{"id":"148f2ed0-c127-469c-89fe-64b11d8b6f8a","arxiv_id":"2607.19923","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"WearWow generates native 2K multi-garment virtual try-on images without masks, using token packing plus dual preference rewards to preserve fabric texture.","lead":"WearWow is a diffusion system that generates virtual try-on photos with several garments at once in native 2K resolution, without hand-made masks. It cuts memory use by packing garment tokens into one canvas and restores fabric detail with a two-part preference reward.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-garment SOTA rests on unvalidated synthetic agnostic dataset; test on real-world try-off distribution.","rationale":"The reader's weakest assumption correctly identifies the single most load-bearing concern: the entire multi-garment evaluation and the mask-free inference rely on a self-constructed dataset with synthesized agnostic models, which are not validated and whose details are in a missing supplementary. If the try-off pipeline is flawed or unrepresentative, the model's reported superiority over commercial baselines is an artifact of training and testing on the same synthetic distribution. I considered the internal inconsistency in the SGR (Eq. 3 uses CLIP cosine similarity, while Section 4.2 says HPSv2 is used) but view it as a reproducibility issue rather than a direct threat to the central claim's core; the dataset validity is more fundamental. The reader's conditional verdict is appropriate: external single-garment results provide some support for the approach, but the multi-garment SOTA remains unverified without dataset release and independent validation. No change to the verdict is needed.","tokens_in":13582,"tokens_out":6343,"duration_ms":62733,"concrete_test":"Release or independently replicate the try-off pipeline, then apply it to a set of real 2K person photos to obtain agnostic models. Compute the FID between this real-world agnostic set and the WearWow-2K training agnostic images; also run WearWow on these real agnostic inputs and compare output FID/LPIPS and human preference scores against the reported WearWow-2K test results. If the agnostic FID is high or output quality drops significantly (e.g., >10% relative), the dataset is not representative and the multi-garment SOTA claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of state-of-the-art native 2K multi-garment synthesis is evaluated only on WearWow-2K, a dataset whose clothing-agnostic target images are produced by an automated try-off pipeline. The paper asserts these synthesized agnostic models are 'flawless' and 'physically correct,' but the dataset details, including the try-off pipeline's architecture and validation, are deferred to a supplementary that is absent from the preprint. Because the same pipeline is used to create both training conditions and test conditions, any systematic artifacts (e.g., smoothing, ghosting, or unrealistic body shapes) would be learned by the model and would inflate the reported FID, KID, LPIPS, and human evaluation scores. If the try-off distribution differs from real user photos—which lack such synthesized artifacts—the mask-free inference would not generalize, and the claimed SOTA would not hold. This concern is load-bearing because without a valid dataset, the multi-garment experiment provides no evidence for the central claim; the single-garment external benchmarks (VITON-HD, DressCode) do not exercise the multi-garment regime or the try-off pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WearWow, an end-to-end, mask-free virtual try-on framework operating natively at 2K resolution for multiple garments. To handle the memory cost of many high-resolution reference conditions, it introduces Adaptive 2D Token Packing (ATP), which packs garment images onto a shared 2D canvas and prunes background tokens. To counteract the over-smoothing of high-frequency fabric textures, it proposes a Multi-dimensional Try-on Reward (MTR) that combines a Cloth Distribution Reward (CDR), based on DPO-style preference pairs, with a Semantic Guidance Reward (SGR) that uses contrastive text anchors. The authors also introduce WearWow-2K, a 100k/2k training/test dataset of native 2K multi-garment triplets with synthesized clothing-agnostic person images. The paper reports state-of-the-art results on WearWow-2K for multi-garment synthesis against commercial APIs and competitive single-garment results on VITON-HD and DressCode.","tokens_in":13900,"tokens_out":3916,"duration_ms":41303,"significance":"If the central claim is validated, WearWow would be a meaningful advance: it directly addresses two real bottlenecks in high-resolution multi-garment try-on — the quadratic attention memory blow-up and the spectral bias of diffusion objectives — and its ATP/MTR formulations are well-motivated and clearly described. The single-garment external benchmarks (VITON-HD, DressCode) give partial independent grounding and show competitive performance. The paper is also careful in ablating its two main components. However, the multi-garment SOTA claim rests entirely on WearWow-2K, whose synthesized agnostic targets are produced by an unpublished try-off pipeline and whose validation is deferred to a missing