{"id":"7ecc0404-2d1c-43f1-b94d-c2878d6a59d5","arxiv_id":"2501.08682","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion-based framework improves spatial and temporal consistency of clothes in virtual try-on videos, reporting the best FID/KID and several video metrics on four public datasets.","lead":"RealVVT is a video virtual try-on system that swaps an uploaded garment onto a moving person while keeping the fabric's color and texture stable across frames. It reports state-of-the-art scores on four datasets and could make fashion e-commerce try-on videos more realistic.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The main temporal-consistency mechanism is internally contradictory: Eq. (9) says 'randomly selected frame' while Sec. 3.4 describes a heuristic rule, and the SOTA claim rests on this unspecified choice. A controlled variant comparison would settle it.","rationale":"The reader's weakest assumption identifies the C&T attention mechanism as the key unvalidated component, and my read agrees. The internal contradiction between the heuristic description and Eq. (9) is concrete: one cannot tell whether the method uses a temporal-order-aware selection or a uniformly random one, and no ablation isolates this choice. Table 6 shows the mechanism produces a dramatic improvement, but that experiment is not designed to discriminate among selection strategies and also changes lambda_N relative to the implementation details. Since the central SOTA claim is specifically about temporal coherence, and the proposed mechanism is both the novelty and the largest reported contributor, the paper's central result cannot be fully verified as written. This does not imply the method is wrong; it implies the evidence is incomplete and conditional. The reader's CONDITIONAL verdict remains appropriate, so I recommend no change. I also note the image-dataset tables show mixed LPIPS results (e.g., Table 2 LPIPS 0.101 versus 0.0786 for DCI-VTON; Table 3 LPIPS 0.0608 versus 0.0432 for WildVidFit), so the 'outperforms in single-image VTO' claim is weaker than stated, but the C&T specification issue is more load-bearing for the video contribution.","tokens_in":876,"tokens_out":828,"duration_ms":65487,"concrete_test":"Run a controlled training comparison on the VVT set with identical data, backbone, and budget for three C&T variants: (a) fixed j=0, (b) uniform random j over all frames, (c) the adjacent-frame-difference heuristic described in Sec. 3.4, evaluating SSIM/LPIPS/VFID_I3D/VFIDResNeXt with standard errors over multiple seeds. Also verify whether the final model used lambda_N=0.01 or 0.1 by re-running the final configuration with both settings; if the random variant does not match or beat the heuristic, Eq. (9) as written is not the causal mechanism behind Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing element is the Clothing & Temporal Consistency attention: it is the paper's main novelty, and Table 6 attributes to it the largest consistency gains (SSIM 0.910 to 0.976; VFID_I3D 4.761 to 2.689). Yet Sec. 3.4 states two incompatible definitions. The prose motivates a 'heuristic rule that selects frames based on adjacent frame differences' with a temporal-order-dependent strategy, while Eq. (9) defines the feature as X_attn_pi = Concat(X_attn_i, X_attn_j, X_attn_c) with j a 'randomly selected frame' over the entire range including the current frame and future frames. No training-time versus inference-time protocol is given, and no comparison to fixed-first-frame selection is provided. If the implemented rule is genuinely random, the stated rationale for temporal coherence fails; if it is the heuristic, Eq. (9) and the algorithm description are wrong. The ablation in Table 6 is also confounded by an unreported lambda_N change: Sec. 4 implementation says lambda_N = 0.01, while the ablation says lambda_N = 0.1. The SOTA claim therefore rests on a component whose exact behavior cannot be reproduced from the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RealVVT, a video virtual try-on framework built on Stable Video Diffusion with a dual U-Net (ReferenceNet + Denoising UNet), and claims state-of-the-art performance on both single-image and video try-on tasks. It introduces three components: an Agnostic Mask-Guided Attention Loss for spatial consistency, a Clothing & Temporal Consistency (C&T) attention mechanism for temporal coherence, and a Pose-guided Long VVT strategy for long videos. Experiments are reported on VVT, VITON-HD, and DressCode (upper/lower/dresses), with