{"id":"94617758-f644-4c3a-b73b-7e5a1153bafa","arxiv_id":"2502.01626","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A mask-free person-to-person virtual try-on model built on FLUX-Fill-dev, trained with pseudo data generated by IDM and a Focus Attention loss.","lead":"This paper proposes MFP-VTON, a mask-free system that transfers a garment from one person's photo onto another person's photo using a fine-tuned diffusion transformer. It is designed for easier virtual try-on, where getting a clean product garment image is often harder than using a photo of someone already wearing it.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The person-to-person evaluation is circular: IDM generates both the pseudo-training targets and the paired test ground truth (Sec. 3.1, 4.2), so Table 1 primarily measures agreement with IDM artifacts rather than real try-on fidelity.","rationale":"Reader's weakest assumption is exactly the load-bearing point. The central claim has two halves: person-to-person and garment-to-person. The garment-to-person half (Table 2) uses VITON-HD real pairs and is not circular, although gains are small and unaccompanied by error bars. The person-to-person half is circular because the pseudo-ground-truth is generated by IDM, the same model family that MFP-VTON is trained to imitate and that also appears as an adapted baseline. Thus paired fidelity metrics are contaminated by train/test generator identity. This is a correctness risk, not a mismatch with consensus. The Focus Attention loss inconsistency (Eq. 1) is real and independently verifiable from the text, but I do not make it the primary attack because even a mis-specified auxiliary loss could leave the model's generative quality largely driven by FLUX-Fill; the empirical evidence is what carries the claim. A cross-generator evaluation or human preference study would settle the issue, and neither is present. The REJECT verdict is therefore justified.","tokens_in":6837,"tokens_out":5469,"duration_ms":50659,"concrete_test":"Rerun the paired person-to-person evaluation in Table 1 with ground truths regenerated by a different garment-to-person model (e.g., CatVTON-FLUX) on the same swap pairs, keeping MFP-VTON and all reported baselines fixed. If MFP-VTON's LPIPS/DISTS/FID advantages shrink, reverse, or are overtaken by IDM/CatVTON variants, the original table measured agreement with IDM's synthetic distribution rather than true fitting fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim rests on Table 1 'paired person-to-person' metrics, where ground-truth images are IDM outputs produced by the same Sec. 3.1 pipeline used to create the training triplets. Since MFP-VTON is fine-tuned to reproduce IDM's synthetic swaps, high SSIM/LPIPS/DISTS on these targets can be achieved by learning IDM's characteristic artifacts (garment warping, texture smoothing, identity drift). The adapted baselines (Seg/TROF + IDM) are not trained on pseudo-data and enter the comparison at an asymmetry; their lower paired scores may simply reflect that they are not optimized against the same generator. No human study or independent real paired P2P dataset is provided, so the 'excels in person-to-person' claim is not established. The Focus Attention loss (Eq. 1) is also internally inconsistent: the first term weights non-garment reference regions and the second weights garment regions of the target, opposite to the stated objective, but this mechanism defect is secondary to the circular evaluation for the performance claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MFP-VTON, a mask-free person-to-person virtual try-on framework built on the FLUX.1-Fill-dev diffusion transformer. Because no large person-to-person paired dataset exists, the authors use the off-the-shelf garment-to-person model IDM to swap garments between pairs from VITON-HD, creating pseudo triplets of the form (reference person, target person, target person wearing the reference garment). The model concatenates the reference person and target person images with a blank inpainting region, and is fine-tuned with a proposed Focus Attention loss. Experiments compare the method against garment-to-person baselines OOTD, IDM, and CatVTON-FLUX after adapting them with segmentation or try-off methods, reporting both paired and unpaired metrics on the person-to-person task as well as garment-to-person metrics.","tokens_in":7006,"tokens_out":9239,"duration_ms":85893,"significance":"The ease-of-use motivation is clear: a mask-free framework that can transfer a garment directly from a worn reference image would remove the need for standard garment product images and garment masks. The recipe is simple and reproducible in spirit, relying on public models (FLUX.1-Fill-dev, IDM) and the standard VITON-HD dataset. The paper also reports a standard set of quantitative metrics, which is a useful starting point. However, the empirical evidence for the central claim is not convincing: the paired person-to-person evaluation is circular, and the Focus Attention loss as written contradicts its stated objective. These issues prevent the current version from establishing the