{"id":"13da72d3-6d0d-4ac3-ac92-713e81b68750","arxiv_id":"2607.09581","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hierarchical diffusion framework generates stable, music-synchronized 720p/30fps dance videos longer than a minute by planning sparse keyframes globally and filling them in locally.","lead":"Wan-Dancer uses a two-stage diffusion framework—global sparse-keyframe planning plus local interpolation—to create dance videos over a minute long at 720p/30fps from music. The framework is aimed at long-form, audio- and text-conditioned dance generation, which current diffusion video models generally cannot sustain beyond roughly 20 seconds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inference assembly as written cannot produce claimed durations: 38 keyframes x 149-frame clips concatenated = 188.7 s regardless of input, with heavy overlap and no stitching rule specified.","rationale":"The reader's weakest assumption—Eq. (3)'s optical-flow loss—is a valid and serious issue: multiplying only the predicted velocity u by a continuous VAE latent w_optical_flow makes the optimum u = v_t / w, not a weighting. That could corrupt training. However, the single most load-bearing concern for the paper's central claim is the inference assembly protocol, because the claim is specifically about generating stable, 720p/30fps videos exceeding one minute. If the described numbers (38 keyframes, 149-frame clips, concatenation) are taken literally, the output length is fixed at ~188.7 s and contains massive overlap for typical minute-scale inputs, so the pipeline as written cannot produce a coherent video of the input music's duration. A charitable reading—that the 149-frame clips are just local context windows and only the intervals between keyframes are kept—could rescue the method, but that step is not stated. This is an internal arithmetic inconsistency in the central method, not a disagreement with external consensus. The concern does not overturn the reader's CONDITIONAL verdict; it reinforces it. The authors need to specify the exact stitching rule and ideally release code/data to verify both the assembly and the loss function. I therefore leave the verdict unchanged at CONDITIONAL, while flagging that the assembly gap is the more fundamental threat to reproducibility of the headline result.","tokens_in":10655,"tokens_out":8974,"duration_ms":103704,"concrete_test":"Implement a mock of the described inference pipeline: for a 60-s and a 160-s input at 30 fps, place 38 keyframes uniformly, form 149-frame clips centered on each keyframe, concatenate in order, and compute output duration and duplicate-frame overlap. If output duration != input duration or adjacent clips overlap without a specified trimming rule, the assembly description is insufficient/incorrect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is an internal inconsistency in the inference assembly protocol (§4.1, Fig. 2). The paper states that the global stage always produces 38 keyframes, that these keyframes 'segment the full sequence into clips of 149 frames centered around each keyframe,' and that 'all generated local video clips are concatenated in order.' For any input duration, concatenating 38 clips of 149 frames at 30 fps yields a fixed output of 38 × 149 / 30 ≈ 188.7 s. For a 60 s input this is roughly 3.1× the music length; for the 160 s example in Fig. 3 it is 188.7 s vs 160 s. Moreover, when keyframes are evenly distributed over D, adjacent 149-frame clips overlap heavily—for D = 60 s, keyframe spacing ≈ 48.6 frames, so adjacent clips overlap by ≈ 99 frames. Naive concatenation would therefore repeat large portions of the dance, producing a stutter/loop rather than a coherent minute-scale video. The text never specifies a trimming or stitching rule (e.g., taking the interval between consecutive keyframes, or using each clip only as a local context window). Without such a rule, the described pipeline cannot be executed to produce a video of the input music's duration, making the central claim—'minute-scale coherent' generation—unreproducible from the method description. This is a concrete arithmetic/algorithmic gap, not a matter of external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Wan-Dancer, a hierarchical diffusion framework for music-to-dance video generation. A global stage generates 38 sparse keyframes from full-track music context, and a local stage generates 149-frame clips centered on each keyframe, which are then concatenated. Additional components include dynamic frame-rate RoPE time embeddings, an optical-flow-weighted loss, motion-speed stratification, and LoRA-based choreography customization. The paper claims stable 720p/30fps videos exceeding one minute across five dance genres, with