{"id":"35073da1-261a-48b7-bad3-b244c743a651","arxiv_id":"2608.05834","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A virtual try-on method conditions a latent diffusion model on pixel-space garment measurements and custom adapters, but validation is only qualitative.","lead":"This technical report describes a virtual try-on system that adds fine-grained labels, such as sleeve length and neckline depth, to garment images so a diffusion model can generate edited photos with user-controlled fit. The work is a student report with qualitative examples, no code, and no quantitative evaluation, so its practical value is not yet established.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Control claim rests on a few unseeded images; Section 8 concedes the one-hot labels may be redundant with text prompts, so the label vector's causal role is unestablished.","rationale":"The reader's weakest assumption identifies redundancy of the one-hot labels as the key risk. I agree that this is a serious risk, and Section 8 itself flags it. However, the paper's control evidence also includes continuous pixel-space measurements (shoulder fraction, hip distance, neckline depth/width) that were not passed as text prompts; Fig. 19 and Fig. 20 are the most direct demonstrations of control. My concern is therefore slightly broader: the entire conditioning vector, not just the one-hot part, lacks controlled quantitative verification. No seeds, repetitions, baselines, or error bars are provided, so the central claim of user-controllable generation is not yet established. This supports the reader's CONDITIONAL verdict rather than changing it, because the mechanism is plausible and some visual evidence is consistent with it. In-scope self-reported limitations in Section 8—the unclear effect of one-hot encoding and the redundancy with text prompts—were weighed explicitly. The proposed test would settle whether label changes causally shift the output distribution beyond stochastic noise and whether one-hot labels add anything over text prompting.","tokens_in":12170,"tokens_out":6331,"duration_ms":68645,"concrete_test":"Fix one garment/person pair and a set of seeds. Generate 20 images for each of: (A) label set A, (B) label set B differing only in shoulder fraction and neckline depth, and (C) label set A repeated with different seeds. Compute pairwise LPIPS distances within and across conditions. If mean across-condition distance is not significantly larger than within-condition distance, seed noise dominates and the claimed control is not demonstrated. Also generate label set A with and without its corresponding text prompt to quantify one-hot redundancy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the trained adapters let a user control fit and style via the label vector—requires that varying the conditioning changes the output distribution in a systematic, label-specific way. The report's evidence is a small set of qualitative images (Figs. 14–20). No inference seeds are reported, no repeated generations per condition are shown, and no quantitative metric (LPIPS, FID, measured garment-length error, etc.) separates label-driven variation from sampling noise. Section 8 explicitly states: 'the effect of the one-hot encoding is unclear' and that garment type and sleeve length were also passed as text prompts during training; since inference in Figs. 14–15 is prompt-free, the differences could reflect prompt-free behavior or seed variation rather than the one-hot vector. For the continuous measurements (shoulder fraction, hip distance, neckline), Figs. 19–20 show single suggestive examples; without seed control or repetition, cherry-picking cannot be excluded. Because the conditioning signal's non-redundant influence is precisely the mechanism asserted in the abstract, the absence of controlled, quantitative evidence is the load-bearing weakness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ControllableClothing, a method for adding fine-grained control to latent-diffusion-based virtual try-on. Garment images are pre-processed with segmentation (GroundingDINO, SAM, OpenPose) to extract continuous measurements (shoulder coverage fraction, hip distance, neckline depth/width) and one-hot class labels (sleeve length, fit, style). These labels are concatenated with an IP-adapter image embedding and passed through a custom trainable MLP; additionally, a custom up-block T2I adapter is introduced. Training on DressCode and VitonHD, the author claims the model can steer fit and style at inference. Evaluation is qualitative: a handful of generated images showing changed conditions, plus a one-image ablation. The author openly acknowledges in Section 8 that the one-hot encoding effect is unclear and that several labels were also passed as text prompts during training.","tokens_in":12340,"tokens_out":2062,"duration_ms":23171,"significance":"If the central claim holds, the method would address a real limitation of VITON systems, which typically produce a single average fit with no user control over sleeve length, neckline depth, or shoulder coverage. The proposed label-extraction pipeline in pixel space is a plausible and reproducible alternative to latent-space captioning, and the custom up-block T2I adapter is a sensible architectural contribution. The author's decision to train on open datasets with open tools and to document limitations honestly (Section 8) counts in the paper's favor. However, the reported evidence does not currently establish the causal control claim, because the evaluation lacks quantitative metrics, seed control, repeated generations, and a held-out split. The paper is therefore best seen as an early technical report with a promising direction rather than a validated method.","major_comments":[{"comment":"The central