{"id":"d5ccc92f-fecf-44cf-af26-a7a556cd11ab","arxiv_id":"2607.28261","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TARS splits videos into clip pairs for self-supervised camera learning, adds text-driven viewpoint labels, and restricts scarce paired training to high-noise timesteps.","lead":"This paper trains video models to re-shoot an existing video with a new camera path and viewpoint without 3D reconstruction, using only ordinary videos plus a small number of paired examples. The trick is to teach camera motion mainly in the noisiest diffusion steps and to use text prompts for viewpoint changes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproven 'shortcut effect' underpins the claim that 60K high-noise-only paired samples suffice; no data-budget/threshold sweep is reported.","rationale":"The reader's weakest_assumption correctly identifies the shortcut effect as the load-bearing unproven assertion. My analysis agrees: the entire data-efficiency argument reduces to this mechanism. The paper provides qualitative figures and inference-time injection ablations that support the timestep-dependent importance of high-noise stages, but those do not directly validate that training with a small amount of paired data restricted to the high-noise regime will elicit synchronized motion. The training procedure itself continues to feed misaligned self-supervised data to the low-noise steps, so it is not obvious that a small nudge at t∈[0.95,1.0] will propagate through the remaining denoising steps. The proposed concrete test is a direct, quantitative check of the exact claim: vary the two key knobs (paired-data budget and high-noise threshold) and measure motion synchronization. If the test confirms saturation, the paper's core claim is credible; if not, the central practical contribution collapses. I also note the paper has independent support: the ablation tables show consistent improvements when adding Stage-2 and when using high-noise injection, and the camera-grid representation is reused from prior work, lending some credibility. However, the reported metrics are mostly LLM-based and undefined, and no code or data is released, which compounds the difficulty of external verification. These issues, together with the shortcut-effect concern, justify the reader's CONDITIONAL verdict. My recommendation is UNCHANGED: the verdict remains CONDITIONAL pending the proposed experiment and metric clarification.","tokens_in":12751,"tokens_out":4060,"duration_ms":60930,"concrete_test":"Run a 2D sweep: starting from the Stage-1 checkpoint, fine-tune with cross-pair budgets of 0K, 10K, 60K, and 120K, each under high-noise thresholds t∈[0.95,1.0], t∈[0.8,1.0], and t∈[0,1.0] (full fine-tuning). Evaluate V-MPGE, R-Prec, and T-Prec on the same held-out set. If 60K at [0.95,1.0] already achieves V-MPGE within ~10% of the 120K/full-range result and adding more paired data yields no significant gain, the shortcut effect and timestep routing are supported. If V-MPGE improves substantially with larger budget or with wider timestep coverage, the 'few cross-pair samples suffice' claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that routing only 60K cross-pair samples to the high-noise regime (t∈[0.95,1.0]) is sufficient to elicit temporally synchronized subject motion. The mechanism for this is the 'shortcut effect' asserted in Stage 2: 'when multiple valid solutions exist, the diffusion model naturally tends to converge to the solution whose motion is temporally synchronized with the input video, as this solution is easier to optimize.' This is presented as an observation without measurement or derivation. If this effect is weak or absent, Stage-2 fine-tuning on 60K pairs restricted to only 5% of the denoising trajectory would be too weak to overcome the misaligned motion supervision in the remaining 95% of timesteps, which continue to be trained with self-supervised clip-split data where V1 and V2 are temporally offset. The paper's ablations (Table 3) compare 'Stage1' vs 'Full' and 'High' vs 'Mid&Low' inference-time source injection, but these do not directly test the training-time data-routing hypothesis: no sweep over cross-pair budget (e.g., 0K, 10K, 60K, 120K) or over the high-noise threshold (e.g., 0.95, 0.8, 0.0) is reported. Without such evidence, the sufficiency of 60K paired clips in the high-noise regime is an unvalidated assumption, and the paper's headline result - minimal paired data without 3D reconstruction - rests on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TARS, a two-stage 3D-free video re-shooting framework. Stage 1 trains on large-scale self-supervised clip pairs created by splitting a video into two parts and using a camera grid and MLLM-generated viewpoint text as conditioning. Stage 2 fine-tunes the model on a small set (60K) of cross-pair videos, but only in the high-noise timestep regime t∈[0.95,1.0], based on the observation that low-frequency structure and camera motion are established early in the denoising process. The authors report state-of-the-art camera accuracy, viewpoint control, spatio-temporal consistency, and visual quality against CamClone, TrajCrafter, and Seedance 2.0.","tokens_in":13095,"tokens_out":4581,"duration_ms":57340,"significance":"If the central claim is substantiated, the paper offers a practical