claimed performance.","major_comments":[{"comment":"The paired person-to-person evaluation is not independent of the training signal. The ground-truth images in Table 1 are generated by IDM using the same procedure described in Sec. 3.1 that created the pseudo training triplets. Since MFP-VTON is fine-tuned to reproduce those IDM-generated swaps, the SSIM, LPIPS, and DISTS columns largely measure agreement with IDM's characteristic artifacts (garment warping, texture smoothing, identity drift) rather than real try-on fidelity. The adapted baselines are not trained against this same generator, so the comparison is asymmetric. No human study or independent real paired person-to-person dataset is provided. The paired columns therefore cannot support the abstract's claim that the method 'excels in person-to-person VTON'; at best, they show that the model reproduces IDM-style outputs. The unpaired FID/CLIP-FID/KID columns remain distributional evidence, but they do not directly measure garment or identity preservation.","section":"Sec. 3.3, Eq. (1)"},{"comment":"The Focus Attention loss is written in the opposite direction of the text. The text says the loss should emphasize the garment of the reference person and the details outside the garment of the target person. However, Eq. (1) weights the attention from the fitting output to the reference person by (1 - M_r), i.e., the non-garment part of the reference image, and the attention to the target person by M_t, i.e., the garment part of the target image. This is exactly the reverse of the stated objective. Because the loss is a claimed contribution and is used during training, this inconsistency must be fixed and the correct form should be verified with an ablation. The equation also contains an unbalanced parenthesis, further obscuring the intended expression.","section":"Sec. 3.3, Eq. (1)"},{"comment":"The paper's wording that the method 'outperforms existing methods across multiple metrics' is stronger than the table shows. In the paired person-to-person columns, 'Ours' has SSIM 0.8688, which is lower than Seg+IDM (0.8727) and TROF+IDM (0.8761). The text should state the exact set of metrics on which improvements are claimed, rather than implying a uniform improvement.","section":"Sec. 4.2, Table 1 (Ours row)"}],"minor_comments":[{"comment":"The caption of Table 2 says 'person-to-person task', but the table reports garment-to-person results; this should be corrected.","section":"Table 2"},{"comment":"The authors state that they 'manually filter out a subset' of the pseudo person-to-person data, but no filtering criteria, statistics, or examples are given. This makes the dataset construction non-reproducible.","section":"Sec. 4.1"},{"comment":"The 'unpaired testing datasets' used for FID/CLIP-FID/KID are not described precisely; the paper should specify which real image distribution is used as the reference for computing these metrics.","section":"Sec. 4.2"},{"comment":"Aside from the mask-direction issue, Eq. (1) has a formatting and parenthesis problem: the first 'mean' is not closed before the second term, and the overall expression is ambiguous.","section":"Sec. 3.3, Eq. (1)"},{"comment":"The authors say the model is fine-tuned on 'both garment-to-person and person-to-person datasets', but Sec. 3.1 only describes construction of the person-to-person pseudo dataset; the garment-to-person fine-tuning data and any associated loss are not described.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The circular evaluation and the inconsistent Focus Attention loss are the two load-bearing problems. If a revision is submitted, it should include an independent evaluation (e.g., a human preference study or a small real paired person-to-person test set) and a corrected, ablated loss. Without those, the central empirical claim is not established, and the paper would be difficult to accept in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible incremental engineering paper, but the evidence for the headline claim is not there yet. The mask-free concatenation idea is sensible, and the authors identify a real usability gap. The problem is that the paired test ground truth is produced by the same IDM model that generated the training data. That makes Table 1 a measure of agreement with IDM artifacts, not genuine try-on fidelity. The baseline comparison is also asymmetric: the baselines are not trained on IDM pseudo-data. And Eq. (1) as written applies the masks in exactly the opposite direction from the stated objective.\n\nWhat's genuinely new: using FLUX-Fill-dev for mask-free person-to-person try-on by concatenating reference and target person with a blank inpainting region is a clean, simple idea. The custom data generation follows Boow-VTON, but adapting it for this task is appropriate. The paper also evaluates with reasonable adapted baselines (segmentation and try-off). The qualitative examples look plausible.