{"id":"261414cf-7d82-4ecd-957b-4b282746ffbe","arxiv_id":"2412.09822","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A DiT-based video try-on framework that reuses the backbone as garment encoder and uses limb-aware dynamic attention to improve temporal consistency.","lead":"This paper proposes Dynamic Try-On, a video virtual try-on system that transfers clothing from a product image onto a moving person in a video. It reuses the diffusion-transformer backbone for garment features and adds limb-aware attention to keep body parts consistent during rapid motion.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LDAM's claimed superiority over 3D full attention rests on a confounded ablation: Table 4 varies DFFM and attention type simultaneously.","rationale":"The reader's weakest assumption focused on the unreleased private dataset and the limited quantitative evaluation on VVT, a reproducibility concern. While valid, that concern affects generalization claims rather than the internal validity of the core technical comparison. A more load-bearing issue is that the paper's headline claim about LDAM outperforming 3D full attention is evidenced by a confounded ablation: Table 4 changes two variables simultaneously and uses an asymmetric insertion depth for the 3D full-attention baseline. This is an internal inconsistency in the support for a stated contribution, not merely missing external validation. Because the paper could still be correct after a clean ablation, the appropriate verdict remains conditional rather than reject. The reader's verdict was already conditional, so this stress-test identifies a different, more direct reason for conditionality; hence agreement_with_reader is 'disagree' on which assumption is weakest.","tokens_in":12401,"tokens_out":6125,"duration_ms":60755,"concrete_test":"Run a controlled ablation on the VVT dataset with DFFM status held constant: (a) w/ DFFM + None, (b) w/ DFFM + 3D full attention inserted in the same first-seven blocks as the paper's baseline, (c) w/ DFFM + LDAM inserted at the same depth, and optionally (d) w/o DFFM + LDAM. Report SSIM, LPIPS, VFID, FVD, and training VRAM at 192x256 for each cell. If (c) does not beat (b) on the video metrics and VRAM, the claim that LDAM surpasses 3D full attention is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's contribution (3) states that LDAM surpasses 3D full attention layers in both VRAM consumption and performance, and Sec. 4.5 claims this is demonstrated by Table 4. However, Table 4 compares a row with 3D full attention but without DFFM against a row with both DFFM and LDAM. Because DFFM itself changes memory and metrics (compare rows 1 and 2: SSIM 0.918 vs 0.915, LPIPS 0.092 vs 0.104, FVD 63.53 vs 66.25), the observed improvement in the LDAM row (SSIM 0.924, LPIPS 0.098, VFID 2.246, FVD 57.49) cannot be attributed to LDAM rather than DFFM. Additionally, Sec. 4.5 states that 3D full attention was inserted only into the first seven DiT blocks to avoid OOM, while LDAM is described as being plugged into the backbone without a matching insertion-depth restriction. The comparison is therefore not controlled for either the garment-preservation paradigm or the attention insertion depth. As reported, the evidence does not isolate the effect of LDAM, so the specific claim of surpassing 3D full attention is unsupported by the paper's own ablation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dynamic Try-On, a DiT-based video virtual try-on framework that avoids a separate garment encoder by reusing the denoising backbone for garment feature extraction and storing the features in a Dynamic Feature Fusion Module (DFFM). It also introduces a Limb-aware Dynamic Attention Module (LDAM) that selects limb tokens from pose keypoints and applies masked self-attention to improve temporal consistency of body parts. The method is evaluated on the VVT benchmark and on a private e-commerce dataset, and is compared against GAN-based, image-based, and video diffusion baselines. The paper claims lower training memory than parallel garment encoders, improved limb consistency, and state-of-the-art video try-on quality.","tokens_in":12648,"tokens_out":7471,"duration_ms":70739,"significance":"If the claims hold, the framework addresses two practical bottlenecks in video try-on: the memory cost of garment encoders and temporal consistency under rapid motion. The external VVT evaluation is a reasonable benchmark, and Table 3 provides a useful memory-scaling comparison. However, the load-bearing comparison of LDAM against 3D full attention is confounded, the quantitative superiority claim is not fully supported by the reported numbers, and the robustness claim for complex postures relies on an unreleased private dataset. The paper does not provide code or data, so reproducibility is limited.","major_comments":[{"comment":"The claim that LDAM surpasses 