{"id":"e2bbbef8-0e27-480b-af39-25b2c392a09b","arxiv_id":"2607.21694","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Oxygen-TryOn performs any-item, multi-reference virtual try-on via understanding-driven generation, reporting state-of-the-art scores on public and internal benchmarks.","lead":"Oxygen-TryOn is a new AI model that can dress a person in multiple clothing items from separate photos — shirts, shoes, bags, hats, jewelry — while keeping the person's face, pose, and background unchanged. It claims to beat both commercial rivals such as GPT-Image-2 and open-source models on fidelity, and is built for e-commerce try-on.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gemini-as-judge circularity: the model is RL-trained with the same Gemini 3.1 Pro rubric that produces the TStars-VTON SOTA scores; independent human/judge re-scoring is needed.","rationale":"The central claim is that Oxygen-TryOn achieves state-of-the-art consistency and realism, matching or surpassing proprietary systems. The most load-bearing evidence for this is the TStars-VTON evaluation, which uses Gemini 3.1 Pro as judge. The same model family is a reward source during RL, creating a direct conflict of interest between training objective and evaluation instrument. This is a textbook reward-overoptimization risk: the policy may learn to satisfy the judge's rubric rather than the underlying human preference the rubric is meant to approximate. The paper's rubric design (harmonic/geometric means) mitigates some per-dimension reward hacking, but does not eliminate the fundamental circularity. The human evaluation, while a genuine attempt at external validation, is too small and too underpowered to settle the question, especially given the tiny overall margin. The omission of TStars-Tryon1.0 from re-evaluation further weakens the SOTA claim, since that model's official number is higher. These are addressable issues, so the conditional verdict stands; the concern does not require a change in verdict, but it does require the stated conditions (independent evaluation, released artifacts, error bars) before the SOTA claim can be accepted.","tokens_in":27024,"tokens_out":3826,"duration_ms":39235,"concrete_test":"Independently re-score a random 200–300 sample subset of TStars-VTON (both single and multi) with human raters and/or a VLM judge not used in training (e.g., GPT-5 or Claude) under the same protocol, and compare rankings/margins to Tables 3–4. If Oxygen-TryOn's lead shrinks or reverses relative to GPT-Image-2/Seedream5 Lite, the SOTA claim is an artifact of Gemini-reward overoptimization. Additionally, report 95% CIs for the human eval (Table 7) and for the subset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline SOTA results on TStars-VTON (Tables 3–4) are produced by Gemini 3.1 Pro acting as judge under the official protocol (Section 4.1). The same Gemini 3.1 Pro is one of two reward sources in the RL stage (Section 3.2.3, 'Rubric-guided Gemini reward'). This creates an evaluator-circularity risk: the policy is directly optimized against the rubric of the same judge that later scores it, so high Gemini scores may reflect reward overoptimization rather than human-perceived try-on quality. The paper's own ablation (Table 6) shows RL moves the needle on the in-house bench (e.g., Model-to-Model item consistency 3.286→3.340, usability 80.44→85.43), and the Gemini reward is a component of that RL; it is therefore plausible that part of the measured gain is specific to the judge. The human evaluation (Section 4.3, Table 7) is not a sufficient control: it compares only 985 samples against two proprietary systems and reports no confidence intervals or significance tests, so it cannot rule out a small, judge-specific inflation (the overall margin over GPT-Image-2 is 3.5502 vs 3.5375). A further aggravator: TStars-Tryon1.0, the closest any-item baseline, cannot be reproduced, and its official Overall (9.37) exceeds Oxygen's 9.36, so the single-item SOTA claim rests entirely on excluding this baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Oxygen-TryOn, a fashion-native foundation model for any-item virtual try-on. It reformulates try-on as a multi-reference, understanding-driven generation task rather than mask-based inpainting, building on JoyAI-Image-Edit with a Qwen3-VL MLLM, a Wan VAE, and a 16B MMDiT. Training uses a three-stage recipe: CPT, large-scale SFT, and RL with a hybrid reward combining an in-house try-on reward model and a rubric-guided Gemini 3.1 Pro judge. The authors report a data engine with over 50M raw images filtered to over 10M, multi-item pairing, and diverse category/subject coverage. Experiments cover DressCode/VITON-HD reconstruction metrics, TStars-VTON single- and multi-item VLM-judged scores, a private Oxygen-TryOn Bench, ablations of CPT/SFT/RL, and a human evaluation against two proprietary systems. The central claim is state-of-the-art consistency and realism for single-item try-on and leading multi-item composition, matching or surpassing proprietary systems and open-source models.","tokens_in":27384,"tokens_out":4701,"duration_ms":49726,"significance":"If the claims are substantiated, the contribution is significant: a unified