{"id":"cf0ba146-6c01-4ef4-90c2-8bf8f842ea7e","arxiv_id":"2510.03550","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DragStream enables real-time drag, deform, and rotate edits on autoregressively generated videos without retraining, by correcting latent drift and selectively filtering context features.","lead":"This paper introduces REVEL, a task where users can drag, deform, or rotate objects in videos while the video is still being generated by an autoregressive diffusion model. The authors propose DragStream, a training-free method that keeps the video model stable during drags by correcting latent-code drift and filtering out distracting high-frequency context.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADSR's moment-matching (Prop. 2, Eq. 4) may erase the drag signal: the paper's own Fig. 2(a) shows drag shifts the mean/variance that ADSR rescales away, and no proof or long-horizon test rules this out.","rationale":"The reader's weakest_assumption identifies exactly this: ADSR assumes neighboring latents' mean/std are a safe reference and that rescaling does not erase the drag signal. That is the most load-bearing point because it targets the core mechanism that distinguishes DragStream from prior latent-optimization methods. If the concern lands, the method's headline claim (training-free, high-quality, seamless streaming drag) fails even under perfect experimental reporting. The alternative concerns -- unreadable quantitative figures, small self-generated benchmark, no released code, and metrics partially measuring the objective -- are real but secondary: they affect verifiability and generalizability, not the internal correctness of the proposed mechanism. The paper provides qualitative evidence and ablations showing ADSR improves the reported metrics, but those metrics (especially DAI) overlap with the optimization objective, so they cannot disambiguate 'ADSR preserves the drag' from 'ADSR improves FVD/FID by reducing drift while the drag is carried by other components.' The proposed test is concrete, uses the paper's own benchmark, and isolates the moment-matching step. The verdict remains CONDITIONAL, as the paper needs the test (and code/data release) to support the central claim; this stress-test does not move the reader's verdict, so UNCHANGED is appropriate.","tokens_in":20510,"tokens_out":4906,"duration_ms":78366,"concrete_test":"Using the released (or reproduced) Self-Forcing setup, take the 204 drag annotations. For each drag: (1) run latent optimization WITHOUT ADSR to obtain z*; decode, track the handle point, and compute ObjMC. (2) Apply only the final ADSR rescaling to z* (using the same neighboring stats) and re-decode; compute ObjMC again. If ObjMC degrades by more than ~20% when the rescaling is applied post hoc, ADSR is erasing drag signal. Additionally, measure the fraction of the drag-induced latent change captured by the first two moments: compute ||mu_z* - mu_ref|| / ||z* - z_init||; if this ratio is substantial (e.g., >0.3), the moment-matching step is removing a principal component of the drag. Repeat on a stream where context latents have already been edited (e.g., two successive drags) to test reference contamination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DragStream achieves high-quality REVEL training-free. The load-bearing mechanism is ADSR (Prop. 2, Eq. 4), which after each optimization iteration rescales the optimized latent z_{T'}^{k'} to the mean/std of neighboring context latents: hat z = (Iter_optim(z) - mu_z)/sigma_z * sigma_ref + mu_ref. The paper's own Fig. 2(a) documents that drag-induced perturbations shift exactly these global statistics (mean, variance, min, max). Thus ADSR removes precisely the component of the drag perturbation that is encoded in the first two moments, while asserting without proof that the intended drag survives in the residual spatial structure. No decomposition is provided: how much of the handle-point displacement is carried by the global moments vs. higher-order/spatial features? If a nontrivial fraction of the drag signal lies in the moments, ADSR partially cancels the user's intended motion -- the very failure mode it claims to fix.\n\nThe evaluation cannot rule this out. It uses only 204 short Self-Forcing-generated clips (Appendix C.1); short clips with a single drag may have modest moment shifts, and no long-horizon or repeated-drag test is reported. In a streaming session with multiple drags, the 'reference' neighboring latents themselves contain prior drags, so the reference distribution also drifts. ADSR then anchors to an already-edited distribution, leaving no fixed point to the original unedited latent manifold. The qualitative demos are coherent, but they do not isolate whether the success comes from ADSR or despite it. A simple post-hoc test (below) would settle whether the rectification step erases drag signal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces REVEL, a streaming drag-oriented interactive video manipulation task, and proposes DragStream, a training-free plug-in for autoregressive video diffusion models. DragStream combines an Adaptive Distribution Self-Rectification (ADSR) strategy, which normalizes the optimized latent to the mean/standard deviation of neighboring context latents, with a Spatial-Frequency Selective Optimization (SFSO) mechanism, which filters keys/values in the frequency domain and gates gradients spatially during iterative latent optimization. The authors claim that this enables both editing and animation operations (translation, deformation, 