.3B setup: use the already-trained DMD expert for both steps, drawing x1 ~ N(0,I), computing xhat0 = g_DMD(x1,1,c), re-noising to x_tau = R_tau(xhat0, epsilon_tau) with tau=0.8, and outputting x0 = g_DMD(x_tau, tau, c). Evaluate with the exact pipeline of Table 1 (same prompts, R=5, ViCLIP/DINO/CLIP diversity, VBench quality). Compare DMD-DMD diversity and quality against DMD, sCM, DUET, and DUET+. Also compute bootstrap confidence intervals over prompts for the diversity averages; if the DUET-vs-DMD diversity gap is not significant, the headline claim weakens regardless of the control.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that assigning an sCM expert to the high-noise step is what preserves diversity in DUET. The reported experiments do not establish this, because the DUET sampler differs from the DMD baseline in two variables simultaneously: the identity of the first-step expert, and a two-step re-noising schedule (Eq. 6) that injects an independent Gaussian epsilon_tau at tau=0.8. This second variable is a source of stochasticity that a native two-step DMD rollout may not possess. The included baselines do not isolate it: sCM and DMD are native paradigms, rCM is a loss-level combination, and DP-DMD replaces the first step with ODE regression rather than with DMD. A DMD-DMD relay—same tau=0.8, same re-noising, same two network evaluations, but with the DMD expert used for both steps—would control for the schedule. If DMD-DMD diversity is comparable to DMD's 0.073, the sCM expert is the causal source; if it approaches DUET's 0.151, the diversity advantage is a schedule artifact and noise-level expert specialization is not the mechanism. This control also sharpens the distribution-gap concern: DMD-DMD has no train-inference gap (p_DMD_tau matches training), so it separates the relay-interface problem from the expert-identity question.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DUET, a two-step video-generation sampler that assigns a consistency-model (sCM) expert to the high-noise interval and a distribution-matching (DMD) expert to the low-noise interval, connected by a re-noising step at switch time tau=0.8. A second variant, DUET+, adapts the sCM expert with a reward-guided GRPO update (CM-GRPO) and the DMD expert by continuing its training on the relay latents. On Wan2.1-T2V-1.3B, the paper reports that DUET preserves roughly twice DMD's same-prompt diversity (average .1512 vs .0727) while raising VBench quality to 83.96 versus DMD's 84.38, and that DUET+ reaches 84.40 quality while retaining diversity .1521. The authors interpret these results as evidence that 'noise-level expert specialization' reconciles quality and diversity in the extreme few-step regime.","tokens_in":17748,"tokens_out":9896,"duration_ms":88376,"significance":"If the empirical claims hold, the paper offers a surprisingly simple design principle for the two-step regime: keep the two dominant distillation paradigms separate and let each act on the noise interval where its inductive bias is most beneficial. The work is significant because it directly challenges the prevailing loss-level combination strategy (e.g., rCM) and because the two-step setting provides a clean stress test of the quality-diversity trade-off. The paper is unusually explicit about its limitations: it acknowledges the relay-interface distribution gap (Eq. 9) and states that the public rCM checkpoint may underperform the official rCM. The evaluation uses external metrics (VBench and embedding-based diversity), and the key hyperparameter tau is ablated rather than used to force the result, so the headline differences are not derived from the method's own inputs.","major_comments":[{"comment":"The attribution of the diversity gain to the sCM expert's identity is underdetermined because the DUET sampler differs from the DMD baseline in two respects: the first-step expert and the intermediate re-noising time (tau=0.8). Since DMD-style multi-step sampling already injects fresh Gaussian noise at its intermediate step, the confound is not the presence of re-noising but the specific noise level used. The authors should run a DMD-DMD control with the same tau=0.8, the same re-noising operator (Eq. 4), and two network evaluations. If this control's diversity stays near DMD's .0727, the sCM expert is the causal source; if it approaches DUET's .1512, the reported diversity advantage is a schedule artifact and the central 'expert specialization' interpretation needs revision.","section":"§5.3 / Table 1; §4.2 Eq. (6)"},{"comment":"The primary quantitative claims lack confidence intervals or significance tests. The diversity gap between DUET (.1512) and DMD (.0727) is large, but the quality gaps are small: DUET (83.96) is within 0.4 points of DMD (84.38), and DUET+ (84.40) is essentially tied with DMD. Without standard errors or bootstrap confidence intervals over prompts/seeds, the reader cannot judge whether the claimed 'DMD-level quality' is distinguishable from noise. Please report uncertainty estimates at least for the headline quality and diversity averages in Table 1.","section":"§5.3 / Table 1"},{"comment":"The rCM comparison relies on a public checkpoint that, as the authors note, was not trained under the rCM paper's settings and may underperform the official rCM. This weakens the Q3 