{"id":"eac25221-97f4-426a-abb8-70cddf5185b4","arxiv_id":"2508.13632","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"OmniTry extends virtual try-on to all wearable objects with a mask-free two-stage pipeline, reporting gains on localization and ID preservation.","lead":"OmniTry is a unified virtual try-on system that handles any wearable object, not just clothes, and does not require a user-drawn mask. It uses a two-stage training pipeline for localization and appearance transfer, and reports improved localization and identity preservation over existing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empty-mask inpainting may not actually learn localization; the abstract provides no evidence that Stage 1 does more than reconstruct inputs or exploit dataset bias.","rationale":"The reader's weakest assumption was that empty-mask inpainting learns plausible object placement and transfers to new objects after few paired examples. My stress-test converges on the same load-bearing concern: with an empty mask, the inpainting objective has no incentive to generate a new object at a semantically valid position unless the training data and loss are carefully designed to enforce it. Without specifying those details or providing stage-wise localization metrics, the abstract cannot support the claim that the first stage learns mask-free localization. Since only the abstract was available and no code or benchmark is public, the evidence is insufficient for ACCEPT or CONDITIONAL. The reader's UNVERDICTED verdict is appropriate; I found no independent reason to increase or decrease confidence, so the verdict remains unchanged. The proposed concrete test would resolve the concern if the authors release the Stage-1 model and benchmark, as promised.","tokens_in":763,"tokens_out":1868,"duration_ms":21890,"concrete_test":"Request the Stage-1 checkpoint and evaluate localization on held-out object classes without any Stage-2 fine-tuning. Compare predicted placement against ground-truth human-annotated positions and against a trivial baseline that always places objects at the most frequent corpus position. Also compute the Stage-1 reconstruction loss on unpaired images: if the model achieves near-zero loss by copying inputs while ignoring the empty mask, it is not learning placement. If public code/checkpoints are unavailable, this claim remains unverifiable; rerunning the same check on any released subset would be a minimal sufficient test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim hinges on Stage 1: repurposing an inpainting model with an empty mask on unpaired portraits should learn mask-free localization of arbitrary wearable objects. The load-bearing concern is that a standard inpainting objective with an empty mask imposes no reconstruction constraint that forces the model to introduce a new object at a semantically appropriate location. With an empty mask, the model can satisfy the loss by reconstructing the input with minimal change, or it can learn a generic prior about where accessories appear in the training corpus. In either case, Stage 1 does not learn a task-specific localizer; reported 'quick convergence' in Stage 2 may then be dominated by the paired fine-tuning data, undercutting the claim that unpaired images provide useful localization pretraining. The abstract states the setup but does not specify the training loss, mask construction, or any localization evaluation before fine-tuning. This is an evidence gap rather than an internal inconsistency, but it is precisely the premise that must hold for the proposed two-stage pipeline to be a genuine advance over ordinary inpainting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The abstract proposes OmniTry, a two-stage framework for mask-free virtual try-on of arbitrary wearable objects (garments, jewelry, accessories). Stage 1 trains an inpainting model on large-scale unpaired portrait images using an 'empty mask' to learn object localization; Stage 2 fine-tunes on paired images to preserve object appearance. The authors report quick convergence with few paired samples and better performance than existing methods on both object localization and ID-preservation, evaluated on a benchmark of 12 wearable-object classes. Code, model weights, and the benchmark are promised for public release. The review is based solely on the abstract, as the full text was not available.","tokens_in":1012,"tokens_out":2083,"duration_ms":22605,"significance":"If the claims hold, OmniTry would broaden VTON from garments to arbitrary wearables, remove the need for segmentation masks at inference, and reduce paired-data requirements through unpaired pretraining. The planned public release of code, weights, and a 12-class benchmark is a concrete contribution that could facilitate reproducibility. However, none of these contributions can currently be assessed: the abstract contains no evaluation protocol, no numerical results, no baseline definitions, and no analysis of why empty-mask inpainting should learn semantically meaningful localization. The significance is therefore potential rather than established.","major_comments":[{"comment":"The central performance claim—'OmniTry shows better performance on both object localization and ID-preservation compared with existing methods'—is stated with no evaluation protocol, no numeric results, no baseline definitions, and no statistical detail. As written, the claim is unverifiable. The full manuscript must provide the benchmark construction, metric definitions, baseline implementations, and comparison tables before the claim can be assessed.","section":"Abstract, first paragraph"},{"comment":"The