{"id":"9105bca0-8b90-4530-a728-a84c8da40ef9","arxiv_id":"2412.10178","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SwiftTry makes diffusion-based video virtual try-on faster and more consistent by shifting non-overlapping video chunks during sampling and caching features across denoising steps.","lead":"This paper presents SwiftTry, a diffusion-based system that replaces a person's clothing in a video with a new garment while keeping frames temporally coherent. It also introduces TikTokDress, a more challenging high-resolution video try-on dataset with diverse backgrounds and movements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported VFID numbers are internally inconsistent, and the paper's own comparison shows ShiftCaching degrading temporal consistency, so the central speed-consistency claim is not established.","rationale":"The reader's conditional verdict is appropriate, but the most load-bearing issue is not primarily the train/inference distribution shift in temporal attention. Instead, the most immediate problem is that the reported numbers contradict each other: Table 3 gives VFIDI3D = 3.589 for SwiftTry on VVT, while the ShiftCaching ablations in Tables 5–6 give VFIDI3D around 8.7–9.0 on the same dataset. This cannot be explained by the stated differences in shift, mask type, or partial computation. Furthermore, Table 4 directly shows ShiftCaching worsening VFIDI3D on TikTokDress, which undercuts the claim that temporal consistency is maintained. These inconsistencies make the central speed-consistency claim impossible to verify from the paper alone, and they are more decisive than the theoretical concern about training with fully computed features. The paper deserves a conditional decision because the method and dataset are plausible and useful, but the authors must release code, checkpoints, and exact evaluation protocols so that the main-table and ablation numbers can be reproduced. If the discrepancies persist under a unified protocol, the headline claim should be substantially weakened.","tokens_in":13986,"tokens_out":7540,"duration_ms":78614,"concrete_test":"Obtain the authors' code and trained checkpoint, then run a single evaluation script on the same VVT split under the exact settings of Table 3 and of each ablation row in Tables 5–6 (fixed/random shift, partial computation, full/half/quarter/causal attention). If the reproduced VFIDI3D values do not match 3.589 for the main setting and ~8.7–9.0 for the ablations, one table is not reproducible. Additionally, compute a bootstrapped 95% confidence interval for the TikTokDress VFIDI3D difference between SwiftTry with and without ShiftCaching; if the interval includes zero or excludes a meaningful improvement, the 'maintains temporal consistency' claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ShiftCaching reduces cost while maintaining temporal consistency is not supported by the paper's own quantitative data. On the VVT dataset, the main result (Table 3) reports SwiftTry VFIDI3D = 3.589, but every ShiftCaching ablation in Tables 5 and 6 on the same dataset reports VFIDI3D values around 8.7–9.0. No stated difference in shift type, attention mask, or partial-computation ratio explains a >5-point gap; this indicates that the main-table numbers and the ablation numbers were produced under different, unreported evaluation protocols, or that one set of numbers is erroneous. On TikTokDress, Table 4 directly compares SwiftTry w/o ShiftCaching (VFIDI3D = 3.865) with SwiftTry with ShiftCaching (VFIDI3D = 4.231), showing that the proposed technique degrades the temporal-consistency metric by roughly 9% while the text claims 'maintaining temporal consistency.' With no error bars, no code, and no released checkpoints, it is impossible to tell whether this degradation is statistically significant or an artifact of a mismatched protocol. Because the headline advantage over baselines in video consistency and speed rests entirely on these numbers, the central claim is currently unverified. The train/inference distribution shift identified by the reader is a plausible mechanism for such degradation, but the more immediate load-bearing problem is that the reported evidence is internally inconsistent and cannot be checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SwiftTry, a diffusion-based video virtual try-on method built from a two-stage pipeline: an image try-on model in the style of StableVITON (Garment UNet, pose control, Reference Attention) and a video stage that adds temporal attention layers initialized from AnimateDiff. For efficient long-video