2D/3D rotation) at any frame during autoregressive generation, and that it integrates seamlessly with existing autoregressive VDMs such as Self-Forcing and CausVid. Experiments are reported on a newly annotated benchmark of 204 Self-Forcing-generated clips, with FVD, FID, ObjMC, and DAI metrics, plus qualitative visualizations and ablations.","tokens_in":20929,"tokens_out":4464,"duration_ms":40806,"significance":"If the claims hold, the paper would make a useful practical contribution: a training-free, model-agnostic mechanism for interactive drag control of autoregressive video diffusion models, with a unified treatment of editing and animation. The REVEL task definition is clear and the qualitative demonstration of drag-induced effects (including rotation and deformation) is compelling. The paper also merits credit for attempting to identify latent drift and context interference as distinct failure modes and for providing component-level ablations. However, the current evidence is not sufficient to establish the central claim: the main quantitative figures are unreadable in the provided PDF, the benchmark is small and self-generated, no code or data are released, and the headline metrics (ObjMC and DAI) are closely aligned with the very objective the method optimizes. The ADSR mechanism also raises a specific correctness risk that is not addressed by any long-horizon or repeated-drag test.","major_comments":[{"comment":"The quantitative results that carry the central claim are not legible in the provided manuscript. Figure 5 (and similarly Figures 6 and 7) renders as a sequence of broken substitution glyphs (e.g., \"/uni00000016/...\") rather than visible bars or numbers, so none of the FVD/FID/ObjMC/DAI values can be read. This is not a cosmetic issue: the main text's assertion of consistent quantitative superiority cannot be verified. The only legible numeric table is Appendix H (Table 2). The authors should provide readable figures, or equivalently tabulate all reported numbers, for every experiment that is used to support the performance claims.","section":"Section 5.1, Figures 5-7"},{"comment":"The DAI and ObjMC metrics are not independent of the DragStream optimization objective. DAI (Eq. 11) measures the latent-feature difference between the original handle region and the target region, which is exactly what the reconstruction loss L_Rec (Eq. 3) minimizes by transplanting source-handle features to the target position. ObjMC similarly measures adherence to the user-specified trajectory that defines the optimization target in Eq. (1). Thus the headline 'better ObjMC/DAI' partially reports the method's own objective. Since the adapted baselines also use latent optimization, the comparison is not wholly meaningless, but the paper should acknowledge this circularity and supplement the evaluation with independent metrics: e.g., user studies, point tracking on held-out points, appearance/identity preservation, or measurements of the actual physical displacement achieved.","section":"Appendix C.2, Eq. (11) vs. Eq. (3)"},{"comment":"ADSR is load-bearing, but its key assumption is unverified. The method rescales the optimized latent z_{T'}^{k'} to the mean and standard deviation of neighboring latents after every iteration. Figure 2(a) shows that drag operations shift precisely these first-order moments (mean, variance, min, max); Eq. (4) then removes that shift. The paper gives no decomposition of how much of the handle-point displacement is carried by the global moments versus the residual spatial structure, and no long-horizon or repeated-drag test that would rule out partial cancellation of the user's intended motion. The evaluation uses only 204 short Self-Forcing clips, which cannot establish that the mechanism is safe under streaming sessions with multiple or long drags. Please add a direct stress test that measures handle-point displacement with and without ADSR, and report whether the method remains stable w","section":"Proposition 2, Eq. (4)"},{"comment":"The paper labels the three components as 'Propositions,' but no proofs or derivations are supplied. Proposition 2 states a heuristic normalization; Proposition 3 states a frequency-domain filtering choice; Proposition 4 states a spatial gradient mask. Each is plausible, and the ablation plots (once readable) may support them empirically, but calling them propositions without formal statements or assumptions overstates their status. Please either provide rigorous statements (with conditions under which the normalization preserves the drag signal, or the filter choice is optimal) or rename them as design choices/heuristics.","section":"Propositions 2-4"},{"comment":"The benchmark is small and self-contained: 204 clips generated by Self-Forcing, annotated by the authors, with no external data, no code, and no data release. No confidence intervals, per-category breakdown, or statistical significance tests are reported. The claim that DragStream is 'seamlessly integrated into existing autoregressive VDMs' is supported only by qualitative results on CausVid in Appendix I. For a new task definition and a proposed method, the evaluation should include: (i) release of code and benchmark annotations; (ii) numerical tables with error bars; and (iii) at least one experiment on a longer stream or a repeated-drag scenario to substantiate the streaming claim.","section":"Appendix C.1"}],"minor_comments":[{"comment":"Typo: 'achievninbg' should be 'achieving'.","section":"Section 3, REMARK 2"},{"comment":"There are two remarks labeled 