subjective evaluations against X-Dancer and MusicInfuser.","tokens_in":11115,"tokens_out":4198,"duration_ms":47787,"significance":"The hierarchical global-to-local idea is timely and plausible, and the dynamic frame-rate RoPE mechanism is an interesting design choice. If the pipeline actually produced minute-scale coherent videos, this would be a meaningful advance over typical short-window video diffusion. However, as written, the inference assembly is internally inconsistent and cannot reproduce the claimed durations, the optical-flow loss in Eq. (3) is not a valid weighted loss, and the quantitative evaluation relies on mean subjective scores without variance or significance tests. These are load-bearing issues, not presentation concerns. The paper also reports honest limitations (identity consistency, semantic alignment, multi-dancer), which is commendable, but the central claim is currently not reproducible from the method description.","major_comments":[{"comment":"","section":"§4.1, Model Inference Details and Fig. 2"},{"comment":"","section":"§3.2, Eq. (3)"},{"comment":"","section":"§4.3, Tables 1–3"},{"comment":"","section":"§4.3, Ablation Studies and Figs. 6–7"}],"minor_comments":[{"comment":"","section":"Table 1"},{"comment":"","section":"Fig. 2"},{"comment":"","section":"§4.1, Dataset"},{"comment":"","section":"§4.1 and Fig. 7(a)"},{"comment":"","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper relies on the authors' own backbones (Wan-I2V/Wan-S2V, refs [7,29]) and a proprietary dataset, and no code or data are released. This is not by itself disqualifying, but it raises the bar for the manuscript's internal consistency, which the inference-assembly gap and Eq. (3) currently fail to meet. The fit with a computer-vision venue is appropriate; the issue is execution, not scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about Wan-Dancer before spending time on it. First, the inference protocol in §4.1 cannot produce the claimed minute-scale output: the global stage always emits 38 keyframes, each keyframe seeds a 149-frame clip, and the paper says the clips are concatenated in order. That gives 38 × 149 / 30 ≈ 188.7 s regardless of music length, and for a 60 s input the 149-frame clips centered on evenly spaced keyframes overlap by roughly 99 frames, so naive concatenation would repeat content. No trimming or stitching rule is given. I checked the text: 'concatenated in order' is all there is. That is a concrete, arithmetic gap in the central claim.\n\nSecond, Eq. (3) multiplies the predicted velocity by the VAE-encoded optical-flow latent and subtracts the ground truth velocity. If that latent is continuous, the minimizer is u = v_t / w, not v_t; if it is supposed to be a mask, the loss should weight (u - v), not u. Either way, the objective as written does not do what the text says it does.\n\nWhat the paper does well: the global-to-local decomposition is a sensible way to attack the 20-second ceiling, and time-mapped RoPE for dynamic FPS is a clean adaptation of Qwen2.5-VL's absolute-time encoding. The five-genre proprietary dataset and the LoRA fine-tuning for specific choreographies are useful engineering contributions, and the qualitative figures show plausible detail at 720p.\n\nBut the evaluation is thin: mean subjective scores in Tables 1–3 come with no variance, no rater count, no significance tests, and the ablations are qualitative. Data and code are not released. None of that by itself makes the method wrong, but combined with the pipeline arithmetic and the loss equation, the paper is not yet reproducible.\n\nMy read: the system may well work, but the write-up has two load-bearing problems that need fixing before the claim 'exceeds one minute' can be taken seriously. I would send it to reviewers because the topic is timely and the method has merit, but I would expect a request for a corrected protocol, a properly specified loss, and real statistical reporting before any acceptance.\n\nRecommendation: engage with it as a cautionary example, not as a reliable method yet.","headline":"Plausible hierarchical framework for minute-scale dance video, but the inference protocol as written cannot produce the claimed durations and the optical-flow loss is mis-specified; evaluation is thin.","tokens_in":11522,"tokens_out":3861,"would_cite":false,"duration_ms":40214,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hierarchical global-to-local pipeline lets a diffusion model generate coherent dance videos of over a minute from music alone.","keywords":["music-to-dance generation","video diffusion","hierarchical generation","keyframe planning","optical flow loss","dynamic frame rate","RoPE embeddings","long-form video synthesis"],"falsifier":"Compute