supplementary. The quantitative support also lacks error bars, significance tests, and a specified human-evaluation protocol. These gaps are load-bearing for the paper's headline claim, so the significance is currently conditional.","major_comments":[{"comment":"The multi-garment benchmark is evaluated only on WearWow-2K, whose target agnostic images are generated by an automated try-off pipeline. The paper states in §3.1 that dataset details are deferred to the supplementary and in §4.1 that detailed distributions and visual examples are deferred to the supplementary, which is not included. Because the same synthetic distribution is used for training and testing, systematic artifacts (smoothing, ghosting, body-shape bias) would be learned by the model and inflate the reported FID/KID/LPIPS/HE. The abstract's claim of 'physically correct' and 'flawless' agnostic models is unsupported. The authors should provide the try-off architecture, manual/automated validation of the agnostic images, and ideally an independent test set using real user try-off photos to demonstrate generalization. This is load-bearing for the central SOTA claim.","section":"§3.1, §4.1, Table 2"},{"comment":"No error bars, confidence intervals, or significance tests are reported for FID, KID, LPIPS, SSIM, or HE. KID is known to be high-variance; differences like 7.54 vs 8.18 (Table 2) may not be statistically distinguishable without repeated trials. The human evaluation protocol is underspecified: how many ratings per comparison, how participants were instructed, whether outputs were randomized and blinded, and what the variance across participants was. The statement 'a clear user preference' (HE 0.2543 vs 0.2229) needs statistical support. Without this, the 'new state-of-the-art' claim is not quantitatively established.","section":"Table 1, Table 2, §4.3 (Human Evaluation)"},{"comment":"The CDR preference pairs are constructed by diffusing a high-quality (chosen) and a low-quality (rejected) sample to the same noise level, but the paper does not specify how the low-quality samples are obtained or labeled. The text only says 'we curate a comparative subset based on the physical material fidelity of generated results.' If the rejected samples are, for example, SFT outputs or over-smoothed variants, this must be stated precisely. The implicit reward R_theta in Eq. (2) depends on this choice: if the reject distribution is too close to or too far from the chosen distribution, the preference signal degenerates. This is a correctness-risk issue for the MTR method and should be clarified.","section":"§3.3, Eq. (1)–(2)"},{"comment":"The multi-garment baselines are a mix of open local models and commercial APIs, but the comparison protocol is not specified. It is unclear what prompts, garment inputs, or reference masks were used for each baseline, and whether each baseline had equal access to the same garment images and textual descriptions. Commercial APIs may have different internal prompt parsing or resolution handling; without a controlled protocol, the comparison may be unfair. Also, reference [3] is cited as GPT-image 1.5 but points to the DALL-E 3 technical report, and [30]/[31] are duplicates; these should be corrected.","section":"§4.3, Table 2 caption"}],"minor_comments":[{"comment":"The figure shows 'Sparse Density: 40% ～70%' and '70% ～ 90%', while the text in §3.2 says the baseline token utilization is approximately 40% and pruning elevates it from 70% to over 90%. Please align the figure labels with the text and clarify the two stages.","section":"Figure 2"},{"comment":"The role of ω is described as a 'relative CFG scaling coefficient,' but the equation itself is a simple margin loss. Please define how ω interacts with CFG or justify the name.","section":"§3.3, Eq. (4)"},{"comment":"The backbone is described as 'pre-trained Qwen-Image-Edit foundation model' with a Qwen2.5VL encoder, but no citation or details of the edit model are given. It would help to state the exact checkpoint and any conditioning interface used.","section":"§4.2"},{"comment":"There are several typos and formatting issues: 'T able 1', 'sparisity', 'W earW ow' in table captions, and inconsistent capitalization in figure labels. A careful proofread is needed.","section":"Throughout"},{"comment":"The release status of WearWow-2K is not stated (e.g., whether it will be public and under what license). This matters for reproducibility and for the community's ability to independently verify the multi-garment benchmark.","section":"§4.1"},{"comment":"The discussion of SSIM favoring over-smoothed outputs is a reasonable caveat, but it is not quantified. Consider reporting a no-reference texture metric or frequency-domain analysis to substantiate the claim that the LPIPS/FID improvements reflect restored high-frequency details rather than other perceptual differences.","section":"§4.4, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid architectural contribution and the single-garment results are encouraging, but the headline multi-garment SOTA is not yet supported because the only multi-garment benchmark is a self-created dataset whose synthetic agnostic images are not validated in the preprint. The authors should be asked to supply the missing