comparison against loss-level combinations. The authors should either train rCM on the same synthetic dataset under the same compute budget, or explicitly qualify all rCM numbers as lower bounds and temper the statement that loss-level combinations cannot achieve the DUET trade-off. The current caveat is disclosed but buried in the appendix; it should be reflected in the main-text claims about rCM.","section":"§5.1 / App. C.1"},{"comment":"The reward function for DUET+ (HPSv3) was selected after observing its effect on the same VBench quality dimensions that are reported as headline results in Table 1. This retrospective selection can inflate the reported DUET+ quality. Please clarify whether the evaluation was split into a validation set for reward selection and a held-out test set, and discuss the potential optimism in the DUET+ numbers. At minimum, state explicitly that the reward ablation is exploratory and not a test-set hyperparameter search.","section":"§5.4 / Table 3"}],"minor_comments":[{"comment":"The text says DUET's diversity is 'slightly lower' than sCM's, but Table 1 shows .1512 vs .1865, a 19% relative reduction. This is not 'slight'; please rephrase or provide a quantitative comparison.","section":"§5.3"},{"comment":"The column abbreviations SC, BC, TF, MS, DD, AQ, and IQ are not expanded in the table caption. Please expand them in the caption or add a footnote for self-contained readability.","section":"Table 1"},{"comment":"Please specify the exact native two-step sampling schedule used for the DMD baseline (intermediate time value and the distribution of the re-noising variable). This is needed to make the Table 1 comparison fully reproducible and to assess the confound discussed in the major comments.","section":"§5.1"},{"comment":"Equation (8) defines the curvature measure only up to proportionality. Please provide the exact normalization or formula used to generate Figure 3, so that tau=0.8 can be reproduced from the teacher trajectory.","section":"§4.2 / Eq. (8)"},{"comment":"The qualitative diversity evidence is presented through first-frame stills. Since the diversity metric is computed on video embeddings, consider adding a temporal-diversity metric or a short qualitative note on motion diversity, or explicitly state that first frames are only a proxy.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for the journal and the core empirical result is interesting. The main issue is the missing DMD-DMD control, which is a standard experimental isolation that would substantially strengthen the central mechanistic claim. I would also require uncertainty estimates for the small quality differences and a handling of the rCM baseline limitation. These are fixable within the paper's scope, so major revision seems appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this is the first paper I've seen that assigns different distillation objectives to different noise steps with independently trained experts, and it gets a real result on a real video backbone. DUET lifts sCM's two-step quality by ~2.5 VBench points while keeping most of its diversity, and DUET+ reaches DMD-level quality with diversity roughly twice DMD's. That's a meaningful win in a regime where the trade-off is sharp.\n\nWhat's genuinely new: the relay of an sCM expert for the high-noise step and a DMD expert for the low-noise step, trained separately, with no loss-weight to tune. The 2D toy experiment showing conflicting noise-to-data mappings is a nice conceptual justification. The Eq. (9) distribution-gap statement is honest, and the DUET+ adaptation—GRPO on the sCM expert, DMD continued on relay latents—is a clever way to handle the interface. The init-then-DMD ablation is exactly the right control to show that the diversity advantage isn't just a warm-start artifact.\n\nSoft spots, in order of severity. First, the rCM baseline is a public checkpoint the authors admit may underperform official rCM. That weakens the \"we beat loss-level combination\" claim, though the sCM/DMD comparison doesn't depend on it. Second, Table 1 has no confidence intervals; diversity averages of .151 vs .073 are large, but VBench point differences of 0.5 are not obviously significant. Third, the DUET+ reward is chosen from the same VBench dimensions used in Table 1; the ablation helps, but selection on the evaluation metric is still in play. Also, no code or checkpoints are released, which makes the numbers harder to check.\n\nThe stress-test note asks for a DMD-DMD relay control to separate expert identity from the re-noising schedule. I'd push back a little: DMD2's native multi-step sampling already uses re-noising between steps, so the schedule may not be a confound. But the paper never describes DMD's two-step rollout, and the control would be cheap. Without it, \"the sCM expert is the causal source\" is a reasonable inference, not a proven fact.