quantitative metrics including SSIM, LPIPS, FID, KID, VFID_I3D, and VFID_ResNeXt, plus qualitative comparisons on ViViD.","tokens_in":14379,"tokens_out":3714,"duration_ms":32519,"significance":"If the method and its reported results are reproducible, RealVVT would be a meaningful advance in video virtual try-on, particularly because it avoids expensive optical-flow or dedicated temporal modules and instead augments attention with a single reference frame and garment features. The paper also provides useful qualitative evidence of improved temporal garment consistency on challenging cases. However, the central novelty (C&T attention) is underspecified and internally inconsistent as written, the ablation supporting it is confounded by an unreported hyperparameter change, and the long-video algorithm has a likely infinite loop in its pseudocode. These issues directly affect the validity of the headline SOTA claims, so the paper needs substantial revision before the results can be trusted.","major_comments":[{"comment":"The definition of the Clothing & Temporal Consistency attention is internally contradictory. Eq. (9) states that j is 'a randomly selected frame' from {0, ..., i-2, i, ..., N-1}, which includes the current frame and future frames. The prose immediately before Eq. (9), however, describes a 'heuristic rule that selects frames based on adjacent frame differences' where for small i the model gathers information from later frames and for large i it uses earlier frames. These are two different mechanisms. No training-time vs. inference-time protocol is given, and no ablation comparing the random choice, the heuristic, and fixed first-frame selection is provided. Since Table 6 attributes the largest consistency gains (SSIM 0.910→0.976, VFID_I3D 4.761→2.689) to this component, the exact implemented behavior must be specified and an ablation with controlled variants must be added.","section":"§3.4, Eq. (9)"},{"comment":"The ablation is confounded by an inconsistent value of λ_N. The implementation details in §4 state 'λ_N is set to 0.01', but the ablation in §4.2 says 'with a fixed λ_N of 0.1'. The four rows of Table 6 that vary λ_agn are therefore not directly comparable to the final C&T row if the two experiments use different λ_N values. The effect attributed to C&T may be partly due to the λ_N change. The authors should rerun the ablation with a single fixed λ_N (or explicitly disclose and justify the discrepancy) and report the full configuration for each row.","section":"§4.2, Table 6 vs §4 Implementation Details"},{"comment":"The Pose-guided Long VVT algorithm is not reproducible as written. First, the thresholds dpose (Algorithm 1) and ddense (§3.5 text) are inconsistent names and neither value is specified. Second, the pseudocode has a control-flow error: if the condition in the while loop is false, j is never incremented, which leads to an infinite loop. Third, the logic 'if ∥pi − pj∥2 < dpose or |i − j| < smax then Ω.insert(i)' appears to select a keyframe when the pose distance is small, which contradicts the stated goal of selecting frames with large pose/motion changes. The authors should provide a correct, unambiguous algorithm with concrete parameter values and a worked example.","section":"§3.5, Algorithm 1"},{"comment":"The claim of a 'clear performance advantage across all datasets' is not supported by the reported numbers. In Table 1, RealVVT's VFID_I3D (2.689) is worse than VITON-DiT (2.498) and GPD-VVTO (1.28). In the image tables, LPIPS is worse than at least one baseline on three of the four datasets: VITON-HD (Ours 0.101 vs. DCI-VTON 0.0786 and WildVidFit 0.0773), DressCode-Upper (Ours 0.0608 vs. WildVidFit 0.0432), and DressCode-Lower (Ours 0.0743 vs. MGD 0.0689). No error bars or significance tests are reported, so it is also unclear whether the numerical gaps are meaningful. The abstract and conclusion should be softened to state where the method is competitive and where it is not, rather than claiming universal state-of-the-art performance.","section":"Tables 1–5 and Abstract/Conclusion"}],"minor_comments":[{"comment":"The sentence 'Spatial Inconsistency: Garments tend to adhere to the target mask’s shape and color...' is duplicated verbatim twice in the list of challenges.","section":"§1, Introduction"},{"comment":"The text says 'prensent' instead of 'present' in §3 (second sentence of the section overview).","section":"§3.4, Eq. (9)"},{"comment":"The threshold name is inconsistent: the text uses 'ddense' while Algorithm 1 uses 'dpose'. Please unify the notation.","section":"§3.5 vs Algorithm 1"},{"comment":"In the visualization section, 'VIVID' is used as an alternative spelling of 'ViViD'; please be consistent.","section":"§4.1"},{"comment":"The text says 'λn ontrols' (typo) and later uses 'λN'; please fix the symbol and its capitalization consistently.","section":"§3.3, Eq. (3)"},{"comment":"The pseudocode formatting is ambiguous (e.g., 'do j+ = 1' inside the if block) and should be rewritten with clear loop semantics and an else branch.