{"id":"409c45e8-7305-4753-8425-ba762f9cac11","arxiv_id":"2411.10187","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffusion-based adapter performs virtual try-on as outpainting from a reference face and garment, reporting FID 5.56 and 7.23 on VITON-HD.","lead":"Try-On-Adapter generates a full try-on photo from just a face crop and a garment image, by imagining the rest of the body instead of editing a standing photo. The method reports state-of-the-art FID scores on VITON-HD and adds text and pose controls, which could make virtual try-on tools simpler for online shoppers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FID-based SOTA claim is not yet load-bearing: TOA's outpainting protocol lets it optimize global distribution matching, so an unconditional or matched-input control is needed before 5.56/7.23 can be read as try-on quality.","rationale":"The reader's weakest assumption identifies the same load-bearing concern I would attack: the FID comparison in Table 1 is apples-to-oranges. TOA's outpainting paradigm changes substantially more pixels than inpainting baselines, and FID only measures closeness of the generated set to the real-image distribution, not whether the conditions were actually used correctly. The paper's own explanation for lower SSIM and higher LPIPS concedes that the paradigms differ in how much content is preserved, which makes the standalone FID comparison insufficient evidence for the SOTA claim. The architecture and qualitative results are plausible, and the outpainting idea is a useful reformulation, so this is not grounds for rejection. The correct remedy is to add a matched-protocol or unconditional-baseline control before the quantitative headline is trusted. I therefore leave the reader's CONDITIONAL verdict unchanged.","tokens_in":12504,"tokens_out":4843,"duration_ms":52849,"concrete_test":"Run TOA on the VITON-HD test pairs under three conditions: (a) face+garment from the same image, (b) face and garment zeroed or ablated, and (c) face+garment from random cross pairs; compute FID, SSIM, and LPIPS for each. If (b) or (c) yields FID within about 1 point of the reported 5.56/7.23, FID is not capturing conditional try-on fidelity. As a second check, give TOA the full standing image as its reference, which the paper states is a special case of facial-area expansion, and compare FID/SSIM/LPIPS to OOTDiffusion under identical inputs; if the FID advantage disappears, the headline claim is a protocol artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1's headline FID scores (5.56 paired, 7.23 unpaired) are computed under different input protocols: TOA receives only a face crop and garment, while every baseline receives a full standing person image and garment. Because FID compares the distribution of generated images to the real test-set distribution, TOA is free to choose pose, body shape, background, and composition that match the dataset prior; inpainting baselines are constrained to preserve all unmasked pixels of their input. The paper itself reports much lower SSIM (0.772 vs 0.862-0.877) and higher LPIPS (0.178 vs 0.071-0.117) and attributes this to the paradigm, but this is exactly the regime in which a low FID can be achieved by ignoring the conditions and producing typical VITON-HD-like images. The small gap between TOA's paired and unpaired FID (5.56 vs 7.23) suggests the metric is not very sensitive to which face/garment pair is provided. Thus the central quantitative assertion of state-of-the-art realism is not established; it needs a matched-protocol or unconditional-baseline control.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Try-On-Adapter (TOA), a virtual try-on method that treats the task as outpainting rather than inpainting. Given only a cropped face image, a garment image, and optional text/pose conditioning, TOA generates a full dressed-person image using a Stable Diffusion backbone augmented with an IP-Adapter-style cross-attention block for face/garment embeddings and a Reference U-Net for garment detail preservation. The authors introduce a two-stage training pipeline (0.2M Internet images for pretraining, VITON-HD for finetuning) and report qualitative comparisons against OOTDiffusion, cross-dataset and low-quality-input experiments, and quantitative results on VITON-HD: FID 5.56 (paired) and 7.23 (unpaired), with SSIM 0.772 and LPIPS 0.178.","tokens_in":1593,"tokens_out":1763,"duration_ms":50361,"significance":"If the central claim were established, the paper would be practically significant: it would show that virtual try-on does not require a high-quality full standing image, and that pose and text editing can be combined with face/garment preservation in one diffusion-based system. The qualitative results support the feasibility of this outpainting paradigm, and the cross-dataset examples with internet garment screenshots are a useful demonstration of real-world applicability. However, the state-of-the-art quantitative claim is not yet load-bearing. The main FID comparison in Table 1 uses different input protocols for TOA and all baselines, and no matched-protocol or unconditional control is provided; the much lower SSIM/LPIPS are attributed to the paradigm but not demonstrated as inherent to outpainting. I also