{"id":"27ee2bc2-c37b-4714-830a-51433a2bae4b","arxiv_id":"2504.13078","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MGT is a unified diffusion model that reconstructs clean garment images from photos of dressed people for three garment classes, and it improves person-to-person virtual try-on when paired with OOTDiffusion.","lead":"This paper extends the virtual try-off task to multiple garment categories with a single diffusion model that reconstructs standardized upper-body, lower-body, and dress garments from clothed-person photos. The model also plugs into a try-on pipeline for person-to-person garment transfer, with claims of competitive quality and improved preservation of the target person's attributes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VITON-HD 'state-of-the-art' claim is unsupported: Table 2 baselines come from different test splits and the table itself shows MGT worse on most metrics.","rationale":"The paper's central contribution is a single diffusion model that reconstructs upper-body, lower-body, and dress garments at quality comparable to per-category models. Table 1 is a reasonable controlled comparison: the authors trained category-specific TryOffDiff variants on DressCode, so the same data, architecture, and conditioning are used across arms. MGT is slightly worse on most metrics but close enough that 'comparable' is plausible, though error bars or multiple seeds would strengthen this. The load-bearing weakness is the VITON-HD generalization claim. The abstract explicitly advertises state-of-the-art VTOFF results on VITON-HD, but Table 2 compares MGT against baseline numbers taken from original papers that used a different, non-deduplicated test split. Since FID and related metrics are distribution-sensitive, MGT's FID improvement cannot be interpreted without recomputing all methods on one shared split. Moreover, on the metrics that are less distribution-dependent, MGT is consistently worse than TryOffDiff and sometimes much worse than TryOffAnyone, so the table is internally inconsistent with the abstract's 'state-of-the-art' phrasing. The reader's identified weakest assumption, the SigLIP adapter, is not the most load-bearing issue: any deficiency in SigLIP-based conditioning would affect the per-category baselines equally, so Table 1 still isolates the multi-garment contribution. The VITON-HD comparison, by contrast, is the sole support for a headline claim and is not a valid apples-to-apples evaluation. The appropriate verdict remains CONDITIONAL rather than REJECT because the multi-garment DressCode result is likely salvageable and the paper releases code and models; the condition is to rerun the VITON-HD baselines on MGT's exact test split and to select inference hyperparameters on a validation set rather than on test sets used for final scoring.","tokens_in":10976,"tokens_out":7133,"duration_ms":58683,"concrete_test":"Re-run TryOffDiff and TryOffAnyone with their released checkpoints on MGT's exact preprocessed VITON-HD test split (1,990 pairs after duplicate/leak removal), using identical preprocessing, the Euler scheduler, and the same hyperparameter-selection protocol (e.g., a held-out validation split) as MGT. If MGT's FID/FDCLIP advantage disappears or its DISTS/LPIPS deficit persists on this shared split, the abstract's state-of-the-art and cross-domain generalization claims should be downgraded to 'competitive on selective metrics' or removed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4.1 states that MGT is evaluated on a deduplicated VITON-HD test set of 1,990 pairs, while Table 2's footnote says baseline values are taken from the original papers. Those papers used VITON-HD's standard test split, which is not the same set after duplicate and test-leak removal. FID, FDCLIP, and KID are distribution-level metrics, so MGT's favorable FID (21.9 vs. 25.1) and FDCLIP (7.0 vs. 9.4) could be artifacts of test-set composition rather than model quality. The table's own numbers also contradict the abstract's 'state-of-the-art VTOFF results on VITON-HD': MGT is worse than TryOffDiff on DISTS (24.7 vs. 23.0), LPIPS (36.3 vs. 32.4), SSIM (78.1 vs. 79.5), and no better on KID (8.9 vs. 8.9), and worse than TryOffAnyone on LPIPS (36.3 vs. 17.2) and KID (8.9 vs. 2.0). The cross-domain generalization claim therefore lacks a valid same-split comparison. The reader's SigLIP-conditioning concern is less load-bearing here because the per-category TryOffDiff baselines in Table 1 use the same SigLIP adapter, so any conditioning insufficiency affects both arms equally; the uncontrolled VITON-HD comparison is the fragile premise for the