\n\nBottom line: the central claim is probably right, but the evidence would be tighter with a DMD-DMD relay and error bars. This deserves a serious referee. I'd engage with it.","headline":"Noise-level expert duet is a simple, plausible way to dodge the quality-diversity trade-off in two-step video distillation, though missing a DMD-DMD relay control leaves the mechanism slightly underdetermined.","tokens_in":18233,"tokens_out":4630,"would_cite":true,"duration_ms":40584,"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":"DUET shows that in two-step video generation, relaying a diversity-preserving sCM step into a quality-oriented DMD step yields DMD-level quality with about twice DMD's diversity at identical inference cost.","keywords":["two-step video generation","few-step distillation","quality-diversity trade-off","consistency models","distribution matching distillation","noise-level expert specialization","reinforcement learning adaptation","Wan2.1"],"falsifier":"Measure the actual gap between the two relay distributions on the Wan2.1 backbone: generate a batch of intermediate latents $x^{\\text{sCM}}_\\tau$ by running the sCM expert and re-noising at $\\tau = 0.8$, and a batch of the DMD expert's native intermediate latents at the same noise level, then compute a feature-space distance such as MMD or Fréchet distance. If the gap is large where the base DUET relay still produces clean videos, the relay-interface explanation of the gains is wrong; if the gap shrinks after the DMD-side retraining in DUET+, the repair story is confirmed. A second decisive test is to rerun the whole DUET pipeline on a different backbone, such as Wan2.1-14B, and check whether the roughly two-fold diversity advantage over DMD survives.","tokens_in":17223,"feed_emoji":"🎬","tokens_out":10662,"duration_ms":82289,"temperature":0.7,"pith_summary":"Two-step video generation currently forces a choice: trajectory-level distillation (consistency models such as sCM) preserves diverse layouts but produces soft, blurry frames, while distribution-level distillation (such as DMD) renders sharp detail but collapses the diversity of compositions and motion. This paper claims that choice is unnecessary. DUET trains the two kinds of students independently and lets each take the step it is best at: the sCM expert handles the high-noise step that fixes structure and layout, and the DMD expert handles the low-noise step that refines appearance. On the Wan2.1 text-to-video model, this relayed two-step sampler reaches DMD-level quality while keeping about twice DMD's same-prompt diversity, and a follow-up reinforcement-learning adaptation (DUET+) retains that diversity edge while matching DMD's overall quality. If the claim holds, assigning different distillation objectives to different noise regimes is a simple way to get quality and diversity jointly in extreme few-step generation, with no fragile blending of conflicting losses.","feed_headline":"Two denoising experts jointly beat video's quality-diversity trade-off","feed_subtitle":"At two steps, an sCM expert lays out diverse structure and a DMD expert sharpens detail — so users get both.","key_machinery":"The load-bearing object is the noise-level expert duet: a relay sampler in which a consistency-model expert $f^{\\text{sCM}}_\\theta$ predicts a clean endpoint from pure noise, the prediction is re-noised to the switch time $\\tau$ via $R_\\tau(\\hat{x}_0, \\epsilon) = (1-\\tau)\\hat{x}_0 + \\tau\\epsilon$, and a DMD expert $g^{\\text{DMD}}_\\phi$ maps that intermediate latent to the final clean video. Because the two experts are trained independently with their native objectives, no tunable ratio between the consistency and distribution-matching losses is needed. The switch time $\\tau = 0.8$ is chosen from the curvature of the teacher's flow-matching ODE trajectory, which is high in the semantic-formation regime and low in the detail-refinement regime. The adaptation stage adds a second mechanism: CM-GRPO reinforcement learning applied to the sCM expert's Gaussian relay kernel, steering structures toward high-reward layouts, plus continued DMD training that uses the sCM expert's own relay latents as the backward simulation, repairing the distribution gap $p_{\\text{sCM}}^\\tau \\neq p_{\\text{DMD}}^\\tau$ that the paper flags as Eq. (9).","core_discovery":"The paper's central claim is that the quality–diversity trade-off between the two dominant few-step distillation approaches is not intrinsic to the two-step budget but a symptom of asking one set of parameters to realize two conflicting noise-to-data mappings. DUET instead splits the denoising schedule by noise level: a coverage-seeking sCM expert takes the high-noise interval $[1,\\tau]$, where layout, composition, and motion are decided, and a mode-seeking DMD expert takes the low-noise interval $[\\tau,0]$, where edges, texture, and color are refined. The two experts are trained separately with their native objectives, so no loss coefficient mediates between them; at inference they relay through a re-noising operator with switch time $\\tau = 0.8$ chosen from the curvature of the teacher's flow-matching trajectory. The measured outcome is that DUET reaches DMD-comparable quality (VBench aggregate 83.96 versus 84.38) while preserving more than twice DMD's same-prompt diversity (0.1512 versus 0.0727), and that DUET+, which steers the sCM expert with CM-GRPO and retrains the DMD expert on the sCM expert's actual relay latents, raises quality to DMD's level (84.40) while keeping the diversity advantage (0.1521).","pith_inferences":["A general pattern this result suggests: whenever one distilled expert consumes another expert's outputs, the consumer should be trained on the actual intermediate