note that the manuscript includes no implementation details, no error bars, and no public code/checkpoints, so reproducibility cannot currently be assessed.","major_comments":[{"comment":"The headline FID comparison is not a fair protocol comparison. Every inpainting baseline receives a full standing person image plus a garment, whereas TOA receives only a face crop and a garment. Because FID measures the distributional match between generated and real VITON-HD images, TOA is free to choose pose, body shape, and background composition to match the dataset prior, while the inpainting baselines are constrained to preserve all unmasked pixels of a fixed input. The reported FID advantage (5.56 paired, 7.23 unpaired) may therefore reflect a difference in input freedom rather than better try-on quality. The authors attribute this to the paradigm, but they do not provide the necessary control: an unconditional or face/garment-only baseline, or a matched-input version of an inpainting method. Without such a control, the state-of-the-art claim is not established.","section":"Quantitative Results / Table 1"},{"comment":"The paper's own SSIM (0.772 vs 0.862–0.877) and LPIPS (0.178 vs 0.071–0.117) show much weaker pixel-level and perceptual similarity to the ground truth, and the text states this is 'mainly due to the difference of paradigms'. This attribution needs quantitative support rather than assertion. The manuscript should either (a) feed the same face-only + garment protocol to a baseline, (b) feed a full standing image to TOA (the paper says the randomly-expanded face can cover the whole image), or (c) provide an ablation that isolates the amount of preserved context. As it stands, the low SSIM/LPIPS could indicate weaker conditioning adherence rather than an inherent property of outpainting.","section":"Quantitative Results / Paired Evaluation"},{"comment":"The paper's quantitative results are not reproducible or fully auditable as submitted. The training schedule relies on 0.2M Internet images, two stages, a random face-expansion range, a classifier-free guidance drop rate, and face/garment extraction with SAM, but the manuscript defers all concrete hyperparameters, dataset sizes, and evaluation details to an Appendix that is not included. Furthermore, FID scores are reported without error bars or multiple seeds; on a 6K test set, the difference between TOA and the next-best FID may be within sampling noise. Please report variance estimates and a detailed implementation description.","section":"Training and Inference / Experimental settings"},{"comment":"The qualitative comparison with OOTDiffusion in Fig. 4 is made under different input modalities (full standing image for OOTDiffusion vs face+garment for TOA) and is evaluated only subjectively. This is a reasonable illustrative comparison, but it should not be extended into a claim of better performance. A user study with side-by-side images or a matched-input quantitative evaluation is needed before the qualitative advantage is used as evidence for the SOTA claim.","section":"Qualitative comparisons"}],"minor_comments":[{"comment":"The caption contains the typo 'VITION-HD'; it should read 'VITON-HD'.","section":"Single Dataset Evaluation / Fig. 3 caption"},{"comment":"The notation 'Ez0,y∼[ε(q0),y]Eq(ϵ)' is malformed; please clarify the joint distribution over z0 and y.","section":"Eq. (4)"},{"comment":"The first method row is labelled 'VITON-HD (Choi et al. 2021)', which is the dataset name rather than a method name; rename it to the actual baseline method to avoid confusion.","section":"Table 1"},{"comment":"The claim that traditional methods are 'just a special case' of TOA because the face region can be expanded to the whole image is not fully supported: a whole standing image reference contains body, background, and garment-pose information that the model is expected to preserve rather than imagine, so the conditional task changes in kind, not just in input size. Please clarify what exactly is recovered when the face box is expanded to the full image.","section":"Constructing the Training Data"},{"comment":"Equation (2) has a likely formatting error in the noise-scale term ('1 − α2(t)p' should presumably be '1 − α2(t)'); please correct the typesetting.","section":"Diffusion Models background"}],"recommendation":"major_revision","confidential_remarks":"The protocol mismatch in Table 1 is the main technical concern. The paper's central contribution—a new input paradigm—is interesting and potentially publishable, but the state-of-the-art claim should be either backed by a matched-protocol control or substantially softened. Please also ensure the Appendix with implementation details is included in the submitted version, since the current text explicitly defers all experimental details to a missing appendix and the quantitative results are otherwise unverifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on virtual try-on: the paper reframes try-on as outpainting from a face crop plus garment, which is a genuinely useful simplification for real-world e-commerce. The architecture is a sensible composition of IP-Adapter-style decoupled cross-attention and an AnimateAnyone-style reference U-Net, with ControlNet for pose. Qualitative results show plausible preservation of face and garment, including when the garment is a street screenshot, which is a nice demonstration of robustness.