inference, it introduces ShiftCaching, which shifts chunk boundaries across denoising steps and partially computes frames using cached features, together with a Masked Temporal Attention (Half Attention) variant designed to handle mismatched cached features. The paper also contributes TikTokDress, a dataset of 817 high-resolution short-form videos with manually refined garment masks, diverse backgrounds, and complex motions. Experiments on VVT and TikTokDress claim that SwiftTry outperforms baselines in video consistency and runs at 2.27 FPS, about 60 percent faster than the re-evaluated ViViD baseline.","tokens_in":14340,"tokens_out":7862,"duration_ms":70272,"significance":"If the claims hold, ShiftCaching would be a practical contribution to long-video try-on inference, and TikTokDress would address a real gap in dataset diversity and mask quality. The paper's concrete assets are a reproducible speed comparison against one re-run baseline, a documented annotation pipeline for TikTokDress, and a set of ablations over attention-mask designs and chunk lengths. However, the quantitative support for the central claim is currently not internally consistent: the VVT main table reports VFIDI3D = 3.589 while all ShiftCaching ablations on the same dataset report values around 8.7-9.3, and the TikTokDress comparison shows the proposed technique degrading VFIDI3D from 3.865 to 4.231. These discrepancies, combined with a mixed protocol for baseline numbers, prevent the paper from establishing its headline speed-consistency claim.","major_comments":[{"comment":"Table 3 reports SwiftTry VFIDI3D = 3.589 on the VVT dataset, but every SwiftTry variant in Tables 5, 6, and 7 on the same dataset reports VFIDI3D between 8.679 and 9.312. No stated difference - fixed vs random shift, partial-computation ratio, attention-mask type, or chunk length - accounts for this more than five-point gap. The main table and the ablations appear to have been produced under different evaluation protocols, or one set of numbers is erroneous. Because the abstract's claim of superior video consistency rests on Table 3, this inconsistency is load-bearing and must be resolved by reporting the exact protocol for each table or recomputing all numbers under one protocol.","section":"Tables 3, 5, 6, 7"},{"comment":"Table 4 directly compares SwiftTry without ShiftCaching (VFIDI3D = 3.865) with SwiftTry with ShiftCaching (VFIDI3D = 4.231) on TikTokDress, a relative degradation of roughly 9 percent in the temporal-consistency metric. The paper nonetheless claims that ShiftCaching 'maintains temporal consistency' (Abstract) and that it 'ensures temporal smoothness' (Conclusion). No error bars or repeated runs are provided, so it is unclear whether the degradation is statistically significant, but as presented the data contradict the stated claim. The authors should report variance across seeds or videos and either demonstrate that the difference is within noise or qualify the claim as a speed-accuracy trade-off.","section":"Table 4"},{"comment":"The temporal attention layers are trained only on fully computed features at a consistent noise level within a chunk, but at inference ShiftCaching feeds half of each chunk with cached features from timesteps t+2 and t+1 (Fig. 3). The Masked Temporal Attention (Half Attention) is introduced to reconcile this mismatch, yet the model is never trained with mixed cached/current inputs or with the mask applied. The paper offers no analysis of the resulting distribution shift (for example, feature distances between cached and current features) and no evidence that the mask is sufficient on videos outside the test set. Given that Table 4 shows degraded VFIDI3D when ShiftCaching is enabled, this untested assumption is directly relevant to the central speed-consistency claim. The authors should address it, for instance by fine-tuning temporal layers with partial computation or by explicitly ablating the mismatch.","section":"Methods (ShiftCaching), Fig. 4, Table 6"},{"comment":"Table 3 mixes a re-run of ViViD (marked with a dagger) with numbers reported by other papers under unspecified protocols, so the comparison is not controlled. The text further states that SwiftTry 'excels in the VFID metric', but Tunnel Try-On has a better (lower) VFIDI3D (3.345 vs 3.589) and better LPIPS (0.054 vs 0.066) in the same table. The claim of outperforming baselines in video consistency is therefore contradicted by the table itself, and the protocol mismatch prevents adjudicating the comparison. The authors should state the evaluation protocol for each baseline, use