claim that user-controllable labels steer generation is supported only by a small set of curated single images. No inference seeds are reported, no repeated generations per condition are shown, and no quantitative metric (e.g., LPIPS, FID, measured garment-length error, or fraction of correct semantic labels) distinguishes label-driven variation from sampling noise. Because Figures 14–15 are generated without text prompts, the differences could reflect prompt-free behavior or seed variation rather than the one-hot vector. I request a controlled evaluation: fix the seed, mask, prompt, and all other inputs, vary one label at a time, and report distributions of a relevant metric over multiple seeds, ideally on a held-out split of DressCode/VitonHD rather than on the training set.","section":"Section 6, Figures 14–20"},{"comment":"The manuscript itself states, 'the effect of the one-hot encoding is unclear' and that 'the type of garment ... was passed as a prompt along with five other random labels' and that sleeve length information was 'added as a prompt.' Since the text prompt already conveyed these attributes, the label vector's non-redundant influence is exactly the load-bearing mechanism asserted in the abstract. The ablation study in Section 7 removes entire modules but does not test whether the label input alone changes the output when the prompt is held fixed. To support the control claim, the author needs an experiment where the label vector is varied while the text prompt (if any) and all other conditioning are held constant; without such a test, the method may reduce to prompt conditioning plus sampling noise.","section":"Section 8, Discussion of one-hot encoding"},{"comment":"The shoulder-fraction demonstration shows a single pair of images with left-frac and right-frac set to different values. This is a suggestive example, but a single unseeded pair cannot exclude cherry-picking, especially because the paper reports that manual changes can have 'unforeseen consequences' (Section 6.1, Figure 14). For a claim that 'setting right-frac to 1 and left to 0' yields asymmetric shoulder coverage, I ask for several seeds per condition and a quantitative measure of shoulder coverage in the output, computed with the same segmentation pipeline used for labels.","section":"Section 6.3, Figure 19"},{"comment":"The waist-mask control experiment also relies on two examples per condition without seeds or a numeric measure of the resulting garment length. Since the comparison is between two generated images with 'the exact same mask' but different waist-mask values, the relevant question is whether this parameter changes the output distribution systematically. Please provide repeated generations and a metric such as the distance from the hem to the mask boundary.","section":"Section 6.4, Figure 20"}],"minor_comments":[{"comment":"The text references 'Figure 24' in the discussion of Figure 13, and the figure numbering in the appendix appears inconsistent (e.g., 'Figure 24' is described as a 'relatively good example' while the appendix orders figures differently); please renumber and cross-check all figure references.","section":"Section 6.1"},{"comment":"Table 1 lists questions and answers from GPT-4o, but the reader cannot tell which of these were actually used as labels versus which were rejected; Section 8 mentions that some question types were rejected. Please clarify the final label set and its encoding in a single table.","section":"Section 4.4"},{"comment":"The dimensions of the up-block T2I intermediate tensors are listed, but the paper does not give the number of training steps, learning rate, batch size, or the exact random chance for adding extra waist mask. Adding these hyperparameters would improve reproducibility.","section":"Section 5.2"},{"comment":"There are several typos and grammatical issues (e.g., 'bu' for 'but', 'the model with not learn', 'looses' for 'loses', 'unpainted' for 'unmasked', 'nlp' for 'MLP' in Section 8). A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This manuscript reads as a candid technical report rather than a fully validated research paper. Its main weakness—qualitative evaluation on curated examples without seed control or held-out data—is fixable within the scope of a revision, but only if the author adds controlled experiments and quantitative metrics. The author's explicit admission in Section 8 that the one-hot encoding's effect is unclear is particularly damaging to the abstract's control claim, so the revision must directly address the redundancy of the label vector with text prompts. I would not recommend rejection because the preprocessing pipeline and the up-block T2I adapter are novel and potentially useful, and the author is transparent about limitations. However, the bar for acceptance should be a demonstration that the label vector has measurable, label-specific influence on the output distribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the arXiv report on Controllable Clothing with interest. The headline is: the central idea is plausible, but as it stands the paper is a candid student report, not a validated method. The abstract overclaims—the evidence for controllable generation is a handful of curated images with no seed control, no repetitions, and no quantitative evaluation.\n\nWhat's actually new is the specific combination: conditioning a VITON pipeline on pixel-space garment measurements (shoulder fraction, neckline depth, hip distance) computed from segmentation keypoints, dynamic mask generation in the dataloader, and a custom T2I adapter for the up-blocks. The author also deserves credit for trying GPT-4 labeling, finding it unreliable, and switching to explicit pixel measurements. The ablation, though single-image, is a reasonable first check, and Section 8 is refreshingly honest about the one-hot encoding's unclear effect and its likely redundancy with the text prompts.