\n\nThe soft spot is the quantitative headline. Table 1 claims SOTA FID (5.56 paired, 7.23 unpaired) but the comparison is apples-to-oranges: TOA receives only a face crop and garment, while baselines get a full standing image and garment. With an outpainting setup, the model is free to choose pose, background, and body shape, and FID measures distribution match to the test set, so a low FID can be achieved by generating typical VITON-HD-looking images without actually conditioning faithfully on the given garment/face. The paper itself reports much lower SSIM (0.772 vs 0.862-0.877) and higher LPIPS (0.178 vs 0.071-0.117), and the small gap between paired and unpaired FID (5.56 vs 7.23) is consistent with the metric being insensitive to which identity/garment pair is used. The authors attribute the lower similarity scores to paradigm difference, which is partly fair, but it doesn't establish that the FID gain is due to better try-on quality rather than protocol. A matched-protocol control (e.g., an unconditional generation FID, or giving the baselines the same face crop) or a condition-omission test would be needed to make the SOTA claim load-bearing.\n\nOther issues are minor: implementation details are thin (training specifics, hyperparameters, FID sample sizes) and no code is released. The text and pose control claims are supported qualitatively but not quantitatively. None of this sinks the core idea; it just means the current draft oversells the numbers.\n\nI'd send this to a serious referee. The paradigm shift is real, the problem is practically motivated, and the shortcomings are addressable. The revision should either re-run comparisons under matched input protocols or explicitly present the FID as a distribution-quality measure of the outpainting setting, not as a head-to-head SOTA claim.","headline":"Outpainting-based try-on is a real idea; the FID SOTA claim is not yet supported by the protocol-mismatched comparison.","tokens_in":13279,"tokens_out":2937,"would_cite":true,"duration_ms":27532,"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":"Try-On-Adapter claims virtual try-on is an outpainting problem: given only a face image and a garment image, a diffusion model imagines the rest of the person and reports FID 5.56 paired and 7.23 unpaired on VITON-HD.","keywords":["virtual try-on","outpainting","image-as-prompt","latent diffusion","Try-On-Adapter","garment preservation","VITON-HD","controllable generation"],"falsifier":"Run TOA and the inpainting baselines under identical input conditions—either give all methods the same minimal face-plus-garment input, or force the same pose and background on every output—and recompute FID on VITON-HD; if TOA no longer beats StableVITON and OOTDiffusion, the central SOTA claim is an artifact of comparing different tasks.","tokens_in":12333,"feed_emoji":"👗","tokens_out":6781,"duration_ms":61488,"temperature":0.7,"pith_summary":"Try-On-Adapter (TOA) tries to establish that virtual try-on should be treated as outpainting, not inpainting. Where earlier systems require a clean full-body standing photo and only fill in the garment region, TOA takes a reference face and a garment image as its only visual inputs and lets the diffusion model imagine pose, body, and background. The paper argues this makes try-on simpler for users and more controllable, since text prompts and pose maps can steer the imagined body. On VITON-HD it reports the lowest FID among compared methods, 5.56 paired and 7.23 unpaired, while explicitly accepting lower SSIM and LPIPS than inpainting baselines because it regenerates more of the image. A sympathetic reading is that the point is practical generation from casual, low-quality inputs rather than pixel fidelity to a reference standing image.","feed_headline":"Face plus garment is enough: try-on becomes outpainting","feed_subtitle":"TOA imagines the whole body from a face crop and a garment shot, reporting the lowest FID on VITON-HD.","key_machinery":"The mechanism that carries the argument is a pair of trainable adapter blocks wrapped around a frozen diffusion U-Net. The first, Comprehension- and Fusion-inspired Cross-attention, encodes the face and garment with CLIP's full hidden states, runs a small attention block over them with learned query, key, and value projections, and injects the result through decoupled cross-attention next to the text embedding. The second, Reference-preserving Self-attention, copies the denoising U-Net as a frozen Reference U-Net and concatenates its feature maps into the self-attention of each denoising layer, which is what keeps fine garment details from being lost. These adapters are trained with denoising score matching on a dataset reconstructed from public try-on images by SAM face crops and CogVLM captions, then fine-tuned in a second stage; ControlNet is added at inference for pose control.","core_discovery":"The central claim is that the try-on task can be reformulated as 'given a face and a garment, imagine the rest,' and that a diffusion model with two lightweight adapters is enough to do it convincingly. The Try-On-Adapter fuses CLIP