3D full attention layers in both VRAM consumption and performance is confounded in Table 4. The 3D full attention row is evaluated without DFFM, while the LDAM row includes DFFM; the two no-attention rows show that DFFM itself changes memory (58.3G vs 48.5G at 192x256) and changes evaluation metrics (SSIM 0.918 vs 0.915, LPIPS 0.092 vs 0.104, FVD 63.53 vs 66.25). In addition, Sec. 4.5 states that 3D full attention was inserted only into the first seven DiT blocks to prevent OOM, while LDAM is plugged into the backbone without an equivalent insertion-depth restriction. The reported comparison therefore does not isolate the effect of LDAM, and the specific claim in Contribution (3) is not supported by the evidence presented.","section":"Sec. 4.5, Table 4"},{"comment":"The text states that Dynamic Try-On outperforms existing methods quantitatively and 'in all aspects,' but Table 2 does not support this: Dynamic Try-On has LPIPS 0.098, which is worse than ClothFormer (0.081) and Tunnel Try-on (0.054), and FVD is missing for ClothFormer and Tunnel Try-on, preventing a complete video-level comparison. Moreover, the table reports no error bars, significance tests, or multiple-seed statistics, and the SSIM margin over ClothFormer (0.924 vs 0.921) is small enough that the superiority claim requires a controlled evaluation protocol.","section":"Sec. 4.4, Table 2"},{"comment":"The claim of robustness to complex human postures is mainly supported by qualitative results on a private e-commerce dataset that is not released and for which no quantitative metrics are provided. The only quantitative evaluation is on VVT, which the paper itself describes as involving simpler motions. As a result, the paper's central robustness claim is not independently verifiable with the submitted evidence.","section":"Sec. 4.1, Sec. 4.3"},{"comment":"The evaluation protocol is under-specified: the paper does not state the number of evaluation samples, sampling steps, classifier-free guidance scale, random seeds, or whether all baselines use the same frame count and post-processing for VFID/FVD. These details are necessary to reproduce the quantitative comparisons and to interpret small metric differences.","section":"Sec. 4.2, Sec. 4.4"}],"minor_comments":[{"comment":"There is a typo: 'MagicAnimate [46] has has demonstrated' should be 'has demonstrated.'","section":"Sec. 2.2"},{"comment":"The citation 'MagicAnimate[2]' appears to be wrong: the MagicAnimate method is reference [46], while [2] is Magic Clothing. In Table 2, 'OOTDiffusion [3]' should be '[44]'.","section":"Sec. 4.4"},{"comment":"References [40] and [41] are duplicates of the same paper (Patch-Routed Spatially-Adaptive GAN) and should be merged.","section":"References"},{"comment":"The row/column checkmarks in Table 4 are hard to parse because the cells are not visually separated; please use explicit component labels for each row, such as 'w/o DFFM', 'w/ DFFM', 'None', '3D Full Attention', and 'LDAM'.","section":"Table 4"},{"comment":"The mapping from pose keypoints to the limb token mask S_l is not specified; please state how keypoint coordinates are quantized to patch tokens and how padding tokens are excluded from the masked self-attention.","section":"Sec. 3.4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable, and the memory comparison for DFFM is informative. The main risk is the confounded LDAM ablation and the overstatement of quantitative superiority. I would ask for a controlled ablation that varies DFFM and attention type independently at matched insertion depth before accepting. The missing protocol details and the reliance on an unreleased private dataset for the robustness claim should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the DFFM idea -- reusing the DiT backbone as the garment encoder and storing features in a bank -- is a genuine, practical contribution. The memory numbers in Table 3 are compelling, and the VVT benchmark results in Table 2 look good against a reasonable set of baselines. Second, the paper's claim that LDAM beats 3D full attention is not supported by the ablation as reported. The stress-test concern is correct: Table 4 compares a row with 3D full attention and no DFFM against a row with both DFFM and LDAM, and the full-attention row was only inserted into the first seven blocks to avoid OOM. That is doubly confounded, so the LDAM-specific advantage is not isolated. The paper does well on framing: the two problems (garment encoding cost, temporal consistency of limbs) are real, and the architecture is a plausible response. The external VVT evaluation is a plus, and the paper is not circular -- the central results are benchmarked independently. What is missing: error bars or significance tests, a controlled ablation for LDAM, and any public code, weights, or the custom dataset that backs the 'complex postures' claim. The qualitative results