\n\nSoft spots: the circular evaluation is the main issue. VITON-HD is split, IDM generates pseudo-triplets for training, and then the test set is run through the same IDM pipeline to create ground truths. The model is fine-tuned to imitate IDM, so high LPIPS/DISTS scores may just reflect learning IDM's warping and smoothing. No human study or real paired person-to-person set. Also the loss equation contradiction: text says focus on reference garment and target non-garment, but Eq. (1) weights reference non-garment and target garment. That's a concrete error that needs fixing or clarification. Minor: no error bars; in Table 1, TROF+IDM actually beats Ours on SSIM, so 'excels' is overstated; Table 2 shows Ours is slightly behind IDM/CatVTON on most paired metrics; and the end-to-end person-to-person baselines (StreetTryOn, PASTA-GAN++) are cited but not compared.\n\nWho this is for: people working on practical VTON pipelines might find the mask-free setup useful. It deserves a serious referee if the authors can fix the evaluation and clarify the loss, but as submitted the central claim is not established. I would send it to peer review because the idea is worth checking and the flaws are fixable, but the authors should be told to redo the evaluation with human raters or an independent paired dataset and to correct Eq. (1).","headline":"A practical mask-free person-to-person try-on framework built on FLUX-Fill-dev, but its headline performance claim rests on a circular IDM-generated evaluation and its Focus Attention loss equation contradicts its own text.","tokens_in":7560,"tokens_out":3682,"would_cite":false,"duration_ms":32971,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MFP-VTON claims that one mask-free diffusion-transformer model can handle both person-to-person and garment-to-person virtual try-on at high fidelity, without needing a clean garment image or a garment mask at inference.","keywords":["virtual try-on","person-to-person try-on","mask-free VTON","diffusion transformer","Focus Attention loss","FLUX.1-Fill-dev","VITON-HD","synthetic training data"],"falsifier":"Collect genuine person-to-person swap images where the same garment is worn by two people in matched poses, then compare MFP-VTON against baselines with human preference judgments and pixel-level fidelity metrics; if human raters prefer the baselines or the generated images reproduce artifacts that appear in the teacher model's synthetic swaps, the central claim of high-fidelity person-to-person try-on would be weakened.","tokens_in":6612,"feed_emoji":"👕","tokens_out":5760,"duration_ms":46555,"temperature":0.7,"pith_summary":"Virtual try-on usually requires a clean product photo of a garment, and getting one from a photo of a person wearing it requires extra segmentation or restoration steps. This paper tries to remove that requirement: given a photo of a person wearing a garment and a photo of another person, the model should directly generate the second person wearing the first person's garment. Because no person-to-person paired dataset exists, the authors synthesize one by using a pretrained garment-to-person model to swap garments between pairs from an existing dataset. They then fine-tune a diffusion-transformer inpainting model with a Focus Attention loss that directs attention to the reference garment and to the target person's non-garment details. The reported experiments indicate the same model also works on the original garment-to-person setup, so one checkpoint covers both tasks.","feed_headline":"Model swaps clothes between two photos with no mask","feed_subtitle":"Training pairs come from an off-the-shelf try-on model; attention supervision keeps pose and hand-held details.","key_machinery":"The mechanism is a mask-free conditioning scheme combined with a Focus Attention loss. The condition image is built by concatenating the reference person, the target person, and a blank patch along the width dimension, and the corresponding mask is all ones except for the blank patch, so the diffusion transformer only inpaints the final region. The Focus Attention loss operates on the attention sub-matrices from the fitting output tokens to the reference and target image tokens: it penalizes attention to the non-garment area of the reference person and to the garment area of the target person, thereby steering the model to borrow the garment from the reference and preserve the exterior of the target. This loss is what makes the mask-free setup trainable without explicit spatial guidance at inference.","core_discovery":"The paper's central claim is that person-to-person virtual try-on can be done without a garment mask and without first restoring a clean garment from the reference person. The input to the model is a horizontal concatenation of the reference person, the target person, and a blank region, with a full-mask image so that inpainting happens only in the blank region. Because the target person's garment area is never masked, the model preserves foreground details such as hands and handheld objects. A Focus Attention loss supervises the transformer's cross-attention maps: queries from the fitting region are pushed to attend to garment pixels of the reference person and to non-garment pixels of the target person. The authors report that on paired and unpaired person-to-person benchmarks the method