recipe for video re-shooting that avoids explicit 3D reconstruction and relies on a modest amount of paired data, while adding text-driven semantic viewpoint control. The underlying timestep-aware analysis is plausible and builds on established coarse-to-fine properties of diffusion models. However, the key mechanistic assertion — the 'shortcut effect' — is not measured or derived, and the quantitative evidence suffers from undefined evaluation metrics and missing sensitivity analyses. The work has clear potential, but its load-bearing empirical claims require strengthening before the contribution can be accepted.","major_comments":[{"comment":"The central claim that 60K cross-pair samples restricted to t∈[0.95,1.0] suffice to elicit temporally synchronized subject motion rests entirely on the asserted 'shortcut effect.' No derivation, measurement, or controlled experiment is provided to show that the model preferentially converges to the temporally synchronized solution. The ablations in Table 3 do not vary the cross-pair data budget (e.g., 0K, 10K, 60K, 120K) or the high-noise threshold (e.g., 0.95, 0.8, 0.0). The 'High' and 'Mid&Low' rows are inference-time source injections, not training-time data-routing experiments. Without such a sweep, the sufficiency of 60K paired clips is not empirically established, and the main contribution of the paper remains unvalidated.","section":"Stage 2: Eliciting Spatio-Temporal Consistency via Few Cross-Paired Data"},{"comment":"Several metrics in Tables 1 and 3 are not defined in the paper: CE, FDR, VDR, FSCS, GCR, LSR, and V-MPGE. The text states that 'all metrics are assessed by Gemini 3.1 Pro,' but does not specify the prompts, the scoring rubric, the number of evaluation samples, or how the LLM-based scores were aggregated. ArcFace is mentioned without indicating which identity embedding model is used and how similarity is computed. Without metric definitions and protocols, the quantitative comparisons are not reproducible and the reported improvements cannot be independently verified.","section":"Evaluation Metrics"},{"comment":"The statement that 'for the vast majority (potentially up to 95%) of the denoising steps, such strict pairing constraints are unnecessary' is a direct restatement of the hand-set threshold t∈[0.95,1.0], not an independently measured finding. The threshold is described as 'empirically defined' but no sensitivity analysis is given. If the threshold were changed, the claimed 95% figure would change accordingly. The Figure 3 visualization and the 'High' vs 'Mid&Low' ablation support a qualitative division of labor, but they do not justify the specific 0.95 threshold or the claim that 95% of training steps can safely ignore paired supervision. A threshold sweep is needed to make this load-bearing claim credible.","section":"Timestep-Aware Data Routing"},{"comment":"The camera-grid representation and the clip-splitting self-supervised strategy are adopted from prior work (Liu et al. 2026, OmniDirector), as cited in the Method section. The abstract and contributions list 'self-supervised training' as a new element, but the only genuinely new component appears to be the timestep-aware routing of cross-pair data. With the routing hypothesis not yet validated by controlled data-budget/threshold experiments, the incremental contribution over the prior work is not clearly demonstrated. The authors should clarify the novelty boundary and provide the missing ablation evidence.","section":"Stage 1: Self-Supervised Learning for Re-Shooting"}],"minor_comments":[{"comment":"There is a formatting typo in the abstract: 'Videore-shootingaimstoregeneratevideoswithcontrollable' should have spaces. The abstract also includes a project page URL, which is helpful.","section":"Abstract"},{"comment":"The acronyms R-Prec and T-Prec are introduced, but the caption of Table 1 does not explain that they are rotation and translation precision. The reference 'DPA-V3 (Lin et al. 2025)' appears to cite 'Depth Anything 3' rather than a camera pose estimator; please verify the correct reference.","section":"Evaluation Metrics / Table 1"},{"comment":"The row labels 'Stage1' and 'Stage2' are ambiguous: 'Stage1' means full self-supervised training without the second-stage fine-tuning, while 'Stage2' means training only on the cross-pair data. Consider renaming to 'Stage 1 only' and 'Stage 2 only' for clarity.","section":"Table 3"},{"comment":"The phrase 'potentially up to 95%' is imprecise and suggests an estimate rather than a definition. Since the threshold is a design choice, state explicitly that 95% corresponds to the chosen t∈[0.95,1.0] interval.","section":"Stage 2"},{"comment":"The evaluation relies heavily on LLM-based metrics (Gemini 3.1 Pro) without any validation or comparison to human judgment. At minimum, a small human study or a description of the prompt used would help calibrate the reliability of these scores.