supplementary, add an independent real-user try-off evaluation, and provide statistical rigor (error bars/significance tests). These are addressable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the core engineering idea is sensible and the single-garment numbers are respectable, but the flagship multi-garment claim is not yet supported. The entire evaluation of the multi-garment regime rests on WearWow-2K, and the crucial part of that dataset — the automated try-off pipeline that produces the clothing-agnostic conditions — is described only by assertion ('physically correct', 'flawless') with validation deferred to a supplementary that isn't here. Because the same synthetic condition type is used in both training and test, the reported FID/KID/LPIPS improvements could simply reflect the model learning the try-off pipeline's artifacts. I don't think this is deliberate gaming — the single-garment results on VITON-HD and DressCode are genuinely competitive, which tells me the underlying system has real capacity. But those benchmarks don't exercise multi-garment composition or the try-off pipeline, so they can't carry the main claim.\n\nWhat's actually new: ATP is a pragmatic token-packing scheme that keeps 2D spatial structure while cutting sequence length — a useful engineering contribution for multi-condition 2K diffusion, and the memory math is straightforward. MTR is a reasonable combination of a velocity-based distribution reward and a semantic margin objective; it borrows DPO and semantic relative preference ideas without adding a new learning principle, but the design is coherent. The reward-hacking concern is handled plausibly with the timestep-dependent weighting.\n\nWhere it's soft, in proportion: (1) the dataset/validation gap above is the big one; (2) no error bars or significance tests, and commercial API baselines are stochastic — one run each is not enough; (3) the human evaluation reports a single averaged HE score per method with no protocol detail; (4) minor internal mismatch — the method text says a frozen CLIP model computes the semantic similarities, but the implementation section says HPSv2, which is a different kind of scorer; (5) there's no sensitivity analysis for β, λ, τ, ω, α_t, or the canvas capacity L, and the free hyperparameters are all hand-set.\n\nWho this is for: people building industrial VTON pipelines will find ATP useful if it ships as code; researchers reading it for the preference-optimization angle will find the MTR recipe worth trying but not revelatory.\n\nMy recommendation: send it to peer review — the system is substantial enough to merit referee time — but don't let the multi-garment SOTA through as-is. If the authors release the dataset, validate the try-off conditions against real images, and add independent multi-garment evaluation, the paper would be a solid contribution. Until then, treat the headline claim as unverified.","headline":"Plausible system, competitive single-garment numbers, but the multi-garment SOTA is built on a self-created dataset whose validation is deferred to a missing supplementary — that claim is unverified as it stands.","tokens_in":14380,"tokens_out":3338,"would_cite":false,"duration_ms":35376,"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":"WearWow claims the first end-to-end, mask-free framework for native 2K multi-garment virtual try-on, enabled by adaptive token packing and a dual-reward preference alignment that rescues high-frequency fabric texture.","keywords":["virtual try-on","2K resolution","multi-garment synthesis","mask-free generation","token packing","preference alignment","diffusion models","texture fidelity"],"falsifier":"Run the released checkpoint on real (non-synthetic) paired photos where the same person is photographed both with and without multi-garment outfits, then compare structural and textural metrics; alternatively, measure the distribution gap between synthesized agnostic images and real agnostic images — a large gap would invalidate the mask-free claim.","tokens_in":13462,"feed_emoji":"👗","tokens_out":5312,"duration_ms":51642,"temperature":0.7,"pith_summary":"WearWow sets out to make native 2K multi-garment virtual try-on practical and realistic in one end-to-end, mask-free system. The paper argues that two obstacles block the field: the quadratic memory blow-up from passing many high-resolution garment images into a diffusion transformer, and the MSE denoising objective's spectral bias, which smooths away micro-textures like wool and denim at 2K. Its answer is Adaptive 2D Token Packing, which reorganizes garment images into a compact 2D canvas and prunes background tokens, plus a Multi-dimensional Try-on Reward, which aligns generation away from over-smoothed outputs using both a distribution-anchoring loss and semantic text guidance. If correct, WearWow would be the first open vertical system to make native 2K multi-garment try-on feasible with texture-faithful, mask-free generation, and the release of the WearWow-2K dataset would provide the first benchmark for it.","feed_headline":"Beats commercial APIs at native 2K multi-garment try-on","feed_subtitle":"A mask-free diffusion pipeline packs garment tokens to cut memory and aligns rewards to restore fabric texture at 