'REMARK 2': one after Proposition 1 and one after Eq. (3). Renumber to avoid confusion.","section":"Section 3 and Section 4.2.1"},{"comment":"'Notions' should be 'Notations'.","section":"Appendix B, Table 1"},{"comment":"The notation definitions contain repeated typos: 'indictor' should be 'indicator' (two occurrences).","section":"Section 4.1"},{"comment":"The cutoff frequency set {0.2, 0.4, 0.6, 1} is given without units (normalized frequency?). Clarify how the Butterworth filter is applied to 2D feature maps and what value of N and the omega_i correspond to.","section":"Proposition 3, Eq. (7)"},{"comment":"The quantitative figures are not only unreadable but also appear to omit the actual numbers in the text. Even after fixing the rendering, please include the numeric values in the caption or body so that the results are accessible to all readers.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is potentially useful and the qualitative results are appealing, but the current manuscript cannot be accepted in this form because the main quantitative evidence is unreadable, the metrics are partially circular, and the ADSR mechanism has an unaddressed failure risk. I would encourage a resubmission with readable quantitative tables, code/data release, independent or at least clearly-deconfounded metrics, and a direct stress test of ADSR under repeated/long drags. If those are provided, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DragStream is a real contribution to the streaming video manipulation space. The REVEL task—dragging any frame during autoregressive generation, supporting both editing and animation with translation, deformation, and rotation—is not covered by prior work, and the paper is honest about what it builds on. The method combines latent optimization with two sensible fixes: ADSR re-normalizes the optimized latent to neighboring frame statistics to combat drift, and SFSO selectively propagates context in both spatial and frequency domains. Ablations show each piece matters, and the qualitative demos, including a second backbone (CausVid), are coherent. That is a solid engineering story.\n\nThe problems are in the evidence. The main quantitative figure is unreadable in the provided PDF, so I can't verify the numbers. No code or data is released. The benchmark is 204 Self-Forcing-generated clips, which is small and self-generated; there's no long-horizon or repeated-drag stress test. That matters because ADSR's moment-matching is exactly the kind of heuristic that could quietly erase the drag signal—their own Fig. 2 shows drag shifts mean and variance, and ADSR rescales those away. It's plausible the residual spatial structure carries the drag, but they don't show that, and across multiple drags the reference statistics themselves drift. This is a load-bearing assumption that needs a dedicated analysis, not just a footnote.\n\nThe circularity concern is fair but less damning than it sounds. ObjMC and DAI measure trajectory/target-region agreement, which is what L_Rec optimizes. So the headline numbers partly reflect the loss. But the metrics still measure user-aligned outcomes; the bigger issue is that they don't compare against a trained trajectory-control baseline. The paper excludes Tora and DragAnything for good reasons (training-free), but that leaves the comparison only against two other latent-optimization methods without ADSR/SFSO, so it can't isolate whether ADSR helps or hurts beyond the ablation.\n\nPropositions 2–4 are heuristics, not theorems—that's fine for an applied paper, but they should be labeled as design choices with empirical support. The support here is partial.\n\nBottom line: this deserves a serious referee. The task formulation and method are worth engaging with, and the flaws are addressable: release code, make the tables readable, add longer-stream evaluation, and test whether moment rectification removes drag signal. I'd conditionally accept, not reject. For people working on interactive video generation and drag-based control, this is a useful read.","headline":"A genuinely new streaming-drag task with a plausible training-free method, but the quantitative evidence is too weak to fully trust the headline claims.","tokens_in":21398,"tokens_out":2327,"would_cite":true,"duration_ms":19723,"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":"DragStream shows that drag-based editing of autoregressively generated video can run on the fly with no training, by repairing latent drift and filtering context interference.","keywords":["streaming video manipulation","drag-based editing","autoregressive video diffusion","training-free control","latent optimization","distribution drift","frequency-domain selection","interactive generation"],"falsifier":"A sustained stress test: run a long stream in which a user applies many successive drags to different objects, and measure the KL divergence between the optimized latent distribution and the neighbor distribution after each drag; if the divergence grows without bound, or object attributes change and the drag halts, the central claim fails. A simpler check is whether a drag on an object far outside the training distribution still succeeds with ADSR applied.","tokens_in":20397,"feed_emoji":"🎬","tokens_out":3333,"duration_ms":28185,"temperature":0.7,"pith_summary":"The paper proposes a new task, REVEL: letting a user drag any frame of a video while it is being generated, with support for translation, deformation, and 2D/3D rotation, in both an edit-a-frame mode and an animate-forward mode. It argues