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is positioned as a novel 3D-free re-shooting method, but a large part of the pipeline (camera grid, self-supervised clip splitting, and the base video diffusion model) is drawn from prior work. The main claim to novelty is the timestep-aware data routing, which is currently supported by an unverified 'shortcut effect' rather than direct ablations. I would advise the editor that acceptance requires either a rigorous controlled study of the data budget and threshold, or a substantially reframed contribution that does not depend on that specific mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: TARS has a genuinely new training-time idea — route scarce paired data to only the high-noise denoising steps and rely on cheap self-supervised clip splits for the rest — and the qualitative results look strong. But the paper doesn't yet give me a reason to believe the specific mechanism that makes 60K pairs enough, and the evaluation is too opaque to verify the SOTA claims.\n\nWhat's new: the timestep-aware data routing itself, plus the MLLM-based semantic viewpoint labeling. The two-stage pipeline is sensible, and the ablation shows that removing either stage degrades performance. The authors are honest that the self-supervised data is temporally misaligned and that cross-pair fine-tuning fixes motion consistency. The visuals in Figure 5 show real capability, especially large camera motion and perspective switching.\n\nWhere it's soft: the \"shortcut effect\" — the assertion that the diffusion model gravitates toward the motion-synchronized solution because it's \"easier to optimize\" — is doing all the load-bearing work and is not measured or derived anywhere. If that's true, it's a great finding; if not, the whole 60K-high-only design collapses. The authors need at least a sweep over cross-pair budget (0K/30K/60K/120K) and over the high-noise threshold (0.95, 0.8, 0.0) to demonstrate the effect exists. Without that, the claim is just a restatement of the chosen threshold.\n\nSecond, the evaluation: CE, FDR, VDR, FSCS, GCR, LSR are never defined, and they're \"assessed by Gemini 3.1 Pro\" — that's not reproducible. R-Prec/T-Prec from DPA-V3 need a reference or formula. No error bars, no code/data release, and the base model is in-house. That undermines the comparisons with baselines, though I don't suspect cheating.\n\nThird, the novelty overlap with OmniDirector is not handled. The camera grid and clip-split self-supervision come from that paper; TARS's contribution is the timestep routing. That's a real contribution, but it should be positioned explicitly as building on their own prior work, not presented as if the self-supervision is new here.\n\nThe \"without paired data\" phrasing in the abstract is also misleading — Stage 2 uses 60K paired clips. Minor, but it needs fixing.\n\nBottom line: this is a solid idea from a capable group, and the paper deserves review. But it's not ready as-is. I'd want the mechanism tested, metrics defined, and the OmniDirector relationship clarified.","headline":"Timestep-aware data routing is a genuinely useful idea, but the paper's central mechanism and its evaluation need more rigor before I'd trust the SOTA claim.","tokens_in":13656,"tokens_out":2641,"would_cite":false,"duration_ms":36268,"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":"Video re-shooting can be taught with just 60K paired clips and no 3D reconstruction, by routing paired data only to the high-noise denoising steps.","keywords":["video re-shooting","camera control","viewpoint control","diffusion timesteps","data scaling","self-supervised learning","camera grid","text-driven control"],"falsifier":"Train the Stage-2 model with the same 60K cross-pair samples but route them only to mid/low timesteps (t<0.95), keeping self-supervised data for the high-noise regime. If temporal synchronization (V-MPGE) remains as good as the proposed routing, the central claim that high-noise steps need the paired data collapses. Alternatively, collect paired videos where the source and target are deliberately out of sync (e.g., shifted by a few frames): under the shortcut-effect hypothesis, the model should still converge to the synchronized interpretation when text-conditioning is informative; if it inste","tokens_in":12583,"feed_emoji":"🎥","tokens_out":4857,"duration_ms":62650,"temperature":0.7,"pith_summary":"This paper argues that the scarcity of paired videos—two recordings of the same scene from different camera paths—is not the bottleneck it seems, because a diffusion model does not need paired supervision at every denoising step. A timestep-wise analysis shows that high-noise denoising steps lay down low-frequency structure: camera motion and subject dynamics. The authors exploit this by training on large amounts of unlabeled video (temporally split into clip pairs) for most steps, and reserving a small set of 60K true cross-pair videos for the high-noise regime only. If correct, this makes 3D-free, text-controllable re-shooting practical with data that is actually collectible, and enables reverse-angle shots and first-/third-person switches that 3D-based methods struggle with.","feed_headline":"60K paired videos suffice for 3D-free video re-shooting","feed_subtitle":"By sending paired supervision only to high-noise denoising steps, TARS keeps motion sync and enables reverse-angle shots.","key_machinery":"The load-bearing mechanism is timestep-aware data routing, realized through a two-stage training scheme. Stage 1 trains on self-supervised clip pairs—random splits of a single video—with a camera grid (a video-format rendering of camera trajectory as moving grid lines in an empty 3D room) as conditioning, to