{"id":"c887ce19-7819-45ef-ada0-f8cece25a699","arxiv_id":"2501.04666","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-part method, synthetic garment-pair data augmentation plus an error-aware Schrödinger bridge refiner, improves virtual try-on quality on VITON-HD and DressCode-Upper.","lead":"This paper improves virtual try-on, the task of putting a product photo of a garment onto a photo of a person, using two ideas: synthetic training pairs extracted from ordinary clothing photos, and a refinement model that adds noise only where the first try-on attempt made mistakes. If it works as reported, it makes generated clothing textures and text sharper, and gives other try-on models a cheap boost by augmenting their training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) is not the inverse of the forward process: it omits the known source term x1 and the x0 normalization, so the EARSB sampler as written does not implement a Schrödinger bridge even when M=1.","rationale":"The strongest claim of the paper is that EARSB improves try-on quality and that synthetic pairs help. The synthetic-pair component has reasonable support: Table 3 shows consistent gains across CAT-DM and StableVITON, and Table 4 indicates the augmentation label matters. The EARSB component, by contrast, is the paper's main novelty and its mathematical presentation is the weakest link. Inverting the forward process in Eqs. (2)-(4) with known x1 and a predicted noise gives a different update than Eqs. (5)-(8); the missing x1 term and normalization are not optional. Because the training loss Eq. (9) trains the network to predict M·ε, a perfect network in Eq. (5) estimates μ_t, not x0. This is a concrete, checkable issue, not a disagreement with the community. If the implementation corrects it, the paper needs revised equations and the experimental gains can be attributed to the stated mechanism; if not, the reported gains are not explained by the algorithm in the text. The reader's conditional verdict is appropriate, but the condition should explicitly require a corrected derivation and released sampler code. I do not find a separate decisive problem in the synthetic-pair augmentation; the filtering bias is a secondary concern that the ablations partially address.","tokens_in":15616,"tokens_out":26124,"duration_ms":237918,"concrete_test":"Rerun the VITON-HD sampling pipeline with the corrected endpoint inversion \\hat x0 = (x_t − (σ_t^2/S_t)x1 − √(Σ_t)ε^r_θ)/(σ̄_t^2/S_t), including the matching x1 term in the x_{t−∆t} interpolation, and compare FID/KID/SSIM to Table 5 and Table 1 for EARSB (w.o. M), EARSB (full), and EARSB+H2G-UH. If the metrics move by more than about 0.5 FID or 0.005 SSIM, the published equations are not the implemented sampler and the reported gains are unverified. A minimal analytic check: take x0=0, x1=1, t=0.8, M=1, and an oracle denoiser; Eq. (5) gives \\hat x0=0.8 while the true inverse is 0, so the subsequent Eq. (7) posterior mean differs from the exact Brownian-bridge posterior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Let S_t = σ_t^2 + σ̄_t^2. Inverting Eq. (2) yields x0 = (S_t/σ̄_t^2)(x_t − (σ_t^2/S_t)x1 − √(Σ_t) M ε). Eq. (5) instead sets \\hat x0 = x_t − √(Σ_t) ε^r_θ. With a perfect denoiser, Eq. (5) returns μ_t = (σ̄_t^2/S_t)x0 + (σ_t^2/S_t)x1, not x0; for example at t=0.8 with S_t=1 it gives 0.2x0+0.8x1. Feeding this biased endpoint into Eq. (7) evaluates the Brownian-bridge posterior mean with the wrong conditioning value, so the reverse trajectory is anchored to the artifact image x1 and does not follow the reverse of Eqs. (2)-(4). This is not a subtlety of the error-map scaling: the missing x1 term is already present when M=1, i.e., in the plain I2SB baseline. The correct DDIM-style sampler with the known source endpoint is \\hat x0 = (x_t − (σ_t^2/S_t)x1 − √(Σ_t) ε^r_θ)/(σ̄_t^2/S_t) and x_{t−∆t} = (σ̄_{t−∆t}^2/S_{t−∆t})\\hat x0 + (σ_{t−∆t}^2/S_{t−∆t})x1 + M√(Σ_{t−∆t}) ε. If the released code follows the text, training and inference are mismatched; if it follows the corrected update, Eqs. (5)-(8) are wrong as written. The central EARSB claim therefore rests on an unstated detail that must be resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes two complementary improvements to virtual try-on. First, it trains a human-to-garment