on the private e-commerce dataset are not reproducible. The contribution list overstates what Table 4 can show; that should be fixed. This is a solid engineering paper for the video try-on community, not a conceptual breakthrough. I would send it to serious peer review, but the revision must include a clean ablation that varies LDAM and DFFM one at a time, with matched insertion depth, and ideally some release of code or data. If the authors cannot do the controlled ablation, the LDAM claim should be toned down to what the data actually supports.","headline":"A genuinely useful memory-saving backbone design for video try-on, but the paper's headline LDAM claim is not backed by its own confounded ablation.","tokens_in":13201,"tokens_out":2326,"would_cite":true,"duration_ms":22741,"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":"A new video try-on framework claims to keep garment detail and limb consistency during rapid motion while using less training memory than prior approaches.","keywords":["video virtual try-on","diffusion transformer","temporal consistency","limb-aware attention","garment feature fusion","dynamic attention","pose-guided generation","video generation"],"falsifier":"Run Dynamic Try-On and a strong baseline such as ViViD on a public benchmark containing fast limb movements and occlusion, training both from the same OpenSora weights with identical data and compute; if ViViD matches or beats its VFID/FVD scores, the central superiority claim fails.","tokens_in":12192,"feed_emoji":"👗","tokens_out":7298,"duration_ms":64253,"temperature":0.7,"pith_summary":"Dynamic Try-On is a diffusion-transformer framework for video virtual try-on that aims to preserve garment details and keep limbs temporally consistent even during rapid, complex motion, while using less training memory than earlier methods. The paper argues that a dedicated parallel garment encoder is unnecessary: the same DiT backbone can first extract garment features and then attend to them through a Dynamic Feature Fusion Module. A second module, the Limb-aware Dynamic Attention Module, uses pose keypoints to index and group limb tokens, applies masked self-attention within each limb group, and restores the updated tokens, so a body part in one frame can attend directly to itself in neighboring frames. On the VVT benchmark the method reports the best SSIM, LPIPS, VFID, and FVD among compared baselines, and the ablation shows LDAM outperforming 3D full attention while consuming less VRAM. If these results hold, the framework offers a practical path to deploying video try-on on ordinary GPUs.","feed_headline":"Dynamic attention tames fast motion in video try-on with less memory","feed_subtitle":"Two modules reuse the DiT backbone and focus on limbs, beating prior methods on VVT at lower training cost.","key_machinery":"The dynamic attention mechanism, combining DFFM and LDAM. DFFM reuses the DiT backbone as the garment encoder: a first forward pass on the garment image saves each block's features into a feature bank, and during denoising the corresponding saved garment feature is duplicated across frames and fused with the denoising feature through a cross-attention residual. LDAM turns human pose keypoints into a limb token mask, then runs masked self-attention separately on the tokens of each limb, restoring updated tokens to their original indices through a zero-initialized linear layer; this lets limbs attend across time without paying the quadratic cost of 3D full attention.","core_discovery":"The paper's central claim is that the two challenges of video try-on—preserving garment detail and keeping limbs temporally consistent under fast motion—can be addressed inside a Diffusion Transformer without an extra garment encoder or expensive 3D full attention. The Dynamic Feature Fusion Module stores intermediate garment features during a first forward pass and injects them via residual cross-attention during denoising, effectively making the backbone its own garment encoder. The Limb-aware Dynamic Attention Module selects tokens belonging to each limb from the pose skeleton, pads and groups them, runs masked self-attention within each group, and writes the results back at their original token locations, which the paper illustrates as enabling frame-to-frame limb correspondence directly instead of through background tokens. The authors state that LDAM 'surpasses 3D full attention layers in both VRAM consumption and performance' and that Dynamic Try-On 'outperforms the existing methods in generating videos, both quantitatively and qualitatively.'","pith_inferences":["LDAM's approach of routing limb tokens through per-limb self-attention could transfer to other articulated animation tasks such as dance generation or avatar control, wherever pose keypoints are available.","Because the complex-posture dataset is private and unreleased, the robustness