outperforms adapted garment-to-person baselines, while remaining comparable on the standard garment-to-person task.","pith_inferences":["The paired metrics in the paper measure agreement with the synthetic ground truths produced by the off-the-shelf model, so a real-world deployment would likely need validation against genuinely captured try-on photos; any systematic artifacts in the teacher model would be inherited by MFP-VTON.","The Focus Attention loss is a general mechanism that could be transferred to other diffusion inpainting tasks where one region must borrow content from another region while preserving the rest, such as face swapping or object replacement.","Because the current training pairs come from VITON-HD, a dataset with limited pose variation, an extension worth testing is whether the method retains its pose preservation when the reference and target persons have large pose differences."],"forward_implications":["Person-to-person try-on can be run directly from two photos, without first segmenting or restoring a clean garment from the reference photo.","Because no target-person garment mask is needed at inference, the method can preserve pose and foreground objects such as phones better than mask-based inpainting, according to the paper.","The same fine-tuned checkpoint handles both person-to-person and garment-to-person try-on, so a single model can serve both workflows.","Pseudo-pair generation from an off-the-shelf garment-to-person model is a viable route to build training data for tasks that lack paired person-to-person examples."],"supporting_citations":[{"why":"FLUX.1-Fill-dev is the diffusion-transformer foundation model that the method fine-tunes for inpainting.","marker":"[2]"},{"why":"VITON-HD supplies the paired person and garment images used to build the pseudo person-to-person dataset and the evaluation set.","marker":"[3]"},{"why":"IDM is the off-the-shelf garment-to-person model used to generate the pseudo ground-truth swaps and also serves as a comparison baseline.","marker":"[4]"},{"why":"In-Context-LoRA provides the width-concatenation conditioning idea for feeding multiple images to the transformer.","marker":"[11]"},{"why":"SAM 2 is used for segmentation to adapt garment-to-person baselines to the person-to-person setting.","marker":"[17]"},{"why":"TryOnDiff is used for try-off garment reconstruction in one of the baseline adaptation pipelines.","marker":"[22]"},{"why":"CatVTON-FLUX is a direct baseline that shares the FLUX-Fill foundation and is compared on both tasks.","marker":"[25]"},{"why":"BOOW-VTON provides the mask-free pseudo-data training methodology that the data preparation follows.","marker":"[31]"}],"fun_headline_variants":["Clothes swap between people, no mask needed","Try-on any outfit from a photo, no mask","MFP-VTON: mask-free person-to-person try-on","Swap outfits in photos without masks","Person-to-person VTON, no mask, high fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The person-to-person ground-truth images used for both training and evaluation are generated by another virtual try-on model, so the whole pipeline assumes those synthetic swaps are accurate enough to serve as correct targets.","fun_headline_variants_meta":{"raw":{"variants":["Clothes swap between people, no mask needed","Try-on any outfit from a photo, no mask","MFP-VTON: mask-free person-to-person try-on","Swap outfits in photos without masks","Person-to-person VTON, no mask, high fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1592,"prompt_tokens":866,"completion_tokens":726,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":652}},"tokens_in":482,"tokens_out":726,"duration_ms":7391,"temperature":1.0,"reasoning_tokens":652,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:46:33.686506+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect genuine person-to-person swap images where the same garment is worn by two people in matched poses, then compare MFP-VTON against baselines with human preference judgments and pixel-level fidelity metrics; if human raters prefer the baselines or the generated images reproduce artifacts that appear in the teacher model's synthetic swaps, the central claim of high-fidelity person-to-person try-on would be weakened.","supporting_citations":[{"cited_title":"Flux.1-fill-dev","cited_arxiv_id":null,"evidence_quote":"FLUX.1-Fill-dev is the diffusion-transformer foundation model that the method fine-tunes for inpainting."},{"cited_title":"Viton-hd: High-resolution virtual try-on via misalignment-aware normalization","cited_arxiv_id":null,"evidence_quote":"VITON-HD supplies the paired person and garment images used to build the pseudo person-to-person dataset and the evaluation set."},{"cited_title":"Catvton-flux","cited_arxiv_id":null,"evidence_quote":"CatVTON-FLUX is a direct baseline that shares the FLUX-Fill foundation and is compared on both tasks."},{"cited_title":"BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training","cited_arxiv_id":"2408.06047","evidence_quote":"BOOW-VTON provides the mask-free pseudo-data training methodology that the data preparation follows."}],"review_version":1}