learn camera dynamics and viewpoint semantics. Stage 2 fine-tunes only on the high-noise timesteps with 60K cross-pair videos, leveraging the claimed 'shortcut effect': the model prefers the temporally synchronized motion solution because it is easier to optimize, a tendency reinforced by text conditioning and classifier-free guidance.","core_discovery":"The paper's central claim is that the division of labor across diffusion timesteps—high-noise steps for coarse spatiotemporal structure, low-noise steps for appearance—allows a data-routing strategy for video re-shooting. By sending only a small amount of strictly paired cross-video data to the high-noise interval t∈[0.95,1.0] and filling the rest of training with self-supervised clip-split data, the model learns accurate camera control, preserves appearance, and—via a hypothesized 'shortcut effect'—acquires temporally synchronized subject motion from the source video without large paired datasets or 3D reconstruction. The paper further introduces text-driven semantic viewpoint control, deco","pith_inferences":["The 'shortcut effect' is the linchpin but is asserted, not measured. A natural extension is to test it directly: if the synchronized solution is genuinely easier to optimize, per-timestep loss curves for synchronized vs. shuffled-motion pairs should diverge in the high-noise regime early in training.","The 95% figure (paired supervision unnecessary for up to 95% of steps) implies a scaling law: for a fixed paired budget, the optimal time interval to allocate them shrinks as the self-supervised corpus grows; one could look for a phase transition in motion synchronization as the high-noise window narrows below [0.95,1.0].","If the routing principle transfers, tasks like human pose transfer, relighting, or object re-posing might similarly need only targeted paired supervision at the timesteps where the corresponding global attribute is decided, rather than full-pipeline pairing.","The text-conditioning reinforcement hypothesis suggests a testable dependency: removing CFG or text prompts should degrade motion synchronization even with paired high-noise training; measuring that drop would separate the two mechanisms."],"forward_implications":["Video re-shooting no longer requires either explicit 3D reconstruction or massive paired datasets; the paired-data bottleneck is relaxed by roughly an order of magnitude (from millions to tens of thousands).","Models trained this way can synthesize previously unseen regions under large camera motions and support reverse-angle re-shooting and perspective switching, because high-noise self-supervised data provides broad priors for novel views.","Text-driven semantic control over shot scale, viewing angle, and perspective becomes a viable alternative to supplying camera parameters, since the self-supervised stage learns the association between language and viewpoint.","The timestep-aware routing principle suggests that any video-to-video task requiring global motion alignment may be trainable with far less paired data than currently assumed, provided the paired supervision is concentrated where the corresponding structure is decided."],"fun_headline_variants":["Timestep-aware data scaling yields 3D-free video re-shooting","Send paired data only to high-noise steps, get 3D-free re-shooting","A small slice of paired data unlocks 3D-free camera control","TARS: text-driven viewpoint control without 3D or paired footage","High-noise steps carry motion; TARS uses that for 3D-free re-shooting"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper asserts, without derivation or measurement, that when multiple valid motions are possible, the diffusion model naturally converges to the motion synchronized with the source video because it is easier to optimize; if this 'shortcut effect' does not hold, Stage-2 fine-tuning would not align subject motion and the entire data-scaling argument collapses.","fun_headline_variants_meta":{"raw":{"variants":["Timestep-aware data scaling yields 3D-free video re-shooting","Send paired data only to high-noise steps, get 3D-free re-shooting","A small slice of paired data unlocks 3D-free camera control","TARS: text-driven viewpoint control without 3D or paired footage","High-noise steps carry motion; TARS uses that for 3D-free re-shooting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2725,"prompt_tokens":752,"completion_tokens":1973,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1870}},"tokens_in":496,"tokens_out":1973,"duration_ms":21861,"temperature":1.0,"reasoning_tokens":1870,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:15:44.861828+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the Stage-2 model with the same 60K cross-pair samples but route them only to mid/low timesteps (t<0.95), keeping self-supervised data for the high-noise regime. If temporal synchronization (V-MPGE) remains as good as the proposed routing, the central claim that high-noise steps need the paired data collapses. Alternatively, collect paired videos where the source and target are deliberately out of sync (e.g., shifted by a few frames): under the shortcut-effect hypothesis, the model should still converge to the synchronized interpretation when text-conditioning is informative; if it inste","supporting_citations":[],"review_version":2}