{"id":"419ee9af-57de-4578-8c5c-f0a78cf32709","arxiv_id":"2411.18350","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TryOffDiff reconstructs canonical garment images from single photos of clothed people and reports large gains over adapted baselines on VITON-HD, with cross-dataset evidence from Dress Code.","lead":"This paper introduces Virtual Try-Off, a task that turns a photo of a clothed person into a clean, standardized garment photo, and proposes TryOffDiff, a Stable Diffusion model conditioned on SigLIP image features. The authors report that TryOffDiff outperforms adapted try-on and pose-transfer baselines on VITON-HD, and they argue that DISTS is a more reliable quality metric than SSIM for this task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on the dataset garment image G being the unique canonical VTOFF target, and this uniqueness is assumed rather than validated; without a human-rating check, the headline DISTS improvements may measure fit to one dataset convention, not general reconstruction fidelity.","rationale":"The reader's weakest assumption is the same one I identify: the dataset's G is treated as the unique canonical VTOFF target, without validation of uniqueness or human agreement. This is genuinely load-bearing because the paper's central contribution is not just a method but a proposed evaluation task, and the headline quantitative wins are all defined relative to G. If G is not the natural unique target, the absolute fidelity claims and the metric-comparison claims lose some force, and the DISTS-based conclusions in Sec. 3.1 and Tab. 4 would need re-grounding. Other concerns, such as the absence of error bars and the lack of baseline evaluations on Dress Code, are real but secondary: the reported margins are large, and the paper discloses that baselines were not evaluated on Dress Code. I credit the authors for cleaning the leaked/duplicate VITON-HD pairs, releasing code and models, and reporting a broad metric suite, which supports the relative comparison even if the task definition needs tightening. The concern does not invalidate the paper; it argues for the same conditional verdict: accept the method as a strong baseline and benchmark contribution, but withhold unconditional acceptance of 'high-fidelity reconstruction' until the target definition and metric ordering are validated with human raters.","tokens_in":13368,"tokens_out":5878,"duration_ms":57788,"concrete_test":"Run a human evaluation on a random sample of VITON-HD test items. First, for garment identity, show the input person image I and two reconstructions (TryOffDiff vs. each baseline) and ask raters which is more faithful; compare human preference rates with the DISTS ordering and test significance. Second, for target uniqueness, show I alongside several plausible canonical renderings of the same garment (or ask raters to rank generated candidates against the dataset G) and measure inter-rater agreement with Fleiss' kappa. If human preference does not reproduce TryOffDiff's DISTS advantage, or if raters disagree substantially about which canonical image is 'the same garment', the headline claim should be restricted to 'best match to dataset G' rather than general reconstruction fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption appears in Sec. 3.1: VTOFF 'aims to generate a standardized garment image G from I', and the model is trained and scored against the dataset's G as though P(G|C) had a single well-defined target. This is never tested. In VITON-HD and Dress Code, G is one product-style rendering; the same physical garment can plausibly be canonicalized with different front-view rotations, sleeve lengths, tuck states, or stretch, and occluded back/side portions are not determined by I. The evaluation then compares every model to this single G, so the reported numbers (e.g., DISTS 20.3 vs. 28.2) measure agreement with the dataset's specific canonicalization, not 'high-fidelity garment reconstruction' in an unqualified sense. The paper's own metric-suitability argument (Fig. 3, Tab. 4) uses hand-picked degradations and provides no human ratings or inter-rater agreement for VTOFF; DISTS is a generic IQA metric and is not validated for judging garment identity or canonical correctness. Consequently, the central claim could partly reflect learning the dataset's canonical style. A model producing a different but equally valid canonical garment would be penalized by the current protocol, and the claimed cross-dataset superiority could similarly be a better match to VITON-HD's canonicalization rather than better fidelity. The paper does usefully clean the dataset and release code/models, and the relative comparisons are internally consistent, but the task definition itself is the least secure link in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Virtual Try-Off (VTOFF), a task of generating a standardized garment image from a single photo of a clothed person, and proposes TryOffDiff, a Stable Diffusion model conditioned by SigLIP image embeddings through cross-attention adapters. Experiments on VITON-HD and Dress Code upper-body garments report that TryOffDiff outperforms adapted VTON, pose-transfer, and image-conditioning baselines on VITON-HD, with ablations supporting the SigLIP encoder, adapter, and pretrained U-Net initialization. The paper also argues that conventional metrics like SSIM are unsuitable for VTOFF and recommends DISTS as the primary