any-item, multi-reference try-on system with a documented data engine and training recipe would be a useful advance over garment-centric inpainting methods. The paper has genuine strengths: the re-evaluation of baselines under a common protocol is transparent; the paired DressCode/VITON-HD results are strong; the architecture choice of separating reference semantics (MLLM) from item appearance (VAE latents) is well motivated; and the detailed description of the data engine and RL design is valuable for reproducibility. However, the headline single-item and multi-item SOTA numbers depend on a judge that is also used as an RL reward source, and the in-house benchmark is not released. These issues are load-bearing for the abstract-level claims, so the current evidence does not yet establish the stated SOTA.","major_comments":[{"comment":"Evaluator circularity: the rubric-guided Gemini 3.1 Pro judge used as one of the two RL reward sources in Section 3.2.3 is the same model used to score the TStars-VTON results in Tables 3–4 under the protocol described in Section 4.1. The policy is therefore directly optimized against the rubric of the judge that later produces the headline scores. This means the high TStars-VTON numbers may reflect reward overoptimization rather than human-perceived try-on quality. The human evaluation in Section 4.3 / Table 7 does not resolve the issue: it covers only 985 samples against two proprietary systems, reports no confidence intervals or significance tests, and the overall margin over GPT-Image-2 is 3.5502 vs 3.5375. Please provide independent human or judge-based evaluation on TStars-VTON (or a public subset), report uncertainty, and separate the evaluation judge from the RL reward model.","section":"§3.2.3 and §4.1"},{"comment":"The single-item SOTA claim is not established because TStars-Tryon1.0, the closest any-item baseline, has an officially reported Overall of 9.37, which exceeds Oxygen-TryOn's 9.36 in the same table. The paper lists this baseline under 'Officially Reported' and excludes it from the highlighted comparison because it cannot be reproduced. But the abstract and Section 1 claim state-of-the-art performance 'across public benchmarks.' Excluding the strongest official baseline changes the comparison set rather than establishing superiority. The claim should be reworded to 'best among re-evaluated baselines,' or the authors should obtain official-scoring results for TStars-Tryon1.0 and include them.","section":"Table 3 and §4.2"},{"comment":"Oxygen-TryOn Bench is an unreleased in-house benchmark, and its scores are produced by a GPT-5 judge with no external validation. The main deployment-oriented claims — usability rates of 86.79% and 85.43% versus 80.35% and 77.58% for GPT-Image-2 — rest entirely on this private benchmark. No error bars, confidence intervals, or significance tests are reported, and some differences are small (e.g., Cloth-to-Model Aesthetics: 3.736 vs. 3.768 for GPT-Image-2). To make the claimed real-world advantage verifiable, the benchmark, judge prompts, and scoring code should be released, and variance or significance information should be reported for at least the headline metrics.","section":"§4.1 and Table 5"},{"comment":"The hybrid reward is described only qualitatively as a 'weighted average' of the Gemini reward and the in-house try-on reward. The fusion weights, normalization, and group-relative reward details are not specified. Since Table 6 shows RL produces large gains (e.g., Model-to-Model usability 80.44→85.43), the relative contributions of the two reward sources are important for interpreting the improvement and for reproducing the recipe. Please report the fusion weights and, ideally, an ablation with each reward source removed.","section":"§3.2.3"}],"minor_comments":[{"comment":"The caption says the TStars-VTON Overall score is 'aggregated by average,' while Section 4.1 states the Overall score is the geometric mean of the four dimensions. Please make these consistent.","section":"Figure 7 caption vs §4.1"},{"comment":"The model name is misspelled as 'Oxygen-T ryOn' in Tables 2 and 3.","section":"Tables 2–3"},{"comment":"Reference [41] duplicates [40] (both are the Qwen-Image technical report), and reference [27] duplicates [3] (both FLUX.2). Please consolidate.","section":"References"},{"comment":"The Overall score is described as a 'weighted average' but the weights are not reported. Please state the weights explicitly.","section":"Table 7"},{"comment":"The text says the model weights will be released, but the paper provides only a project page and no direct release link or license information. Please clarify the release status.","section":"§6.3"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection. The architecture and training recipe are plausible, and the paired reconstruction results are strong, but the headline SOTA claims are not yet supported because the evaluation judge coincides with an RL reward source and the closest official single-item baseline is excluded from the highlighted comparison. These issues are fixable with additional independent evaluation and more cautious claims, so they fall within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the most complete any-item try-on system I've seen from an industrial group, and the technical recipe is real engineering. The headline numbers, though, are scored by a judge the model was optimized against, and the closest baseline's official score is higher than theirs. Those are load-bearing caveats, not quibbles.