load-bearing premise is that repurposing an inpainting model with an empty mask learns mask-free localization of arbitrary wearable objects. This is not self-evident: an inpainting objective with an empty mask imposes no reconstruction constraint that forces insertion of a new object at a semantically suitable location; the model could satisfy the objective by reconstructing the input with minimal change or by exploiting dataset biases about typical accessory placement. The abstract does not specify the training loss, mask construction, or any localization evaluation before fine-tuning. Without such details, the claim that Stage 1 contributes task-specific pretraining, and that Stage 2's 'quick convergence' is attributable to it, remains an evidence gap rather than an established result.","section":"Abstract, Stage 1 (empty-mask inpainting)"},{"comment":"The 'comprehensive benchmark consisting of 12 common classes of wearable objects' is not described: which classes, how images were sourced (in-shop vs. in-the-wild), class balance, train/test splits, paired versus unpaired construction, and the specific metrics for object localization and ID-preservation are all unspecified. This prevents independent replication and comparison. The abstract should at minimum state the evaluation protocol or refer to a detailed experimental section.","section":"Abstract, evaluation benchmark"}],"minor_comments":[{"comment":"'Quick convergence even with few paired samples' is vague. Please specify the number of paired samples, fine-tuning iterations/epochs, and the convergence criterion or learning-curve comparison.","section":"Abstract, Stage 2"},{"comment":"'Mask-free setting' should be clarified: does it mean no input segmentation mask at inference, no mask at all during training, or only the empty-mask trick? This affects the practical contribution.","section":"Abstract, terminology"},{"comment":"For arbitrary wearables, 'ID-preservation' is ambiguous: does it measure preservation of the object's appearance identity, the person's identity, or both? A precise metric definition is needed.","section":"Abstract, metrics"}],"recommendation":"uncertain","confidential_remarks":"This review is based only on the abstract, as the full text was not available. My 'uncertain' verdict reflects an evidence gap, not a finding of error: the central claims are plausible but unsupported in the provided text. If the full manuscript contains the missing protocol, numerical results, and an analysis of the empty-mask mechanism, the paper may well be publishable. I recommend the editor ensure the complete submission is made available to future referees, since abstract-only review cannot support accept or reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the idea is genuinely new in the VTON space. Existing work mostly assumes clothes and masks; OmniTry removes both and extends to jewelry, accessories, and other wearables. The trick of repurposing an inpainting model with an empty mask to learn mask-free localization from unpaired portraits is neat, and the two-stage design—unpaired pretraining then paired fine-tuning—is a sensible answer to scarce paired data. The promised benchmark of 12 wearable classes with in-shop and in-the-wild images is also a useful contribution if it actually ships.\n\nWhat I can't assess from the abstract: there are no numbers, no baselines, no evaluation protocol. The sentence about better localization and ID-preservation is a claim, not a result. The quick-convergence observation is similarly unquantified. None of that is evidence of error, but it means the central performance claims are unverifiable from the available text.\n\nThe stress-test concern about empty-mask inpainting is a real risk, not a manufactured one. With an empty mask, an inpainting objective can be satisfied by reconstructing the input with minimal change, or by exploiting a dataset prior about where accessories appear, without learning a task-specific localizer. The paper may well address this with a different loss, positional embeddings, or a localization evaluation between stages—but the abstract doesn't say. So I'd call it an evidence gap rather than a demonstrated flaw.\n\nIf the full paper addresses that gap with reasonable diagnostics, this is a solid contribution for people working on fashion image editing and VTON. The idea is coherent and the authors promise code and data, which is good practice. I would not cite it based on the abstract alone, and I'd want to see the full experiments before trusting the mechanism. But it's not a desk-reject; it deserves a serious referee if the full manuscript reports the experiments properly.","headline":"A clever mask-free extension of virtual try-on to all wearables, with a live question about what the empty-mask inpainting actually learns; the abstract alone can't support the performance claims.","tokens_in":1468,"tokens_out":2138,"would_cite":false,"duration_ms":23604,"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":"OmniTry claims a single mask-free pipeline can put any wearable object—jewelry, glasses, hats—on a person, learning placement from unpaired portraits and appearance from a few paired examples.","keywords":["virtual try-on","mask-free","inpainting","unpaired data","object localization","identity preservation","wearable accessories","benchmark"],"falsifier":"Take a category deliberately withheld from both the unpaired pretraining and the paired fine-tuning, then run OmniTry on it. If the model places the object in anatomically implausible locations (e.g., a ring floating off the finger) as often as a baseline inpainting model, the