metric.","tokens_in":13646,"tokens_out":4939,"duration_ms":42989,"significance":"The VTOFF task is timely and relevant to e-commerce and generative-model evaluation, and the paper makes a useful contribution by formalizing the task, cleaning the VITON-HD dataset, and releasing code and models. The architecture is simple and the ablation study is informative. However, the central empirical claims are weakened by two gaps: no baseline numbers are reported on Dress Code despite the abstract claiming superiority there, and the evaluation protocol assumes a unique canonical garment target without human validation. The relative comparison on VITON-HD is internally consistent and the presented model is plausible, but the absolute 'high-fidelity reconstruction' and metric-reliability claims need stronger support.","major_comments":[{"comment":"The abstract and §4 state that experiments on Dress Code show TryOffDiff outperforms baselines, but Table 2 contains only TryOffDiff rows, and the text says baselines were not evaluated on Dress Code due to poor initial performance. This does not support the claim. Please add Dress Code baseline results or revise the abstract, introduction, and conclusion to restrict the outperformance claim to VITON-HD.","section":"Abstract; §4 Quantitative Results; Table 2"},{"comment":"The paper defines VTOFF as generating a standardized garment image G and then trains and evaluates against the dataset's G as if it were the unique target. The same physical garment can be canonically presented in different front-view rotations, sleeve states, tuck states, or stretch, and occluded back/side regions are underdetermined. No human-rater study, inter-rater agreement, or multi-target analysis is provided. Therefore, the headline numbers (e.g., DISTS 20.3 vs. 28.2) measure fit to one dataset-specific canonicalization; the unqualified 'high-fidelity reconstruction' claim requires validation that the target is indeed the natural unique canonical garment.","section":"§3.1 Problem Formulation; evaluation protocol"},{"comment":"The paper's argument that DISTS is more reliable than SSIM for VTOFF is based on a small set of hand-picked degradation examples and on metric values themselves, not on human perceptual judgments. Without a user study that collects ground-truth similarity ratings for VTOFF outputs, the claim that SSIM 'inadequately reflects reconstruction quality' and that DISTS is 'reliable' is a conjecture. At minimum, report confidence intervals and statistical significance for the metric differences in Tables 1 and 2.","section":"§3.1 Performance Measures; Fig. 3; Tab. 4"},{"comment":"The adapted baselines appear to be placed at a disadvantage relative to TryOffDiff: CatVTON is given a plain white reference image and a handcrafted mask, IP-Adapter is initialized from a mask-like image, and ViscoNet is given masked conditioning without fine-tuning, whereas TryOffDiff is trained on the target task. This makes it difficult to separate task-adaptation quality from method quality. Please provide an oracle/upper-bound or a fine-tuned baseline comparison, or explicitly discuss this limitation as a bound on the strength of the comparative claims.","section":"§4 Baseline Approaches; App. C"}],"minor_comments":[{"comment":"Appendix C contains the typo 'architecturla' in the IP-Adapter description; please correct it.","section":"App. C"},{"comment":"The abbreviation FDCLIP is used without definition in the text; please clarify that it is a CLIP-based Fréchet distance.","section":"Tables 1 and 2"},{"comment":"The SAM point-prompt selection used in baseline post-processing is not specified; please describe it for reproducibility.","section":"App. C"},{"comment":"References [21] (DressCode, a text-to-garment method) and [30] (Dress Code dataset) are easily confused; please consider introducing explicit notational or stylistic distinctions.","section":"References"},{"comment":"The selection criteria for the hand-picked distortion cases in Figure 3 are not stated; a brief description of how representative cases were chosen would strengthen the metric-suitability discussion.","section":"Fig. 3 and Tab. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is best framed as a task proposal plus a strong first method on VITON-HD. The Dress Code claim should be corrected or supported before publication. The lack of human evaluation is the main risk; a revision that adds a small perceptual study and either re-runs baselines on Dress Code or tempers the corresponding claims would materially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real contribution. The paper formalizes virtual try-off, cleans the VITON-HD benchmark (removes duplicates and train/test leakage), builds a straightforward SigLIP-conditioned diffusion pipeline, and shows it beats adapted VTON/pose-transfer baselines on every metric they report on VITON-HD. It also releases code and models. That is more than a lot of papers in this space.\n\nThe new thing is the task definition and the evaluation setup, not the architecture. Using SigLIP tokens in cross-attention of Stable Diffusion is a sensible adaptation, and the ablations show each component helps. The cross-dataset experiment is also useful, even though it only evaluates their own model. I agree with the reader's conditional verdict.