headline generalization and state-of-the-art claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Multi-Garment TryOffDiff (MGT), a latent diffusion model for virtual try-off that reconstructs catalog-style garment images from photos of clothed people. The model extends the authors' prior TryOffDiff by replacing text conditioning with SigLIP image features processed by a lightweight adapter and by adding learned class embeddings for upper-body, lower-body, and dresses, combined with the timestep embedding via element-wise addition (Eq. 5). Experiments compare MGT with category-specific TryOffDiff variants on DressCode (Table 1), with published baselines on VITON-HD (Table 2), and as a garment estimator feeding OOTDiffusion for person-to-person virtual try-on (Table 3). The central claims are that a single unified model achieves performance comparable to per-category models on DressCode, that it generalizes across domains to VITON-HD, and that it supports practical p2p-VTON.","tokens_in":11338,"tokens_out":5439,"duration_ms":46201,"significance":"If the DressCode result holds, the paper fills a clear gap: existing open VTOFF models handle only upper-body garments, whereas MGT is the first unified multi-category model. The Table 1 design—same training split, same architecture, only the unified-vs-specialized factor varied—is the right way to measure the cost of unification, and the release of demo, code, and models is a reproducible contribution. The limitations section honestly acknowledges that texture and logo fidelity remain challenging and that only three categories are covered. However, the cross-dataset and state-of-the-art claims are currently supported by an uncontrolled test-set comparison, and the main 'comparable performance' conclusion lacks variance or significance information; the overall contribution is useful but the evidence is partially overstated.","major_comments":[{"comment":"MGT is evaluated on a deduplicated VITON-HD test set of 1,990 pairs, while the baseline numbers in Table 2 are taken from the original papers, which used the standard VITON-HD test split. Because FID, KID, and FDCLIP are distribution-level metrics, the favorable FID (21.9 vs. 25.1) and FDCLIP (7.0 vs. 9.4) results can be artifacts of test-set composition rather than model quality. This comparison does not support the cross-dataset generalization claim as stated.","section":"Section 4.1, Table 2"},{"comment":"The abstract claims 'state-of-the-art VTOFF results on VITON-HD,' but the same table shows MGT is worse than TryOffDiff on DISTS (24.7 vs. 23.0), LPIPS (36.3 vs. 32.4), SSIM (78.1 vs. 79.5), and equal on KID (8.9 vs. 8.9). The claim should be revised to 'competitive' or substantiated by a same-split re-evaluation of the baselines.","section":"Table 2, Abstract"},{"comment":"Inference hyperparameters (guidance scale and number of steps) are analyzed and selected by measuring FID and DISTS on DressCode-test, the same test set from which all reported quantitative results are computed. This is a test-set selection procedure; a validation split or nested protocol is needed before the reported numbers can be taken as unbiased estimates of MGT's performance.","section":"Section 4.2, Fig. 4"},{"comment":"The paper does not ablate the class-embedding mechanism that is the paper's main technical novelty. Without a multi-category baseline trained without the class conditioning (or an alternative conditioning), it is not established that the element-wise addition in Eq. (5), rather than simply training one model on all categories, is responsible for the multi-garment behavior.","section":"Section 3.2, Eq. (5)"},{"comment":"No standard errors, confidence intervals, or significance tests are reported, and the differences are small (e.g., DISTS 22.2 vs. 21.6 for upper body, SSIM 80.2 vs. 80.8). The claim of 'comparable performance across all categories' needs variance information across seeds or bootstrapping to be established.","section":"Table 1"}],"minor_comments":[{"comment":"The limitations paragraph concedes that fine-grained texture recovery and logo preservation remain challenging and that only three garment categories are supported; this qualification should be reflected in the abstract's strong reconstruction language.","section":"Section 5, Limitations"},{"comment":"Several metric entries for TryOffAnyone are missing (MS-SSIM, CW-SSIM, FDCLIP); please state whether these were not reported in the original paper or are not applicable.","section":"Table 2"},{"comment":"The adapter description says 1,024 SigLIP tokens are reduced to n = 77 conditioning embeddings; please clarify how the reduction is performed (learned projection, pooling, or truncation) and why 77 tokens were chosen.","section":"Section 4.1"},{"comment":"For p2p-VTON, the authors say ground-truth are unavailable and they randomly pair garments across individuals; the exact random pairing and whether the same pairs were used for all methods should be specified for reproducibility.","section":"Section 4.1"},{"comment":"The notation in Eq. (3) uses c both as the index set variable and in E_c; this is harmless but should be cleaned for clarity.