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper does not appear to be circular: the evaluation uses standard external datasets and metrics, and the proposed loss is an auxiliary training term. The main concerns are reproducibility and internal consistency. Given the centrality of the C&T mechanism to the claimed gains, the authors must provide a precise specification and a controlled ablation. If the implementation is genuinely as described in Eq. (9) (random selection), the stated temporal-coherence rationale fails; if it is the heuristic in the prose, the equation and the algorithm description are wrong. This is fixable within the scope of a major revision, so I do not recommend rejection at this point."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the stress-test note — I agree with it, and it lands on reading. The contradiction between Eq. (9) and Section 3.4 is real and sits at the load-bearing joint of the paper. Eq. (9) says the reference frame j is randomly selected from the full range including future frames; the prose describes a heuristic rule based on adjacent-frame differences with a temporal-order-dependent strategy. Those are different mechanisms, and the paper never says which one was trained and evaluated. Table 6 attributes the largest consistency gains to this module (SSIM 0.910 to 0.976; VFID-I3D 4.761 to 2.689), so the ambiguity matters. The lambda_N value also changes between Section 4 (0.01) and the ablation (0.1). That might be a typo, but it muddies the ablation. And dpose and smax in Algorithm 1 are never specified.\n\nThe paper does some things well. The max-attention modification to the agnostic loss is a small but sensible change over the uniform-mask variant. The long-video keyframe replacement with DensePose-distance-based selection is a reasonable zero-shot extension, borrowing from Rerender-A-Video. On VVT, the numbers show real improvement over public baselines: SSIM 0.976 vs 0.949, LPIPS 0.037 vs 0.068. The supplementary limitations paragraph about segmentation and DensePose jitter is honest and useful.\n\nThe soft spots beyond the internal contradiction: the claim of 'clear performance advantage across all datasets' does not hold. On image datasets, LPIPS is worse than WildVidFit on VITON-HD (0.101 vs 0.077), worse than WildVidFit on DressCode-Upper (0.0608 vs 0.0432), and worse than LaDI-VTON on DressCode-Lower (0.0743 vs 0.0596). The FID/KID wins are consistent, but that is not across-the-board superiority. There are no error bars or significance tests. Also, VFID-ResNeXt for VITON-DiT (0.187) and Ours (0.0913) are an order of magnitude lower than all other methods, which suggests a metric implementation difference that needs explanation.\n\nWho is this for? Someone building a video try-on system or a diffusion-based video editing pipeline will find the recipe useful. But the paper cannot be reproduced from the text alone, and the central mechanism's description is internally inconsistent. It deserves a serious referee, but the referee should ask for a corrected description of C&T attention, a comparison to fixed-first-frame selection, the missing hyperparameters, and code or detailed per-run numbers. I'd accept it for peer review with the expectation of major revision. If the authors clarify the mechanism and tone down the overclaim, this could be a solid system paper.","headline":"A useful video try-on recipe with real gains on public benchmarks, but the paper's central temporal-attention module is described two incompatible ways and the SOTA claim overshoots the evidence.","tokens_in":14890,"tokens_out":3997,"would_cite":true,"duration_ms":34850,"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":"RealVVT claims that video virtual try-on can be made photorealistic and temporally stable by feeding one randomly selected frame's attention features plus the reference garment into the denoising U-Net, and reports state-of-the-art scores…","keywords":["video virtual try-on","temporal consistency","diffusion