identical masks/pose inputs where feasible, and revise the claim to match the reported numbers.","section":"Table 3 and 'Comparisons with Prior Approaches'"}],"minor_comments":[{"comment":"The shift value Delta is introduced only through the example in Fig. 3 with Delta = 4; the paper does not specify how Delta is chosen or whether it is tuned per dataset. Please define the shift schedule precisely.","section":"ShiftCaching Technique"},{"comment":"The 'Random Shift' variant in Table 5 is not defined in the main text; it is unclear how random shifts are sampled (per step, per chunk) and whether the random seed is fixed across runs.","section":"Table 5"},{"comment":"No error bars, confidence intervals, or repeated-run statistics are reported for any metric. Given the small differences in Tables 4 and 6, this makes it impossible to assess the significance of the reported improvements or degradations.","section":"All quantitative tables"},{"comment":"The main text says the mask 'set specific values in the attention matrix to 0', while the supplementary caption says white cells represent 0 and gray cells represent -infinity. Please align the mask semantics (0 vs -infinity before softmax) across the main text and supplement.","section":"Supplementary, Fig. 9 and main text"},{"comment":"The paper does not mention the release of code or checkpoints. Given the internal inconsistencies in the reported numbers, public release of the evaluation code and model weights would substantially strengthen the reproducibility of the claims.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea, ShiftCaching, is worth knowing about, but the paper's own numbers don't support its headline claim. If you read it, focus on Section 4 and Table 4.\n\nWhat's genuinely new: ShiftCaching—shift non-overlapping chunks by a fixed delta each denoising step, cache features from earlier timesteps, and use a masked temporal attention to keep the distribution shift from blowing up. That's a real engineering contribution, and the idea of training only the temporal layers and then doing this at inference is sensible. The TikTokDress dataset (817 videos, diverse backgrounds/movements, manually corrected masks) is a useful resource, even if it's built on top of an existing TikTok dataset.\n\nThe problems are in the evaluation. On VVT (Table 3), SwiftTry gets VFIDI3D 3.589, but every ablation in Tables 5-7 reports VFIDI3D around 8.7-9.0 for the same method and dataset. No protocol difference is stated. That's a 5-point gap with no explanation. Either the main table and the ablations used different metrics or subsets, or one set of numbers is wrong. Either way, you cannot trust the quantitative comparisons until this is cleared up.\n\nSecond, Table 4 on TikTokDress directly contradicts the abstract: SwiftTry with ShiftCaching gets VFIDI3D 4.231 while without it gets 3.865. So the proposed technique makes temporal consistency worse by roughly 9% on the one dataset where it's directly ablated. The paper says it 'maintains temporal consistency' but the only clean comparison says otherwise. The speed gain is real (2.27 vs 1.18 FPS), but the consistency claim is not supported.\n\nThird, the comparison to baselines is mixed-protocol. ViViD is re-run, but Tunnel Try-On, ClothFormer, etc. are reported numbers from other papers, likely under different settings. And Tunnel Try-On actually has better VFIDI3D (3.345) and LPIPS (0.054) than SwiftTry, so the claim of 'excelling in video consistency' is overstated.\n\nOn the positive side, the ablations of masking strategies are informative, and the authors do identify the distribution-shift problem with caching—they just don't solve it convincingly. No code or data is released, and there are no error bars, which makes it worse.\n\nBottom line: this is a plausible method and a dataset that could become a benchmark, but the current submission is not reliable as a quantitative claim. It deserves a serious referee, but only if the authors supply the missing evaluation details. I'd send it to review with a request for major revision, not desk-reject it.","headline":"A plausible inference-time caching trick and a useful new dataset, but the evaluation numbers are internally inconsistent and the central speed-consistency claim is not supported by the paper's own tables.","tokens_in":14818,"tokens_out":2436,"would_cite":false,"duration_ms":23693,"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":"The paper claims that reusing cached denoising features across shifted chunks makes video virtual try-on temporally consistent and about 60 percent faster than