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an honest extension of the authors' own TryOffDiff, but the VITON-HD comparison in Table 2 is not apples-to-apples and the state-of-the-art wording in the abstract is not supported by the paper's own numbers. I would advise the editor to require a same-split comparison or a downgraded claim, and to ask for the class-embedding ablation and variance estimates. The DressCode multi-garment finding is plausible and the release of code helps, so I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: MGT is a real, if modest, contribution. Adding learnable class embeddings to the timestep embedding of the authors' own TryOffDiff is a simple trick, but it works for the stated purpose. Table 1 is a fair same-split comparison against per-category TryOffDiff variants on DressCode, and it shows that one unified model can handle upper-body, lower-body, and dresses with only minor trade-offs. That is the paper's actual result, and it holds. The code, models, and demo being public is a real plus, and the p2p-VTON integration with OOTDiffusion is a useful practical demonstration even if it is not the main event.\n\nThe soft spot is exactly where the stress-test lands, and I think the stress-test is right. The VITON-HD \"state-of-the-art\" claim in the abstract and conclusion is unsupported. Table 2 compares MGT's numbers on a deduplicated 1,990-pair test set against baseline numbers taken from original papers that used the standard VITON-HD split. FID, FDCLIP, and KID are distribution-level metrics; you cannot compare them across different test sets. And even taking the table at face value, MGT is worse than TryOffDiff on DISTS, LPIPS, and SSIM, and no better on KID. So the cross-dataset generalization claim is weaker than the prose suggests. That is a fixable problem—report the baselines on the same deduplicated split, or drop the claim—but as written it is an overstatement.\n\nThe other issues are minor by comparison. Figure 4 looks like inference hyperparameters were tuned on the test set; that should be a validation split or clearly labeled as sensitivity analysis. There are no error bars or significance tests, which is common in this subfield but worth noting. The reader's worry about SigLIP conditioning being a fragile load-bearing assumption is, I think, less convincing: the per-category baselines in Table 1 use the same SigLIP adapter, so any conditioning insufficiency hits both arms equally. The real fragile premise is the uncontrolled VITON-HD comparison.\n\nBottom line: this paper deserves a serious referee. The core multi-garment result is genuine and the evaluation on DressCode supports it. A major revision that fixes the VITON-HD comparison and tempers the claims would make it solid. I would bring it to a reading group if people care about fashion-image synthesis, and I would cite it if I were working on VTOFF.\n\nRecommendation: accept for peer review, with the expectation of revision.","headline":"A modest, legitimate multi-garment VTOFF extension with solid DressCode evidence, but the VITON-HD state-of-the-art claim does not survive contact with the same-split comparison.","tokens_in":11847,"tokens_out":1540,"would_cite":true,"duration_ms":14649,"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":"MGT claims that a single diffusion model with class-specific embeddings can reconstruct upper-body, lower-body, and dress garments at quality comparable to per-category models, while generalizing across datasets.","keywords":["virtual try-off","multi-garment reconstruction","latent diffusion","SigLIP conditioning","class-specific embeddings","person-to-person virtual try-on","DressCode","VITON-HD"],"falsifier":"Take the DressCode test set, split it into plain garments and garments with visible logos or complex patterns, and compare MGT's DISTS scores on each subset against a category-specific TryOffDiff model; if MGT degrades substantially more on the logo/pattern subset, the SigLIP conditioning is the bottleneck. A second decisive test would be freezing the class embedding and varying only the reference