claim can only be independently tested after a public benchmark with fast or occluded limbs is established.","The DFFM idea of reusing the backbone as its own reference encoder may extend to other conditional generation tasks where the conditioning image shares the latent space, such as video inpainting or object-driven animation."],"forward_implications":["If the reported VVT numbers hold, Dynamic Try-On is the strongest video try-on method on SSIM, LPIPS, VFID, and FVD among the compared approaches.","Because DFFM removes the parallel garment encoder, training memory scales with the backbone's own growth rather than a duplicated encoder; the paper reports DFFM fitting a 44-block model where the prior paradigm runs out of memory.","Because LDAM is sparse and pose-driven, it offers a cheaper route to temporal coherence than inserting 3D full attention, and the ablation reports it beating full attention on all four metrics.","Initialized from OpenSora weights and trained in three stages, the framework suggests that similar staged fine-tuning on other DiT video backbones could reproduce the gains."],"supporting_citations":[{"why":"Provides the VVT dataset used for the main benchmark and the warping-based baseline that diffusion methods must beat.","marker":"[5]"},{"why":"Concurrent UNet-based video try-on baseline with a full garment encoder; the main method to be outperformed.","marker":"[8]"},{"why":"Uses half of the backbone as a garment encoder for fashion video try-on; the paradigm DFFM claims to replace.","marker":"[18]"},{"why":"Uses a full-backbone garment encoder and serves as a recent video try-on baseline.","marker":"[45]"},{"why":"OpenSora provides the pretrained DiT weights that initialization and the three-stage training build on.","marker":"[14]"},{"why":"Cited to justify the need for stronger temporal attention than separated spatial/temporal modules.","marker":"[48]"},{"why":"The DiT architecture itself, which the backbone and both modules modify.","marker":"[28]"},{"why":"VITON-HD pairs are merged into the first-stage image training data.","marker":"[4]"},{"why":"The FVD metric used to evaluate video quality and temporal consistency.","marker":"[31]"}],"fun_headline_variants":["Video try-on without extra encoder: dynamic attention saves memory","Limb-focused attention keeps video try-on stable in fast motion","DiT backbone doubles as garment encoder, cutting VRAM and boosting quality","Dynamic attention replaces 3D full attention for smoother video try-on"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The robustness claim for complex postures rests on a private, unreleased e-commerce dataset; the public quantitative evaluation uses only the VVT dataset, whose motions are relatively simple.","fun_headline_variants_meta":{"raw":{"variants":["Video try-on without extra encoder: dynamic attention saves memory","Limb-focused attention keeps video try-on stable in fast motion","DiT backbone doubles as garment encoder, cutting VRAM and boosting quality","Dynamic attention replaces 3D full attention for smoother video try-on"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":1987,"prompt_tokens":945,"completion_tokens":1042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":969}},"tokens_in":561,"tokens_out":1042,"duration_ms":10199,"temperature":1.0,"reasoning_tokens":969,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:40:30.610977+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Dynamic Try-On and a strong baseline such as ViViD on a public benchmark containing fast limb movements and occlusion, training both from the same OpenSora weights with identical data and compute; if ViViD matches or beats its VFID/FVD scores, the central superiority claim fails.","supporting_citations":[{"cited_title":"Vivid: Video virtual try-on using diffusion models, 2024","cited_arxiv_id":null,"evidence_quote":"Concurrent UNet-based video try-on baseline with a full garment encoder; the main method to be outperformed."},{"cited_title":"Fashion-vdm: Video diffusion model for virtual try-on","cited_arxiv_id":null,"evidence_quote":"Uses half of the backbone as a garment encoder for fashion video try-on; the paradigm DFFM claims to replace."},{"cited_title":"Tunnel try-on: Excavating spatial- temporal tunnels for high-quality virtual try-on in videos","cited_arxiv_id":null,"evidence_quote":"Uses a full-backbone garment encoder and serves as a recent video try-on baseline."},{"cited_title":"Open-sora: Democratizing efficient video production for all","cited_arxiv_id":null,"evidence_quote":"OpenSora provides the pretrained DiT weights that initialization and the three-stage training build on."},{"cited_title":"Viton-hd: High- resolution virtual try-on via misalignment-aware normalization","cited_arxiv_id":null,"evidence_quote":"VITON-HD pairs are merged into the first-stage image training data."}],"review_version":1}