models","attention mechanism","garment transfer","long video generation","pose guidance","Stable Video Diffusion"],"falsifier":"A controlled ablation that swaps the randomly selected frame in Eq. (9) for a fixed first frame, combined with a stress test on clips with fast limb motion or occlusion, would show whether the random-frame attention itself carries the reported VFID gains or whether the improvements come from the mask-guided loss and training setup.","tokens_in":13914,"feed_emoji":"👗","tokens_out":4857,"duration_ms":43332,"temperature":0.7,"pith_summary":"The paper sets out to show that video virtual try-on—putting a chosen garment onto a person in a moving clip—can be made temporally stable without optical flow or extra temporal modules. Its claim is that a single attention change, plus a mask-focused loss and a pose-based keyframe strategy, is enough to keep the same shirt, dress, or trousers on the body without flicker, shape drift, or texture loss. A sympathetic reader should care because garment flicker and accumulation of errors are exactly what block virtual try-on from being usable in shopping and fitting-room settings. The authors report that their system, RealVVT, outperforms previous single-image and video try-on models on VITON-HD, DressCode, VVT, and ViViD across most reported metrics.","feed_headline":"Random-frame attention stabilizes video virtual try-on","feed_subtitle":"One extra frame's features keep the same garment on a moving body; RealVVT reports top scores on four benchmarks.","key_machinery":"The load-bearing object is the Clothing & Temporal Consistency Attention (C&T), defined by Eq. (9): for each frame i, the key and value pair fed to self-attention is Concat(Xattn_i, Xattn_j, Xattn_c), where j is a randomly selected frame from anywhere else in the clip and Xattn_c is the ReferenceNet's features of the target garment. This single concat carries the entire temporal-consistency claim: it connects every frame's denoising to one other frame and to the fixed garment image, with no flow or temporal layers. Two supporting mechanisms are the Agnostic Mask-Guided Attention Loss (Eq. 4), which maximizes attention in the mask region and suppresses it outside, and the Pose-guided Long VVT strategy (Algorithm 1), which picks keyframes by L2 distance between DensePose frames and replaces keyframe latents with previously denoised outputs.","core_discovery":"RealVVT's central claim is that a video virtual try-on system can achieve photorealistic, temporally coherent results by treating the task as exemplar-based video inpainting on Stable Video Diffusion and modifying only the attention pathway. Instead of warping the garment with optical flow or adding a dedicated temporal module, it concatenates the current frame's key and value features with those of one randomly selected frame and with the reference garment's features, so every frame is denoised while looking at the same clothing evidence. An agnostic-mask-guided attention loss pushes the model to concentrate generation inside the masked clothing region, and a pose-guided keyframe replacement strategy extends generation to long videos. The paper reports that this combination gives state-of-the-art SSIM, LPIPS, FID, KID, and VFID scores on VITON-HD, DressCode, VVT, and ViViD.","pith_inferences":["If the random-frame attention genuinely beats a fixed first-frame control, which the paper does not ablate, it would imply that temporal consistency in generative video can be bought with a one-frame attention change rather than motion modules—a result worth testing on other object-transfer tasks like face swapping or style transfer.","The DensePose-distance keyframe selection is the least constrained piece; a natural extension is to replace the unspecified thresholds dpose and smax with a learned motion-saliency score, and to test the strategy on dance or sports footage where pose estimates jitter.","Because the paper's own limitation section reports that precomputed masks and DensePose from existing datasets are noisy and jittery, a clean test of RealVVT's robustness would be to feed it hand-corrected segmentation on a few clips and see how much of the reported VFID gain comes from input quality rather than the model itself."],"forward_implications":["On the paper's own numbers, RealVVT beats StableVITON, ViViD, VITON-DiT, Tunnel Try-on, and GPD-VVTO on the VVT dataset for SSIM, LPIPS, and VFID-ResNeXt, and beats all listed image methods on FID and KID for VITON-HD and DressCode.","Because the