image across categories: if outputs remain correct without the class label, the class embedding is not doing the claimed work; if outputs become ambiguous without it, the embedding is necessary.","tokens_in":10775,"feed_emoji":"👗","tokens_out":4279,"duration_ms":36055,"temperature":0.7,"pith_summary":"This paper introduces Multi-Garment TryOffDiff (MGT), a diffusion model that turns a photo of a clothed person into a clean, catalog-style image of a single garment. The central claim is that one unified model can reconstruct upper-body garments, lower-body garments, and dresses with quality close to that of models trained separately on each category. The authors also claim that MGT, trained only on the DressCode dataset, produces competitive results on the VITON-HD dataset without retraining. Achieving this would matter because virtual try-off currently requires a separate model per garment type, and because it enables person-to-person virtual try-on by supplying standardized garments extracted from ordinary photos.","feed_headline":"One model reconstructs tops, pants, and dresses from photos","feed_subtitle":"Trained only on DressCode, it matches per-category models and transfers to VITON-HD without retraining.","key_machinery":"The load-bearing mechanism is the combination of SigLIP token conditioning with class-conditioned timestep embeddings. SigLIP features from the reference photo are projected by a lightweight adapter, computing $C(I) = (\\mathrm{LN} \\circ \\mathrm{Linear} \\circ \\mathrm{SigLIP})(I)$, and injected into the U-Net's cross-attention layers as keys and values, following the IP-Adapter design. A 1,280-dimensional learnable embedding for the garment class is added element-wise to the timestep embedding ($e_t + e_c$) before it modulates each residual block, so the same network can steer generation toward the requested garment type. The SigLIP encoder and VAE are frozen; only the adapter, the class embeddings, and the U-Net are finetuned.","core_discovery":"MGT is claimed to be the first unified model for multi-garment virtual try-off. It replaces text conditioning in a Stable Diffusion v1.4 latent diffusion backbone with SigLIP image features injected into cross-attention, and it adds a learnable embedding per garment class ('upper body', 'lower body', 'dress') to the timestep embedding in every residual block. On DressCode, MGT's scores are close to those of category-specific TryOffDiff models across all three categories (e.g., DISTS 22.2 vs 21.6 for upper body), and on VITON-HD it achieves better FID (21.9 vs 25.1) than a baseline trained on that dataset. The paper therefore positions MGT as a single model that replaces several specialized ones without a significant quality loss.","pith_inferences":["If SigLIP features are the real bottleneck, then swapping in a higher-resolution or spatially denser encoder (e.g., patch-level features rather than a 77-token sequence) should improve logo and fine-texture fidelity without changing the architecture's class-conditioning mechanism.","The element-wise addition of class embeddings to timestep embeddings is a generic conditioning channel; the same mechanism could encode color, fabric, or silhouette attributes, turning VTOFF into an attribute-controllable generator.","The success on VITON-HD suggests that paired catalog data may be less necessary than assumed: a model trained on one domain could generate training pairs for another, potentially bootstrapping VTOFF in domains where product shots are rare.","Layered clothing is the natural stress test: the paper explicitly excludes it, and the class labels cannot specify which layer to extract, so a two-garment reference image with a jacket over a shirt would likely require explicit layering or ordering cues."],"forward_implications":["One unified model can replace three per-category VTOFF models, lowering training and deployment cost for e-commerce catalog generation.","VTOFF models trained on one dataset can be applied to another dataset without finetuning, as shown by MGT's transfer from DressCode to VITON-HD.","Pairing MGT with a VTON model yields person-to-person try-on that avoids direct texture and skin-tone leakage from the source person, because garment extraction and person rendering are separated.","The class-conditioning trick is a small change to the timestep embedding and can likely be extended to new garment categories or attributes as annotated data becomes available.","MGT can turn user-generated photos into standardized product images, making catalog creation faster than studio photography."],"supporting_citations":[{"why":"TryOffDiff provides the