that this can be done without finetuning the video diffusion model, treating each drag as an iterative optimization of the latent code at a chosen denoising step. Two obstacles stand in the way: drag-induced perturbations accumulate and push the latent off-distribution, halting the drag, and context frames interfere with the edited result, producing artifacts. The paper's method, DragStream, counters the first by rescaling the optimized latent to match the mean and standard deviation of neighboring latents, and the second by selectively propagating context in the frequency and spatial domains. A sympathetic reader cares because the result is a plug-in control layer for existing autoregressive video models that avoids the hundreds or thousands of GPU-hours required by finetuning-based alternatives.","feed_headline":"Drag-edit video frames mid-generation, no retraining","feed_subtitle":"A training-free plugin corrects latent drift and context artifacts, enabling real-time drag control of autoregressive video diffusion.","key_machinery":"ADSR (adaptive distribution self-rectification) is a moment-matching step that, after each latent optimization iteration, rescales the dragged latent so that its mean and standard deviation equal those of the preceding neighboring latents at the same denoising timestep, preventing cumulative drift. SFSO (spatial-frequency selective optimization) is a two-part mechanism: switchable frequency-domain selection, which applies a Butterworth filter with a randomly chosen cutoff frequency to the cached self-attention keys and values so that each iteration mixes low-frequency robustness with high-frequency detail, and criticality-driven spatial-domain selection, a Gaussian filtering map that suppres","core_discovery":"The central claim is that streaming drag control over autoregressive video diffusion models is achievable at inference time alone, with no training, as long as two failure modes are addressed. DragStream first identifies that each drag optimization step perturbs the latent distribution; over iterations this drift breaks the denoiser, so it rectifies the latent after every step by renormalizing to the mean and standard deviation of the preceding neighboring latents at the same timestep (ADSR). It also observes that context frames, while essential for visual consistency, carry high-frequency details that dominate the optimization and create artifacts, so it filters the cached keys and values i","pith_inferences":["If ADSR generalizes beyond short clips, local moment matching could become a general recipe for any optimization-based latent editing in autoregressive diffusion, not just drag operations.","The switchable-frequency idea may transfer to other context-conditioned generation tasks where high-frequency leakage from reference frames causes artifacts, such as long-context text-to-video or image-to-video.","A natural testable extension is to apply ADSR to other perturbation sources (text prompts, camera controls, pose edits) to see whether the same drift-repair mechanism holds.","The 'anytime, anything' claim is only established for the tested regime: 204 short Self-Forcing clips; longer streams or out-of-distribution content could break the neighboring-latent reference assumption."],"forward_implications":["Users could correct a frame mid-generation without regenerating the video, because the edit propagates through the context window to subsequent frames.","The method is model-agnostic: demonstrated on two different autoregressive video diffusion models, it could become a standard inference-time control layer.","Because it is training-free, adding drag control costs only inference-time optimization, making interactive video manipulation feasible on a single GPU.","The unified operation format (editing and animation, each with translation, deformation, and rotation) provides a common interface for future drag-style video manipulation systems."],"fun_headline_variants":["Drag video as it generates, no retraining","Training-free streaming drag for video diffusion","Drag anything, anytime: live video editing","Mid-generation drag edits for video, zero training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"ADSR assumes that the mean and standard deviation of a few neighboring latents are a safe reference distribution and that rescaling the optimized latent to those statistics does not erase the drag signal.","fun_headline_variants_meta":{"raw":{"variants":["Drag video as it generates, no retraining","Training-free streaming drag for video diffusion","Drag anything, anytime: live video editing","Mid-generation drag edits for video, zero training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000755,"raw_usage":{"total_tokens":3213,"prompt_tokens":785,"completion_tokens":2428,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":2372}},"tokens_in":529,"tokens_out":2428,"duration_ms":71949,"temperature":1.0,"reasoning_tokens":2372,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:37:18.185051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A sustained stress test: run a long stream in which a user applies many successive drags to different objects, and measure the KL divergence between the optimized latent distribution and the neighbor distribution after each drag; if the divergence grows without bound, or object attributes change and the drag halts, the central claim fails. A simpler check is whether a drag on an object far outside the training distribution still succeeds with ADSR applied.","supporting_citations":[],"review_version":1}