\n\nThe soft spots are real but addressable. The abstract overclaims: 'outperforms adapted pose transfer and VTON baselines' is only true on VITON-HD. On Dress Code they explicitly say baselines were not run because of poor initial results. That should be fixed in a revision, but the honesty in the body is a point in their favor.\n\nThe bigger issue is the one the stress-test note flags: the target G is assumed to be the unique canonical garment image. The paper's own framing says VTOFF has a 'clear and precise target,' but G is one dataset-specific product-style rendering. The same garment could reasonably be canonicalized differently, and the model is trained and scored against that single G. Without human ratings or an analysis of ambiguity, the headline DISTS numbers measure fit to this particular canonicalization, not general reconstruction fidelity. This does not invalidate the benchmark — within the protocol the comparisons are fair — but it does mean the phrase 'high-fidelity garment reconstruction' is too strong as a general claim.\n\nAlso, the DISTS argument is built on hand-picked degradations and averaged dataset numbers, but no human study. That's a weakness, not a fatal one, because the degradation patterns are clear and they do run them over the full test set. Still, a small perceptual study would strengthen the metric recommendation.\n\nAll in all, this deserves serious peer review. The task is new, the benchmark is useful, and the method is a credible baseline. I'd ask for (1) a human-rating check on canonical target ambiguity, (2) a corrected abstract or actual Dress Code baselines, and (3) error bars. I'd cite this if I worked on garment generation.","headline":"Solid new task and benchmark with an honest method, but the 'standardized garment' target is assumed unique and the abstract overstates the Dress Code results.","tokens_in":14208,"tokens_out":2543,"would_cite":true,"duration_ms":23788,"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":"TryOffDiff turns a photo of a clothed person into a clean, catalog-style garment image, outperforming all baselines it was tested against.","keywords":["virtual try-off","garment reconstruction","diffusion models","image conditioning","SigLIP","DISTS","VITON-HD","Dress Code"],"falsifier":"Generate or collect a set of person photos for which two different canonical garment images are both plausible (e.g., the same garment photographed in different standard poses or with different acceptable fittings), then run TryOffDiff and ask human raters which output matches the garment's true appearance. If a substantial share of raters prefer an output that is not the dataset's ground truth, or if the DISTS distance between the two plausible targets is comparable to the model's error, the claim that the task has a well-defined evaluation target is undercut.","tokens_in":13131,"feed_emoji":"👕","tokens_out":4795,"duration_ms":38848,"temperature":0.7,"pith_summary":"This paper introduces Virtual Try-Off (VTOFF), the task of turning a single photo of a clothed person into a clean, standardized garment image of the kind used in e-commerce catalogs. It proposes TryOffDiff, a Stable Diffusion model whose text conditioning is replaced by image features from a SigLIP encoder, so the denoising process is guided by the visual details of the person's garment. On VITON-HD and Dress Code, the method reports better scores than adapted virtual try-on and pose-transfer baselines on every metric, including a 7.9-point improvement over the strongest baseline on the perceptual DISTS score. The paper also argues that SSIM and its variants are poor measures of garment reconstruction quality and that DISTS should be used instead. A sympathetic reader would take away that high-fidelity garment extraction from a single photo is feasible with a straightforward diffusion adaptation.","feed_headline":"From one clothed photo, TryOffDiff renders a clean garment image","feed_subtitle":"It beats the strongest virtual try-on baseline by 7.9 points on the perceptual DISTS metric and transfers across datasets.","key_machinery":"The central mechanism is image-conditioned latent diffusion: SigLIP-B/16-512 extracts a sequence of 1024 visual tokens, a linear layer plus layer normalization reduces them to 77 embeddings, and those embeddings replace text embeddings as the keys and values in the cross-attention layers of the denoising U-Net, while the SigLIP encoder and the VAE remain frozen. This lets a pretrained text-to-image diffusion model be repurposed for image-guided garment reconstruction without architectural changes. The second piece of machinery is the DISTS metric, which combines VGG-based structural and textural similarity and is used as the primary evaluation measure because SSIM fails to penalize color distortions and missing details.","core_discovery":"TryOffDiff claims that a latent diffusion model can reconstruct canonical garment shapes, textures, patterns, logos, and occluded regions from a single clothed-person image by conditioning the denoising U-Net on the full token sequence of a SigLIP image encoder rather than on text. The authors finetune Stable Diffusion v1.4 with a linear-plus-layer-norm adapter that compresses SigLIP's 1024 tokens to the 77 slots the cross-attention layers expect, and they report that this outperforms adapted pose transfer, view synthesis, and virtual try-on baselines on all full-reference and no-reference metrics. On VITON-HD-test, TryOffDiff reaches a DISTS of 20.3 versus 28.2 for CatVTON; in cross-dataset evaluation, a model