prior diffusion-based approaches.","keywords":["video virtual try-on","diffusion models","temporal consistency","feature caching","video inpainting","ShiftCaching","TikTokDress dataset","inference acceleration"],"falsifier":"Take a trained SwiftTry model and run it on a long clip with fast arm movements or a sudden lighting change exactly at a chunk boundary; if the cached-feature halves show visible texture smearing, flicker, or garment distortion that the fully computed frames do not, the cached-feature assumption is violated. A quantitative version would compare the Half Attention variant against full computation on such clips: if VFID_I3D on the cached halves is systematically worse than on fully computed frames, the temporal layers have not absorbed the distribution shift.","tokens_in":13797,"feed_emoji":"👗","tokens_out":8450,"duration_ms":76946,"temperature":0.7,"pith_summary":"SwiftTry claims that video virtual try-on—replacing the garment on a moving person while keeping the rest of the video intact—can be made both fast and temporally consistent by viewing it as conditional video inpainting and by avoiding the expensive overlapping-chunk stitching used by previous diffusion-based methods. Its inference scheme, ShiftCaching, processes non-overlapping chunks whose boundaries shift at each denoising step, and for half the frames it reuses cached features from earlier steps instead of recomputing the deep UNet blocks. A Masked Temporal Attention variant keeps those cached features from contaminating fresher ones. The paper also introduces TikTokDress, a high-resolution 817-video benchmark with complex backgrounds and movements. On VVT and TikTokDress, the method reports better video-consistency metrics than ViViD while running at 2.27 FPS, about 60 percent faster.","feed_headline":"Video try-on runs 60% faster by caching denoising features","feed_subtitle":"A diffusion try-on model reuses half-computed features across shifted chunks to keep long videos flicker-free at 2.27 FPS.","key_machinery":"ShiftCaching is the central mechanism. At each denoising step, the video is divided into non-overlapping chunks of N frames, and the chunk boundaries shift by a fixed offset of four frames between steps, so different frames meet at chunk edges over the sampling trajectory. To cut cost, only every other frame is fully propagated through the UNet; the other half is computed partially with features cached from the previous fully computed step (t+1 for one half, t+2 for the other). Masked Temporal Attention, specifically the Half Attention variant, then restricts the softmax so that lower-quality cached features can attend to high-quality fresh features while fresh features do not attend to stale ones, preventing errors from flowing backward through time while preserving temporal coherence.","core_discovery":"On its own terms, the discovery is that the temporal-coherence problem in video try-on splits into two parts that can be solved separately: giving the diffusion model a temporal attention pathway so frames actually exchange information, and then feeding that pathway with features that are cheap but not stale. The authors show that by shifting chunk boundaries by four frames at each denoising step and caching the features of every other frame from the previous step, the model escapes the overlap-recompute trade-off documented in their Table 1. The remaining distribution shift—half the chunk's features come from timestep t+2 and half from t+1—is handled by a Half Attention mask that lets stale features read, but not write into, fresher features. With this machinery, the paper reports consistent long-video try-on at 2.27 FPS on an A100, outperforming baselines on VFID-based video-consistency metrics on both VVT and TikTokDress.","pith_inferences":["The same cached-feature trick should transfer to other video diffusion tasks that already have temporal attention, such as human animation or object editing; the only task-specific piece is the Half Attention mask's assumption that fresher features are always the more reliable ones.","A natural testable extension is to train temporal attention directly on mixed cached and fresh inputs; if that closes the distribution gap, the cache ratio could be raised beyond 50 percent for even larger speedups.","The supplement's mask-error analysis implies that a large share of try-on artifacts in current datasets comes from agnostic-mask mistakes rather than the generative model, so improving automatic segmentation may yield bigger quality gains than further architecture work.","The paper's fixed shift of four