distribution it will see at inference; the paper's relay-latent retraining is the repair, and the same repair likely applies to any pipelined multi-expert sampler.","Curvature-based switch-time selection could be automated into an adaptive schedule for step budgets beyond two, or even chosen per prompt, rather than fixed at 0.8.","The same high-noise/low-noise role split could extend beyond text-to-video: image distillation, autoregressive video models, and other modalities with trajectory-level and distribution-level students of one teacher are natural testbeds.","One testable extension is to combine DUET's role split with explicit mode-collapse diagnostics on the DMD side; if the diversity gain is really structural, adding repulsive corrections should be unnecessary, whereas if it is fragile, such corrections would close the remaining gap."],"forward_implications":["At the same cost of two network evaluations, DUET delivers the sharp appearance of DMD and the diverse layouts of sCM together, so practitioners no longer have to choose between the two in the two-step regime.","Because the experts are trained independently with their native losses, the method needs no tuned ratio between consistency and distribution-matching objectives, avoiding the gradient conflicts that loss-level combinations such as rCM must fight.","The diversity that the duet preserves gives downstream preference optimization room to work: DUET+'s CM-GRPO steering and relay-latent DMD retraining improve quality while keeping the diversity gain, unlike the common init-then-DMD recipe whose inherited diversity collapses.","Switch time $\\tau = 0.8$ is selected from the teacher's ODE curvature, a backbone-agnostic criterion, so the same division of labor should transfer to other flow-matching video models without re-tuning."],"supporting_citations":[{"why":"Supplies sCM, the trajectory-level consistency-distillation method used as the diversity-preserving high-noise expert of the duet.","marker":"(Lu and Song 2025)"},{"why":"Supplies DMD2, the distribution-matching distillation objective and training setup used as the quality-oriented low-noise expert.","marker":"(Yin et al. 2024a)"},{"why":"Provides the Wan2.1-T2V-1.3B teacher backbone used for all distillation and evaluation, and the Wan2.2 hard-routed MoE design that motivates noise-level expert specialization.","marker":"(Wan et al. 2025)"},{"why":"rCM is the loss-level-combination baseline DUET must beat, and it also provides the training data and the JVP consistency kernel used to train the sCM expert.","marker":"(Zheng et al. 2025)"},{"why":"DP-DMD is the step-splitting baseline it is compared against, and its same-prompt diversity metric is the paper's evaluation of diversity.","marker":"(Wu et al. 2026a)"},{"why":"Supplies CM-GRPO, the consistency-model GRPO objective used in DUET+ to steer the sCM expert toward high-reward structures.","marker":"(Lu, Zuo, and Deng 2026)"},{"why":"Supplies HPSv3, the human-preference reward used as the default reward in the RL-guided expert adaptation.","marker":"(Ma et al. 2025)"}],"fun_headline_variants":["Split the denoising: one expert for layout, one for detail","DUET pairs distillation experts for two-step video quality and diversity","A duet of experts gives two-step video both quality and diversity","Two steps, two experts: DUET ends video's quality-diversity trade-off"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The relay works only if the intermediate latents that the sCM expert produces and that are re-noised at time $\\tau$ stay close enough to the latents the DMD expert saw during its own training; the paper flags this gap explicitly as $p_{\\text{sCM}}^\\tau \\neq p_{\\text{DMD}}^\\tau$, and the DUET+ adaptation exists precisely to repair it.","fun_headline_variants_meta":{"raw":{"variants":["Split the denoising: one expert for layout, one for detail","DUET pairs distillation experts for two-step video quality and diversity","A duet of experts gives two-step video both quality and diversity","Two steps, two experts: DUET ends video's quality-diversity trade-off"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3571,"prompt_tokens":1092,"completion_tokens":2479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":708,"completion_tokens_details":{"reasoning_tokens":2401}},"tokens_in":708,"tokens_out":2479,"duration_ms":16223,"temperature":1.0,"reasoning_tokens":2401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:37:11.261238+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual gap between the two relay distributions on the Wan2.1 backbone: generate a batch of intermediate latents $x^{\\text{sCM}}_\\tau$ by running the sCM expert and re-noising at $\\tau = 0.8$, and a batch of the DMD expert's native intermediate latents at the same noise level, then compute a feature-space distance such as MMD or Fréchet distance. If the gap is large where the base DUET relay still produces clean videos, the relay-interface explanation of the gains is wrong; if the gap shrinks after the DMD-side retraining in DUET+, the repair story is confirmed. A second decisive test is to rerun the whole DUET pipeline on a different backbone, such as Wan2.1-14B, and check whether the roughly two-fold diversity advantage over DMD survives.","supporting_citations":[],"review_version":1}