method is built on Stable Video Diffusion with only the first convolution layer changed to accept 9 channels, the same recipe can be transferred to other image-to-video diffusion backbones.","The joint image-video training means one checkpoint handles both single-image and video try-on, so a retailer could use the same model for product photos and motion previews.","The Pose-guided Long VVT strategy is zero-shot: it does not retrain for length, so arbitrarily long clips can be generated by concatenating overlapping segments."],"supporting_citations":[{"why":"Supplies the pretrained Stable Video Diffusion backbone that RealVVT fine-tunes for video virtual try-on.","marker":"[1]"},{"why":"Supplies the VITON-HD dataset and the clothing-agnostic person representation used for image benchmarks.","marker":"[38]"},{"why":"Supplies the DressCode dataset used for multi-category image try-on evaluation.","marker":"[7]"},{"why":"Supplies the ViViD high-resolution video dataset and is a primary video baseline.","marker":"[44]"},{"why":"Supplies the VVT video dataset used for the main quantitative video comparison.","marker":"[4]"},{"why":"Supplies the evaluation paradigm and is a video baseline compared against.","marker":"[16]"},{"why":"Supplies a strong recent video virtual try-on baseline that RealVVT claims to outperform.","marker":"[13]"},{"why":"Supplies StableVITON, the main single-image baseline on VITON-HD and DressCode.","marker":"[20]"},{"why":"Supplies the keyframe replacement strategy that the Pose-guided Long VVT adapts for long videos.","marker":"[8]"}],"fun_headline_variants":["One random frame's features keep clothes consistent in video try-on","Attention to one random frame keeps the same garment on moving body","Video try-on gets stability by conditioning attention on a random frame","RealVVT: random-frame attention for stable video try-on"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the attention features of a single randomly picked frame, added to the garment reference, are enough to keep the clothing from flickering during arbitrary motion; if that frame is unrepresentative, the temporal coherence gains disappear.","fun_headline_variants_meta":{"raw":{"variants":["One random frame's features keep clothes consistent in video try-on","Attention to one random frame keeps the same garment on moving body","Video try-on gets stability by conditioning attention on a random frame","RealVVT: random-frame attention for stable video try-on"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2735,"prompt_tokens":897,"completion_tokens":1838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1767}},"tokens_in":513,"tokens_out":1838,"duration_ms":13577,"temperature":1.0,"reasoning_tokens":1767,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:20:11.925235+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled ablation that swaps the randomly selected frame in Eq. (9) for a fixed first frame, combined with a stress test on clips with fast limb motion or occlusion, would show whether the random-frame attention itself carries the reported VFID gains or whether the improvements come from the mask-guided loss and training setup.","supporting_citations":[{"cited_title":"Viton-hd: High- resolution virtual try-on via misalignment-aware normaliza- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies the VITON-HD dataset and the clothing-agnostic person representation used for image benchmarks."},{"cited_title":"Dress code: High-resolution multi-category virtual try-on,","cited_arxiv_id":null,"evidence_quote":"Supplies the DressCode dataset used for multi-category image try-on evaluation."},{"cited_title":"Single stage virtual try-on via deformable attention flows,","cited_arxiv_id":null,"evidence_quote":"Supplies the VVT video dataset used for the main quantitative video comparison."},{"cited_title":"Gpd-vvto: Preserving garment details in video virtual try- on,","cited_arxiv_id":null,"evidence_quote":"Supplies a strong recent video virtual try-on baseline that RealVVT claims to outperform."},{"cited_title":"Stablevi- ton: Learning semantic correspondence with latent diffusion model for virtual try-on,","cited_arxiv_id":null,"evidence_quote":"Supplies StableVITON, the main single-image baseline on VITON-HD and DressCode."},{"cited_title":"Rerender a video: Zero-shot text-guided video-to-video translation,","cited_arxiv_id":null,"evidence_quote":"Supplies the keyframe replacement strategy that the Pose-guided Long VVT adapts for long videos."}],"review_version":1}