\n\nWhat's new: the reformulation of try-on as multi-reference, understanding-driven generation rather than mask-based inpainting is a real conceptual shift, and the T-axis MRoPE conditioning without reference tags or index embeddings is a clean way to support variable reference counts. The three-stage CPT-SFT-RL recipe with a hybrid reward is documented in enough detail for others to adapt; the ablation in Table 6 is genuinely instructive. On DressCode and VITON-HD, the re-evaluated paired/unpaired numbers are strong, and the practice of re-running baselines under their own environment and reporting both official and reproduced scores is transparent. The qualitative gallery is impressive, spanning non-garment items and multi-reference compositions.\n\nSoft spots: (1) Evaluator circularity. TStars-VTON is scored by Gemini 3.1 Pro, and Gemini 3.1 Pro is also one of the two RL reward sources. The paper goes some way with anchored rubrics and harmonic/geometric aggregation to resist hacking, but the policy is still optimized against that judge's rubric, so the observed margin on that judge cannot be read as human-perceived quality. The human study is just 985 samples, against two proprietary systems, with no intervals or tests, and the margin over GPT-Image-2 (3.5502 vs 3.5375) is razor-thin. (2) The SOTA claim sits oddly next to TStars-Tryon1.0's official overall 9.37, above their 9.36. They list that number but then compute best/second only among re-evaluated methods, which lets them claim first. Should be qualified. (3) The benchmark, reward model, code, and weights are unreleased, and no error bars anywhere. These are addressable. (4) Minor: the five-plus reference limit is stated in limitations, so the any-item claim is really 'up to four.'\n\nWho's it for: anyone working on try-on, fashion editing, or reward-based diffusion. It deserves a serious referee; the evaluation section needs work, but the underlying approach is credible and the public-benchmark reconstruction results are strong. I'd recommend: send to peer review, with the referee specifically asked to verify the judge circularity and to demand independent human/judge rescoring.","headline":"Strong engineering and a genuinely different multi-reference conditioning scheme, but the headline SOTA claims lean on a self-judged Gemini protocol and an unreleased benchmark; worth refereeing, not worth believing as-is.","tokens_in":27942,"tokens_out":2163,"would_cite":true,"duration_ms":22586,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Oxygen-TryOn claims to be the first any-item, multi-reference virtual try-on model to match or beat leading proprietary systems.","keywords":["virtual try-on","any-item try-on","multi-reference generation","mask-free try-on","data engine","CPT-SFT-RL","hybrid reward","fashion foundation model"],"falsifier":"Run a pre-registered human preference study on the same 1,000 Oxygen-TryOn Bench samples plus a matched sample of TStars-VTON items, with annotators blind to model identity and with no overlap between raters and the reward-model training; if humans do not rank Oxygen-TryOn above GPT-Image-2 and Seedream5 Lite by margins comparable to the judged gap, the central SOTA claim is not supported.","tokens_in":26870,"feed_emoji":"👗","tokens_out":6112,"duration_ms":60087,"temperature":0.7,"pith_summary":"The paper sets out to turn virtual try-on from a narrow, garment-only, mask-based inpainting task into a general 'understanding-driven' generation task: given one or more reference images of any wearable item—product shots or photos of someone wearing it—plus a target subject image, the model dresses the subject while preserving both item detail and identity. It claims that a fashion-native foundation model trained with a three-stage CPT–SFT–RL recipe and a dedicated data engine reaches state-of-the-art consistency and realism on single-item try-on and leads multi-item composition, matching or surpassing proprietary systems such as Nano Banana Pro, GPT-Image-2, and Seedream5 Lite, as well as open-source FLUX.2. The paper positions this as the first system to deliver any-item, multi-reference try-on at this fidelity, and it documents the data engine and training recipe for reproducibility. A sympathetic reader would care because the claim, if true, moves virtual try-on from constrained studio prototypes to a universal, deployable tool for fashion, including accessories, shoes, bags, and full-outfit composition.","feed_headline":"Oxygen-TryOn beats top proprietary models at any-item try-on","feed_subtitle":"Multi-reference, mask-free training keeps every item and the subject intact where general editors break down.","key_machinery":"The load-bearing design is the multi-reference conditioning scheme: subject and references are laid out as one token sequence, with the target region denoised from noise and reference regions kept as clean conditioning, separated only by intervals along the