\n\nThe soft spots are exactly where the reader's report puts them. The load-bearing claim is that the label vector causally steers the output, but the evidence doesn't separate label-driven variation from sampling noise. No seeds, no repeated generations per condition, no metric like LPIPS or FID, no measured garment-length error. Figures 19–20 are single suggestive examples. The one-hot labels in Figures 14–15 changed the image, but the garment also turned purple—an unforeseen side effect that suggests the conditioning is not yet well-understood. Section 8 concedes the one-hot encoding may have contributed little because the same information was passed as text. And the continuous labels, while more promising, are not rigorously evaluated either. The ablation is one image, so it's a demonstration, not a result.\n\nSo my take: the idea is worth pursuing, but the paper does not currently support its abstract. It should not be accepted as a scientific claim without held-out evaluation, seed-fixed repeated generation, and a comparison to at least IDM-VITON. As a technical report it is a useful read for anyone building VITON conditioning, and the labeling pipeline is a good starting point. I would send it to a serious referee rather than desk reject it, but the referee should push hard for controlled quantitative results. A major revision is the right outcome, not acceptance as-is.","headline":"Plausible VITON conditioning idea with honest limitations, but the control claim is unvalidated; worth a referee to demand quantitative evidence.","tokens_in":12859,"tokens_out":3045,"would_cite":false,"duration_ms":29922,"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":"Adding precise fit labels lets a virtual try-on model generate many versions of one garment instead of a single average fit.","keywords":["virtual try-on","latent diffusion models","controllable image generation","IP-adapter","T2I adapter","garment fit labeling","segmentation-based measurement","mask-conditioned generation"],"falsifier":"Hold the garment image, person image, mask, and text prompt fixed, then sweep one continuous label such as shoulder fraction from zero to one or neckline depth from positive to negative across several random seeds; if the generated images show no systematic change in shoulder coverage or neckline, the label channel is not learned. A second check is to retrain with label values randomly shuffled and compare outputs to the trained model; indistinguishable outputs would mean the label vector is being ignored.","tokens_in":11916,"feed_emoji":"👗","tokens_out":6189,"duration_ms":56085,"temperature":0.7,"pith_summary":"This technical report argues that virtual try-on can be made controllable, not just photorealistic. The author's method augments the training data for a latent diffusion try-on model with labels such as sleeve length, shoulder coverage, neckline depth and width, and garment length relative to the hip, then trains small adapters that take these labels as extra conditioning. At inference the user can adjust the labels to change the fit and styling of the generated garment. The report also shows that continuous pixel-space measurements are more useful than one-hot semantic classes, and that an extra-mask boolean can tell the model to regenerate the waist area. If the method holds up, online retailers could show garments with the true fit rather than an idealized or average fit.","feed_headline":"New labels let virtual try-on users steer fit and style","feed_subtitle":"Adding precise fit labels lets one garment generate many versions, so shoppers see the true fit, not an average one.","key_machinery":"The machinery is the label-injection path around a frozen latent diffusion model. An IP-adapter, which feeds image embeddings into cross-attention, provides the semantic garment embedding, and a trainable MLP/residual module named ControllableClothing adds the concatenated label vector to that embedding. A custom up-block T2I adapter, a light-weight module that does element-wise addition at up-sampling tensors, injects fine detail. The labels themselves come from a pre-processing chain built from open-source segmentation and keypoint models: distances are measured in pixel space from segmentation masks relative to body keypoints such as hips, chest, shoulders, elbow, and wrist, then normalized by body height or shoulder width and standardized across the dataset. Dynamic masks are created in the data loader, and a boolean label tells the model whether an extra mask region below the garment was added.","core_discovery":"The central claim is that pairing garment images with precise fit labels lets a pretrained latent diffusion model generate a diverse, user-steerable set of try-on images, instead of a single average fit. The report constructs a label vector from one-hot classes and continuous floats computed in pixel space—shoulder fraction, hip distance, normalized neckline depth and width, and sleeve-length class—and injects it into an IP-adapter embedding through a small trainable residual module, while a custom up-block T2I adapter supplies detail. Qualitative experiments show that changing the continuous values changes shoulder styling, neckline depth, garment length, and waist regeneration, and that the numeric labels behave better than the one-hot ones. The author is careful to call the evidence preliminary: the up-block adapter was trained for only a fraction of the total epochs, and the unique contribution of the one-hot encoding is unclear because garment type and sleeve length were already supplied as text.","pith_inferences":["A natural ablation the paper did not report would train with the