encodings of the face and garment into the cross-attention layers of a frozen Stable Diffusion U-Net, alongside the text prompt, and uses a second Reference U-Net to feed garment appearance into self-attention so patterns and textures survive generation. Training data is built by cropping faces from standing images with SAM and captioning each image with CogVLM, then a two-stage schedule pretrains on web images and fine-tunes on VITON-HD. The paper reports that this recipe preserves the given face and garment, accepts text edits such as 'pink dress,' supports pose control through ControlNet, and can even consume a full standing image as its reference because random face-crop expansion makes that a special case.","pith_inferences":["The paper leaves implicit that the fusion attention is not clothing-specific; the same face-plus-item recipe could plausibly extend to hats, bags, shoes, and other accessories, which the conclusion names as future work.","A fairer protocol test that the paper does not run would compare all methods on identical minimal input (face plus garment only), or with pose and background fixed; without that, the FID advantage is entangled with the outpainting model's freedom to invent the rest of the scene.","If the paradigm scales, the product-level consequence would be a shift in try-on from warping a given model photo toward synthesizing a new photo of the user from a phone selfie, with identity preservation as the main constraint."],"forward_implications":["A user would only need a selfie-style face crop and a product garment photo, not a clean full-body standing image, to get a realistic try-on.","Garment color and style become editable through text prompts, and human pose becomes controllable through ControlNet, in a way that inpainting try-on methods do not naturally support.","Because the model is free to generate pose, background, and body proportions, outputs can be more diverse than fidelity-to-input inpainting results.","Randomly enlarging the face crop lets the same trained model optionally accept a full standing reference image, making inpainting-style input a special case of the proposed paradigm.","Garment screenshots and lower-quality internet images remain usable, since the garment enters the model through CLIP-level features rather than pixel-aligned warping."],"supporting_citations":[{"why":"Supplies the decoupled cross-attention image-as-prompt mechanism that TOA extends to fuse face and garment.","marker":"Ye et al. 2023"},{"why":"Supplies the Reference U-Net design used to keep garment details in self-attention.","marker":"Hu et al. 2023"},{"why":"Provides the latent diffusion backbone and rich priors that TOA adapts without retraining from scratch.","marker":"Rombach et al. 2022"},{"why":"Provides VITON-HD, the dataset and benchmark used for fine-tuning and all reported quantitative comparisons.","marker":"Choi et al. 2021"},{"why":"The main open-source baseline that TOA compares against and claims to outperform in qualitative and FID results.","marker":"Xu et al. 2024"},{"why":"Provides CLIP, the vision encoder that extracts face and garment embeddings for the image-as-prompt conditioning.","marker":"Radford et al. 2021"},{"why":"Used to crop facial regions from standing images, creating the reference face inputs for training.","marker":"Kirillov et al. 2023"},{"why":"Provides CogVLM, used to caption training images so the model gains text-guided editability.","marker":"Wang et al. 2023"},{"why":"Used to add pose control via ControlNet on top of the adapter.","marker":"Zhang et al. 2023"}],"fun_headline_variants":["Try-On-Adapter: outpainting try-on from face and garment","Face + garment is all you need: TOA outpaints the rest","Outpainting paradigm for try-on, no full-body input required","TOA: state-of-the-art virtual try-on via outpainting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that FID comparisons across the two paradigms are fair: TOA was free to choose pose, background, and body proportions to match the test distribution, whereas inpainting baselines had to stay close to a supplied standing image; if FID rewards that freedom, the reported advantage is protocol-driven rather than a sign of better rendering.","fun_headline_variants_meta":{"raw":{"variants":["Try-On-Adapter: outpainting try-on from face and garment","Face + garment is all you need: TOA outpaints the rest","Outpainting paradigm for try-on, no full-body input required","TOA: state-of-the-art virtual try-on via outpainting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2493,"prompt_tokens":986,"completion_tokens":1507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":1429}},"tokens_in":602,"tokens_out":1507,"duration_ms":10054,"temperature":1.0,"reasoning_tokens":1429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:52:03.295334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TOA and the inpainting baselines under identical input conditions—either give all methods the same minimal face-plus-garment input, or force the same pose and background on every output—and recompute FID on VITON-HD; if TOA no longer beats StableVITON and OOTDiffusion, the central SOTA claim is an artifact of comparing different tasks.","supporting_citations":[],"review_version":1}