temporal axis of multimodal RoPE—no tags, segment embeddings, or masks. Item appearance travels in VAE latents; item identity and placement travel through the MLLM's semantic tokens and the natural-language instruction. On top of this, the training recipe is carried by a data engine that manufactures item–subject–result triplets, and by the RL stage's hybrid reward that fuses an in-house try-on reward model with a rubric-guided multimodal","core_discovery":"Oxygen-TryOn's central claim is that the right way to build a general try-on model is not to prompt a general image editor or inpaint inside a mask, but to train a fashion-native generator that reasons about what each reference item is, where it goes on the body, and how items layer and occlude. On the paper's own terms, the model—built on a multimodal LLM plus a multimodal diffusion transformer—accepts a variable number of references, encodes them into a shared token sequence, and synthesizes a dressed subject in a single pass, with the instruction binding each item to its position. The paper reports that on public benchmarks (DressCode, VITON-HD, TStars-VTON) and on its in-house Oxygen-Try","pith_inferences":["Because the same rubric-guided judge is used both as an RL reward signal and as the scorer for the headline benchmark numbers, the reported margins may partly reflect alignment to that judge's rubric rather than to human perception; an independent, pre-registered human study with a judge not involved in training would settle whether real users see the same gap.","If the mask-free, understanding-driven formulation is the real source of the gains, the recipe should transfer beyond fashion—for example, to product visualization, interior/decor transfer, and other object-on-subject generation tasks where identity and object fidelity must both be preserved.","The paper's cross-domain demos (anime characters, oil paintings, film stills) suggest the model has learned item–subject wearing relations rather than a studio-photo prior; a systematic test on stylized domains would show how far that generalization extends.","The stated bottleneck of five-plus references predicts a concrete scaling path: apply the same CPT–SFT–RL recipe to a base model with longer context or larger capacity, and evaluate whether multi-item consistency improves in step."],"forward_implications":["Any wearable item—shoes, bags, hats, jewelry, as well as clothing—can be transferred onto a subject from either clean product shots or in-the-wild worn-on photos, in one pass.","Multi-item composition lets a user assemble a full outfit (top, bottom, coat, hat, shoes, bag) while the model resolves layering and occlusion, not just swapping one garment.","The same pass can follow general editing directives such as pose or background changes, so try-on and photo editing no longer require separate models.","The documented data-engine and CPT–SFT–RL recipe gives the open community a path to reproduce or surpass closed proprietary try-on quality.","The paper's own limitation section says five or more references still cause item confusion and dropped items, so the claimed capability is currently bounded at around four references."],"fun_headline_variants":["Oxygen-TryOn beats top proprietary models at any-item try-on","Fashion-native try-on model outdoes GPT-Image-2, Seedream5","Mask-free multi-reference try-on: new SOTA on public benchmarks","Any-item try-on unified model surpasses leading open and closed systems","Oxygen-TryOn: single pass, any item, beats Nano Banana Pro"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the rubric-guided judge used to produce the headline state-of-the-art scores is an honest external measure of try-on quality—but that same judge also supplies part of the RL reward during training; if the model has fitted the judge's rubric rather than human perception, the claimed margins over competitors would shrink or vanish under human evaluation, and the paper's own human study covers only 985 samples against two systems.","fun_headline_variants_meta":{"raw":{"variants":["Oxygen-TryOn beats top proprietary models at any-item try-on","Fashion-native try-on model outdoes GPT-Image-2, Seedream5","Mask-free multi-reference try-on: new SOTA on public benchmarks","Any-item try-on unified model surpasses leading open and closed systems","Oxygen-TryOn: single pass, any item, beats Nano Banana Pro"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001171,"raw_usage":{"total_tokens":4743,"prompt_tokens":874,"completion_tokens":3869,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":3770}},"tokens_in":618,"tokens_out":3869,"duration_ms":24094,"temperature":1.0,"reasoning_tokens":3770,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:03:11.150681+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a pre-registered human preference study on the same 1,000 Oxygen-TryOn Bench samples plus a matched sample of TStars-VTON items, with annotators blind to model identity and with no overlap between raters and the reward-model training; if humans do not rank Oxygen-TryOn above GPT-Image-2 and Seedream5 Lite by margins comparable to the judged gap, the central SOTA claim is not supported.","supporting_citations":[],"review_version":1}