frames and chunk length of sixteen are both tuned hyperparameters; ablating the shift size on high-motion videos would show whether the consistency guarantee degrades gracefully when motion is fast."],"forward_implications":["Long video try-on no longer needs to recompute overlapping frames, so the cost of temporal consistency becomes roughly independent of overlap size, which is what pushes inference to 2.27 FPS.","Because temporal attention is the only module trained in the second stage and only it is masked at inference, the approach transfers the image try-on model's garment fidelity directly to video.","On the VVT benchmark, the method reports VFID_I3D of 3.589 and VFID_RN of 0.534, improving over ViViD's 6.788 and 0.853 while keeping LPIPS and SSIM competitive.","With TikTokDress, evaluation moves to higher-resolution, in-the-wild clips, so the reported gains are tied to harder backgrounds and movements than previous public benchmarks.","Matching training and inference chunk length matters: the paper finds that 16-frame chunks give the best results, so the temporal layers are specialized to that chunk size."],"supporting_citations":[{"why":"ViViD is the primary video try-on baseline and the comparison point for both temporal consistency and the 1.41-to-2.27 FPS speedup; its released code is re-evaluated on VVT.","marker":"Fang et al. 2024"},{"why":"StableVITON supplies the two-stage image try-on architecture with Garment UNet and Main UNet inpainting that SwiftTry extends with temporal layers.","marker":"Kim et al. 2023"},{"why":"AnimateDiff provides the initialization and pseudo-3D temporal attention design used to add temporal coherence to the image model.","marker":"Guo et al. 2023"},{"why":"DeepCache contributes the observation that adjacent denoising steps share high-level features, motivating partial computation with cached UNet features.","marker":"Ma, Fang, and Wang 2024"},{"why":"AnimateAnyone supplies the Reference Attention mechanism through which garment features are injected into the Main UNet, and its combination with StableVITON is a compared baseline.","marker":"Hu et al. 2023"},{"why":"VVT is the standard evaluation benchmark and the source of the VFID metrics; FW-GAN is also an earlier video try-on baseline.","marker":"Dong et al. 2019b"},{"why":"VITON-HD is the dataset used to train the first-stage image try-on model and provides high-resolution garment and person pairs.","marker":"Choi et al. 2021"},{"why":"Stable Diffusion supplies the pretrained latent diffusion weights from which the Main and Garment UNets are initialized.","marker":"Rombach et al. 2022"},{"why":"SAM 2 is the base segmenter for garment and arm masks in TikTokDress, and its prompt sensitivity motivates the frame-selection algorithm and manual corrections.","marker":"Ravi et al. 2024"}],"fun_headline_variants":["Shift-caching makes video try-on faster and stable","Diffusion try-on reuses cached features for speed","Cached denoising features power flicker-free try-on video","Try-on video gets 60% faster with shift-caching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that temporal attention layers trained on fully computed features remain reliable when, at inference, half the frame features in a chunk are reused from earlier denoising steps that the network has never been trained on.","fun_headline_variants_meta":{"raw":{"variants":["Shift-caching makes video try-on faster and stable","Diffusion try-on reuses cached features for speed","Cached denoising features power flicker-free try-on video","Try-on video gets 60% faster with shift-caching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000408,"raw_usage":{"total_tokens":2127,"prompt_tokens":962,"completion_tokens":1165,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1097}},"tokens_in":578,"tokens_out":1165,"duration_ms":9901,"temperature":1.0,"reasoning_tokens":1097,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:15:01.406460+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained SwiftTry model and run it on a long clip with fast arm movements or a sudden lighting change exactly at a chunk boundary; if the cached-feature halves show visible texture smearing, flicker, or garment distortion that the fully computed frames do not, the cached-feature assumption is violated. A quantitative version would compare the Half Attention variant against full computation on such clips: if VFID_I3D on the cached halves is systematically worse than on fully computed frames, the temporal layers have not absorbed the distribution shift.","supporting_citations":[],"review_version":1}