trained on Dress Code and tested on VITON-HD still beats every VITON-HD baseline. The authors also show that SSIM ranks clearly distorted images above perceptually faithful ones, and present DISTS as the more reliable evaluation tool.","pith_inferences":["The same conditioning recipe could extend to other object classes with standardized views, such as shoes, bags, or furniture, provided paired single-object/canonical-image data are available.","The 77-token bottleneck is likely why the paper notes difficulty with intricate logos and text; increasing token capacity or using a higher-resolution image encoder might close that gap.","A human perceptual study comparing TryOffDiff output, ground truth, and alternative canonical renderings would directly test whether DISTS's ranking matches human preference, which the paper does not include.","If VTOFF outputs are used to train or evaluate try-on models, artifacts in reconstruction could propagate; reporting failure cases on complex patterns is a step toward bounding that."],"forward_implications":["If the reported numbers hold, VTOFF becomes a practical way to generate catalog-ready garment images from customer or street photos, reducing the need for studio product photography.","VTOFF gives generative model evaluation a well-defined target image, so reconstruction fidelity can be measured directly instead of relying on distribution-level metrics like FID.","The cross-dataset results imply the approach transfers across pose and domain shifts (full-body to half-body), which matters for real-world deployment.","The DISTS finding suggests that existing VTON evaluations using SSIM may be overstating quality, and that perceptual structure-texture metrics should become standard.","VTOFF and VTON form a cycle, so a try-off model could be used to synthesize training data or impose consistency losses for try-on models."],"supporting_citations":[{"why":"Provides the pretrained latent diffusion backbone that TryOffDiff finetunes and adapts.","marker":"[35]"},{"why":"Supplies the visual features that replace text conditioning in cross-attention.","marker":"[61]"},{"why":"Defines the DISTS metric used as primary evaluation and to expose SSIM's failures.","marker":"[11]"},{"why":"The strongest adapted VTON baseline, which TryOffDiff outperforms by 7.9 DISTS points.","marker":"[9]"},{"why":"Source of VITON-HD person-garment pairs for training and testing.","marker":"[7]"},{"why":"Provides Dress Code pairs used for within- and cross-dataset evaluation.","marker":"[30]"},{"why":"Representative image-conditioned diffusion baseline adapted for VTOFF.","marker":"[59]"}],"fun_headline_variants":["One photo to clean garment: TryOffDiff beats baselines","TryOffDiff extracts garment from single photo, outperforms VTON","From one clothed person to clean garment via diffusion","Clean garment from one photo: TryOffDiff sets new standard","Single-photo garment extraction wins with diffusion model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the dataset pair (person photo, clean garment image) has a single correct 'try-off' target, so reconstruction quality can be measured by comparing the generated image with that one ground truth; if several canonical garments are equally valid, the reported scores measure agreement with a dataset-specific choice rather than garment fidelity.","fun_headline_variants_meta":{"raw":{"variants":["One photo to clean garment: TryOffDiff beats baselines","TryOffDiff extracts garment from single photo, outperforms VTON","From one clothed person to clean garment via diffusion","Clean garment from one photo: TryOffDiff sets new standard","Single-photo garment extraction wins with diffusion model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1248,"prompt_tokens":933,"completion_tokens":315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":234}},"tokens_in":549,"tokens_out":315,"duration_ms":3993,"temperature":1.0,"reasoning_tokens":234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:17:19.418668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate or collect a set of person photos for which two different canonical garment images are both plausible (e.g., the same garment photographed in different standard poses or with different acceptable fittings), then run TryOffDiff and ask human raters which output matches the garment's true appearance. If a substantial share of raters prefer an output that is not the dataset's ground truth, or if the DISTS distance between the two plausible targets is comparable to the model's error, the claim that the task has a well-defined evaluation target is undercut.","supporting_citations":[{"cited_title":"Viton-hd: High- resolution virtual try-on via misalignment-aware normalization","cited_arxiv_id":null,"evidence_quote":"Source of VITON-HD person-garment pairs for training and testing."},{"cited_title":"Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"Representative image-conditioned diffusion baseline adapted for VTOFF."},{"cited_title":"Image quality assessment: Unifying structure and texture similarity","cited_arxiv_id":null,"evidence_quote":"Defines the DISTS metric used as primary evaluation and to expose SSIM's failures."},{"cited_title":"Catvton: Concatenation is all you need for virtual try-on with diffusion models","cited_arxiv_id":null,"evidence_quote":"The strongest adapted VTON baseline, which TryOffDiff outperforms by 7.9 DISTS points."}],"review_version":1}