claim that stage one learns reusable localization from empty-mask inpainting is not supported.","tokens_in":708,"feed_emoji":"🕶️","tokens_out":3375,"duration_ms":30244,"temperature":0.7,"pith_summary":"The paper sets out to broaden virtual try-on from clothes to any wearable object and to remove the need for manually drawn masks. To avoid costly paired data, it trains an inpainting model on large sets of unpaired portraits with an empty mask, letting the model decide where an object belongs. A short fine-tuning stage on a few paired images then transfers the exact object's appearance. The authors report that the resulting system, OmniTry, localizes objects better and preserves their identity better than existing methods on a 12-class benchmark, and that the first stage converges quickly with very few paired samples.","feed_headline":"OmniTry fits any wearable item on a portrait without masks","feed_subtitle":"Two-stage pipeline learns placement from unpaired photos, then preserves identity with few paired examples.","key_machinery":"The key mechanism is the repurposing of an inpainting model with an empty mask. Instead of providing a segmentation mask that tells the model where to place the item, the model receives an empty mask (an all-zero mask covering the entire image) and must infer the natural placement of a wearable object from the portrait itself. This empty-mask inpainting acts as a self-supervised localization learner on unpaired data; the subsequent paired fine-tuning transfers identity.","core_discovery":"OmniTry's central claim is that mask-free virtual try-on of arbitrary wearable objects is achievable through a two-stage recipe. In the first stage, an inpainting model is repurposed: given an empty mask and a portrait, it is trained on unpaired images to hallucinate wearable objects in plausible positions, which teaches the model object localization. In the second stage, this pretrained model is fine-tuned on paired images so that the actual object's appearance is preserved. The paper argues that this separation—localization from unpaired data, appearance from paired data—makes the method practical because paired try-on data is scarce for new categories, and that it outperforms existing app","pith_inferences":["A testable extension: if stage one truly learns localization semantics from empty-mask inpainting, the same pipeline may transfer to objects not in the 12 benchmark classes (e.g., temporary tattoos or footwear) without any paired data for those classes.","The empty-mask inpainting trick could probe whether the underlying model already encodes object–body spatial priors; ablating by category (e.g., objects never seen in pretraining) would reveal whether the localization is learned from unpaired portraits or emerges from the base model.","The reported quick convergence is consistent with the view that stage one acts as a strong prior; a critical test is measuring how few paired samples are needed per new category and whether that number grows with object complexity.","One limitation the paper does not discuss explicitly: the mask-free setting trades automation for control, since users cannot specify a desired placement when the model's localization does not match their intent."],"forward_implications":["If correct, virtual try-on no longer needs per-category segmentation masks, simplifying deployment for glasses, jewelry, hats, bags, and similar accessories.","The two-stage design means adding a new wearable category requires only a small number of paired examples, since stage one already knows how to localize.","The method claims to preserve object identity better, meaning the rendered item looks like the actual product photo, which matters for e-commerce.","The benchmark with 12 classes provides a common testbed for mask-free virtual try-on of non-clothing items.","Quick convergence suggests the unpaired pretraining is doing real work; a few paired samples suffice to adapt appearance."],"supporting_citations":[],"fun_headline_variants":["No masks, any wearable: OmniTry's virtual try-on","OmniTry: try on jewelry, accessories, anything—no mask","Mask-free virtual try-on for all wearable objects","Two-stage training lets OmniTry place any accessory","OmniTry: mask-free try-on for any wearable item"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that an inpainting model trained with an empty mask on unpaired portraits will learn where wearable objects belong, and that this localization knowledge transfers to new objects after fine-tuning on only a few paired images.","fun_headline_variants_meta":{"raw":{"variants":["No masks, any wearable: OmniTry's virtual try-on","OmniTry: try on jewelry, accessories, anything—no mask","Mask-free virtual try-on for all wearable objects","Two-stage training lets OmniTry place any accessory","OmniTry: mask-free try-on for any wearable item"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1811,"prompt_tokens":774,"completion_tokens":1037,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":953}},"tokens_in":518,"tokens_out":1037,"duration_ms":10459,"temperature":1.0,"reasoning_tokens":953,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:56:13.063287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a category deliberately withheld from both the unpaired pretraining and the paired fine-tuning, then run OmniTry on it. If the model places the object in anatomically implausible locations (e.g., a ring floating off the finger) as often as a baseline inpainting model, the claim that stage one learns reusable localization from empty-mask inpainting is not supported.","supporting_citations":[],"review_version":1}