w_optical_flow for typical training frames: if it contains values other than 0 and 1, the loss in Eq. (3) does not weight the velocity error as claimed, and the reported optical-flow ablation could be re-run with a true mask to see if the gains vanish.","tokens_in":10581,"feed_emoji":"💃","tokens_out":2866,"duration_ms":29219,"temperature":0.7,"pith_summary":"The paper aims to break the 20-second ceiling on music-to-dance video generation by splitting the task into a global keyframe planner and a local frame refiner. It claims this hierarchical decoupling, together with time-aware position embeddings, an optical-flow loss, and motion-speed control, yields stable 720p/30fps dance videos over a minute long across five genres. A sympathetic reader would care because long-form coherent dance video from audio has been out of reach; if the claim holds, the approach removes the main temporal bottleneck for end-to-end dance synthesis.","feed_headline":"Generates minute-long dance videos from music","feed_subtitle":"A global keyframe planner plus local refiner beats the 20-second ceiling for music-to-dance video.","key_machinery":"The load-bearing mechanism is the keyframe mask plus dynamic frame-rate injection: a mask that is all-zero except for sparse anchor frames tells the shared diffusion transformer whether to plan the global structure or refine local motion. Time-mapped RoPE embeddings convey absolute time, letting the same network operate at 3–15 fps and align keyframes to any music duration. The optical-flow loss and motion-speed stratification are supporting controls for detail during fast motion.","core_discovery":"The central claim is that minute-scale coherent music-to-dance video is achievable by a global-to-local decomposition: a sparse global stage plans 38 keyframes across the whole track, and a local stage fills in dense frames around each keyframe, with absolute time injected through time-mapped RoPE embeddings so the model knows the true temporal spacing. The paper reports that this produces temporally stable, identity-consistent videos lasting 60–160 seconds, with rhythmic alignment and genre control.","pith_inferences":["The global keyframe stage might be reusable as a standalone choreography planner, decoupling motion semantics from rendering.","The optical-flow loss formulation, as written, has an ambiguity that could make the reported gains partly an artifact; a corrected weighting may be needed.","The hierarchical scheme suggests a general recipe for taming diffusion drift: plan sparsely, refine locally in parallel; could apply to other long-form conditional generation such as narrated action sequences."],"forward_implications":["If the claim is right, music-to-dance generation no longer needs 3D skeletons or short clips; end-to-end video synthesis can cover full songs.","The global keyframes can serve as a parseable choreography sketch, enabling editing and customization (LoRA) without retraining.","The dynamic frame-rate trick may extend to other long video generation tasks, not just dance.","The measured gains over prior end-to-end methods suggest beat-aligned, prompt-controlled dance video could become a practical creative tool."],"fun_headline_variants":["Global keyframe planning yields minute-long music-to-dance videos","Two-stage hierarchy overcomes diffusion limit for dance videos","Beat the 20-second barrier with hierarchical music-to-dance","Minute-scale dance videos via global plan and local refinement"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The optical-flow term in the training loss is assumed to act as a binary mask that focusses the velocity error where motion is large; if it is a continuous multiplier, the loss's optimum would rescale the target velocity and corrupt the very motion it is meant to protect.","fun_headline_variants_meta":{"raw":{"variants":["Global keyframe planning yields minute-long music-to-dance videos","Two-stage hierarchy overcomes diffusion limit for dance videos","Beat the 20-second barrier with hierarchical music-to-dance","Minute-scale dance videos via global plan and local refinement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1236,"prompt_tokens":698,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":470}},"tokens_in":442,"tokens_out":538,"duration_ms":5748,"temperature":1.0,"reasoning_tokens":470,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:30:32.084414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute w_optical_flow for typical training frames: if it contains values other than 0 and 1, the loss in Eq. (3) does not weight the velocity error as claimed, and the reported optical-flow ablation could be re-run with a true mask to see if the gains vanish.","supporting_citations":[],"review_version":3}