2048x1536.","key_machinery":"Adaptive 2D Token Packing (ATP) — an algorithmic step that treats each reference garment as a 2D item, packs K items onto a unified canvas by geometric scaling and 2D bin-packing, then uses valid-region masks to prune empty background tokens after patchification; it preserves the original 2D spatial priors via 2D positional encodings and bounds attention cost to O((N_base+L−m)²). Multi-dimensional Try-on Reward (MTR) — a post-training alignment objective combining Cloth Distribution Reward (CDR), an implicit reward defined by the difference between trainable and frozen-reference velocity prediction errors on chosen vs rejected noisy samples, and Semantic Guidance Reward (SGR), a margin-based","core_discovery":"WearWow is a generative diffusion framework claiming state-of-the-art results for native 2K (2048×1536) multi-garment virtual try-on without user-provided masks. It identifies two scaling barriers: self-attention cost grows as O((N_base + K·N)²) when K reference garments are concatenated, making 2K multi-garment training infeasible on current hardware; and the MSE-based diffusion objective acts as a low-pass filter at high resolution, producing plastic-like fabrics. The framework answers with ATP, which packs heterogeneous garment images onto a single 2D latent canvas by 2D bin-packing and then prunes background void tokens, cutting the effective sequence length while preserving 2D positiona","pith_inferences":["If MTR's dual-reward design is the real cause of texture recovery, the same recipe (implicit distribution reward plus semantic margin loss) could transfer to other high-resolution generative tasks where MSE flattening is a problem, such as material synthesis or video super-resolution.","ATP's prune-and-pack idea generalizes to any conditional input with large uninformative regions, not just flat-lay garments; the memory savings should scale with foreground sparsity rather than image count alone.","The central risk is domain transfer: because both training and the 2K test set come from the same synthetic try-off pipeline, a real-user study with true paired before/after photos is the natural next validation. A large distribution gap between synthesized agnostic images and real photos would weaken the mask-free claim.","Independent reproduction should report actual peak VRAM as a function of K at 2K; the paper's complexity bound is plausible but the hardware numbers are not shown, and real-world multi-garment prompts often exceed the 1–6 item range of the dataset."],"forward_implications":["End-to-end training at native 2K with multiple garment conditions becomes computationally tractable, removing the need for cascaded super-resolution and its structural hallucinations.","Micro-level fabric texture (wool, denim, fleece) can be recovered at high resolution by combining distribution-anchored and semantic-guided preference alignment, counteracting MSE spectral bias.","Mask-free inference works: the model learns occlusion and depth ordering directly from synthesized agnostic training targets, so users do not need to provide segmentation masks.","The WearWow-2K dataset (≈100k native 2K triplets, 1–6 items, including footwear and accessories) becomes a benchmark for ultra-HD multi-garment try-on evaluation.","On both single-garment and multi-garment tests, WearWow reports better FID, KID, LPIPS, SSIM and human preference scores than existing open and commercial baselines."],"fun_headline_variants":["Mask-free 2K multi-garment try-on via token packing","Packing garment tokens cuts memory for 2K try-on","Reward alignment restores textures in 2K virtual try-on","Native 2K multi-garment try-on without masks","Token packing makes 2K multi-garment try-on feasible"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that WearWow-2K's 'clothing-agnostic' target images, produced by an automated try-off pipeline and asserted to be physically correct and flawless (details deferred to the supplementary), are faithful stand-ins for real user photos; if they contain artifacts or distributional bias, both mask-free generation and the reported 2K results will not transfer to real-world inputs.","fun_headline_variants_meta":{"raw":{"variants":["Mask-free 2K multi-garment try-on via token packing","Packing garment tokens cuts memory for 2K try-on","Reward alignment restores textures in 2K virtual try-on","Native 2K multi-garment try-on without masks","Token packing makes 2K multi-garment try-on feasible"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000491,"raw_usage":{"total_tokens":2280,"prompt_tokens":804,"completion_tokens":1476,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1400}},"tokens_in":548,"tokens_out":1476,"duration_ms":12125,"temperature":1.0,"reasoning_tokens":1400,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:17:44.351560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released checkpoint on real (non-synthetic) paired photos where the same person is photographed both with and without multi-garment outfits, then compare structural and textural metrics; alternatively, measure the distribution gap between synthesized agnostic images and real agnostic images — a large gap would invalidate the mask-free claim.","supporting_citations":[],"review_version":1}