label vector present versus the same labels removed or shuffled while keeping text prompts fixed; near-identical outputs would mean the claimed control mechanism reduces to the text prompt.","Because the current labels are global scalars, the architecture cannot pin a boundary to an exact pixel row; combining label injection with spatial conditioning such as a hemline mask is a likely next step for precise geometric placement.","The segmentation-based measurement pipeline is itself a load-bearing contribution, so the method's reliability is tied to segmentation quality; an extension would test robustness by adding controlled noise to shoulder and neckline masks during training.","If the waist-mask boolean generalizes, the same boundary-regeneration trick could apply to necklines, cuffs, or hemlines, turning the model into a general 'regenerate from this boundary' tool for garment edits."],"forward_implications":["If the method works as described, a virtual try-on system can expose controls for shoulder coverage, neckline depth, garment length relative to the hip, and sleeve length, and each control changes the generated outfit rather than only the prompt.","Because the position labels are measured in pixel space and normalized, they give the model explicit geometric guidance that a latent text or image embedding alone does not carry; the paper states this is why these features work well despite little training.","The mask-plus-boolean mechanism makes it possible to regenerate a waist or crop-top area upward from the mask boundary, letting a user dress a person in a shorter garment than the one in the source photo.","The paper's critique of one-hot semantic labels implies that future versions should encode attributes such as sleeve length as continuous variables between zero and one."],"supporting_citations":[{"why":"Supplies the IP-adapter embedding path into cross-attention that the paper augments with its label vector.","marker":"[12]"},{"why":"Supplies the T2I-adapter design and pretrained down-block adapter used to inject spatial detail.","marker":"[6]"},{"why":"Provides the baseline IDM-VITON parallel U-Net architecture with IP-adapter that this work builds on and pivots from.","marker":"[9]"},{"why":"Provides the VITON-HD dataset of paired person and garment images used for training and evaluation.","marker":"[5]"},{"why":"Provides the Dress Code dataset of paired person and garment images used alongside VITON-HD.","marker":"[18]"},{"why":"Supplies the human-parsing segmentation approach used for masking and for measuring garment positions on the body.","marker":"[17]"},{"why":"Supplies open-set object detection used to locate and label garment bounding boxes during pre-processing.","marker":"[19]"},{"why":"Supplies segmentation masks used to create the dynamic masking and pixel-space measurements.","marker":"[20]"}],"fun_headline_variants":["Precise fit labels make virtual try-on images steerable","Labels let shoppers steer virtual try-on fit and style","One garment, many fits: precise labels for try-on","Virtual try-on gets precise length and style controls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the custom label vector has its own influence on what the diffusion model generates; if the model already gets the same information from the text prompt and the garment image, the whole control mechanism adds nothing.","fun_headline_variants_meta":{"raw":{"variants":["Precise fit labels make virtual try-on images steerable","Labels let shoppers steer virtual try-on fit and style","One garment, many fits: precise labels for try-on","Virtual try-on gets precise length and style controls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1167,"prompt_tokens":822,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":280}},"tokens_in":438,"tokens_out":345,"duration_ms":3592,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:40:08.225878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold the garment image, person image, mask, and text prompt fixed, then sweep one continuous label such as shoulder fraction from zero to one or neckline depth from positive to negative across several random seeds; if the generated images show no systematic change in shoulder coverage or neckline, the label channel is not learned. A second check is to retrain with label values randomly shuffled and compare outputs to the trained model; indistinguishable outputs would mean the label vector is being ignored.","supporting_citations":[{"cited_title":"Improving diffusion models for authentic virtual try-on in the wild, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the baseline IDM-VITON parallel U-Net architecture with IP-adapter that this work builds on and pivots from."},{"cited_title":"Viton-hd: High-resolution virtual try-on via misalignment-aware normalization, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the VITON-HD dataset of paired person and garment images used for training and evaluation."},{"cited_title":"Dress code: High-resolution multi-category virtual try-on, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the Dress Code dataset of paired person and garment images used alongside VITON-HD."},{"cited_title":"Self-Correction for Human Parsing","cited_arxiv_id":"1910.09777","evidence_quote":"Supplies the human-parsing segmentation approach used for masking and for measuring garment positions on the body."},{"cited_title":"Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies open-set object detection used to locate and label garment bounding boxes during pre-processing."},{"cited_title":"Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick","cited_arxiv_id":null,"evidence_quote":"Supplies segmentation masks used to create the dynamic masking and pixel-space measurements."}],"review_version":1}