base VTOFF architecture, task formulation, and the per-category baseline that MGT must match.","marker":"[32]"},{"why":"SigLIP is the frozen image encoder whose token features supply the garment shape, texture, and pattern conditioning.","marker":"[42]"},{"why":"IP-Adapter supplies the design for projecting image features into cross-attention keys and values.","marker":"[40]"},{"why":"Stable Diffusion v1.4 provides the pretrained latent diffusion backbone whose U-Net and VAE are reused.","marker":"[27]"},{"why":"DressCode is the training dataset with three garment categories and paired person-garment images.","marker":"[22]"},{"why":"VITON-HD is the cross-dataset evaluation target with upper-body pairs.","marker":"[8]"},{"why":"OOTDiffusion is the VTON model used in the p2p-VTON integration experiments.","marker":"[39]"},{"why":"CatVTON is the dedicated p2p-VTON baseline compared in the integration study.","marker":"[10]"}],"fun_headline_variants":["One model handles tops, pants, and dresses for try-off","Unified try-off model for multiple garment types","Multi-garment try-off in one diffusion model","One model, three garment categories: try-off unified","First unified virtual try-off for tops, bottoms, dresses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that a frozen SigLIP encoder, through a single linear adapter, captures enough garment shape, texture, and pattern detail that the model can reconstruct high-fidelity catalog images without warping, segmentation, pose, or text guidance.","fun_headline_variants_meta":{"raw":{"variants":["One model handles tops, pants, and dresses for try-off","Unified try-off model for multiple garment types","Multi-garment try-off in one diffusion model","One model, three garment categories: try-off unified","First unified virtual try-off for tops, bottoms, dresses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00061,"raw_usage":{"total_tokens":2836,"prompt_tokens":936,"completion_tokens":1900,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1822}},"tokens_in":552,"tokens_out":1900,"duration_ms":12936,"temperature":1.0,"reasoning_tokens":1822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:14:59.810369+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the DressCode test set, split it into plain garments and garments with visible logos or complex patterns, and compare MGT's DISTS scores on each subset against a category-specific TryOffDiff model; if MGT degrades substantially more on the logo/pattern subset, the SigLIP conditioning is the bottleneck. A second decisive test would be freezing the class embedding and varying only the reference image across categories: if outputs remain correct without the class label, the class embedding is not doing the claimed work; if outputs become ambiguous without it, the embedding is necessary.","supporting_citations":[{"cited_title":"Tryoffdiff: Virtual-try-off via high-fidelity gar- ment reconstruction using diffusion models","cited_arxiv_id":null,"evidence_quote":"TryOffDiff provides the base VTOFF architecture, task formulation, and the per-category baseline that MGT must match."},{"cited_title":"Sigmoid loss for language image pre-training","cited_arxiv_id":null,"evidence_quote":"SigLIP is the frozen image encoder whose token features supply the garment shape, texture, and pattern conditioning."},{"cited_title":"Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models","cited_arxiv_id":null,"evidence_quote":"IP-Adapter supplies the design for projecting image features into cross-attention keys and values."},{"cited_title":"High-resolution image syn- thesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Stable Diffusion v1.4 provides the pretrained latent diffusion backbone whose U-Net and VAE are reused."},{"cited_title":"Dress code: High- resolution multi-category virtual try-on","cited_arxiv_id":null,"evidence_quote":"DressCode is the training dataset with three garment categories and paired person-garment images."},{"cited_title":"Viton-hd: High-resolution virtual try-on via misalignment-aware normalization","cited_arxiv_id":null,"evidence_quote":"VITON-HD is the cross-dataset evaluation target with upper-body pairs."},{"cited_title":"Towards scalable unpaired virtual try-on via patch-routed spatially- adaptive gan","cited_arxiv_id":null,"evidence_quote":"OOTDiffusion is the VTON model used in the p2p-VTON integration experiments."},{"cited_title":"Catvton: Concatenation is all you need for virtual try-on with diffusion models","cited_arxiv_id":null,"evidence_quote":"CatVTON is the dedicated p2p-VTON baseline compared in the integration study."}],"review_version":1}