model that converts a single image of a clothed person into a canonical product-view garment image, creating (human, synthetic garment) pairs that are filtered and used to augment try-on training under a real/synthetic flag to mitigate the domain gap. Second, it introduces EARSB, a refinement module built on the Image-to-Image Schrödinger Bridge (I2SB): a weakly supervised classifier (WSC) predicts a spatial error map from an initial try-on output, and that map is used to reweight the diffusion noise, condition the denoiser, and guide the reverse sampler. Experiments on VITON-HD and DressCode-Upper report improved FID, KID, SSIM, and LPIPS over GAN and diffusion baselines, and a user study prefers EARSB in 59% of cases on average.","tokens_in":16017,"tokens_out":10228,"duration_ms":91763,"significance":"If the results hold, the synthetic-data contribution is practical and well ablated: Table 3 shows that H2G-UH improves CAT-DM and StableVITON, and Table 4 isolates the benefit of flag-conditioned augmentation over plain mixing and pretraining. The error-map idea is also interesting; Table 5 demonstrates that a meaningful map outperforms a random map and an unweighted I2SB baseline. The paper also ships code, which is valuable for reproducing the experiments. The main weakness is mathematical: the sampling equations in §4.2 do not, as written, invert the forward process, which undermines the central claim that EARSB is a valid Schrödinger bridge. The quantitative results are reported as single numbers without error bars or significance tests, which makes some of the smaller reported gains difficult to evaluate.","major_comments":[{"comment":"The proposed sampler does not invert the forward process in Eqs. (2)-(4). If the noise prediction is perfect (ε^r_θ = ε^r), Eq. (5) gives \\hat{x}_0 = μ_t(x0,x1) = (σ̄_t²/(σ̄_t²+σ_t²)) x0 + (σ_t²/(σ̄_t²+σ_t²)) x1, not x0. The correct denoising estimate must remove the known x1 term and the normalization, e.g. \\hat{x}_0 = (x_t − (σ_t²/(σ̄_t²+σ_t²)) x1 − √Σ_t ε^r_θ) / (σ̄_t²/(σ̄_t²+σ_t²)). Consequently Eq. (7) conditions the Brownian-bridge posterior on the wrong value, and the sampler remains anchored to the artifact image x1. This is not a consequence of the mask M; it already occurs when M is all ones, so the statement that EARSB reverts to I2SB when M=1 is not supported. Please correct the equations or clearly report the actual sampler used in the released code; as written, the theoretical framing of EARSB as a Schrödinger bridge is invalid.","section":"Sec. 4.2, Eqs. (5)-(8)"},{"comment":"The paper introduces the spatially reweighted noise ε^r = M·ε and a reverse update that again injects M·√Σ_t ε, but it provides no derivation that this reweighted stochastic process is a valid Schrödinger bridge or that the reverse update is consistent with the forward marginals of Eqs. (3)-(4). Since the central contribution is an \"Error-Aware ... Schrödinger Bridge,\" the authors should either supply the supporting derivation or explicitly characterize the M-reweighting as a heuristic modification of I2SB and evaluate it on that basis.","section":"Sec. 4.2, Eqs. (3)-(4) and (6)"},{"comment":"All quantitative results are single measurements without error bars, confidence intervals, or significance tests. Several of the key gains that support the synthetic-data claim are small (e.g., Table 3: StableVITON unpaired FID 8.25→8.17; Table 5: EARSB w.o. M vs. full, unpaired FID 8.48 vs. 8.42 with identical SSIM/LPIPS to three decimals), so it is difficult to know whether these differences are reliable. Please report variance over multiple seeds or a paired significance test for the central comparisons.","section":"Tables 1-5 and Fig. 6"}],"minor_comments":[{"comment":"The final loss is written as \"Lins + Lpat,\" which appears to be a typo for \"Limg + Lpat\"; the patch-level loss also omits an explicit averaging or summation over spatial locations.","section":"Sec. 4.1, Eq. (1)"},{"comment":"The sentence \"When M is all zeros (i.e., no error), x1 is believed to be perfect x1 does not need to be refined\" has a missing comma and reads incorrectly; please rephrase.","section":"Sec. 4.2, after Eq. (8)"},{"comment":"The GP-VTON row reports KID 0.86 for VITON-HD unpaired but 0.07 for paired, which is atypical compared with the other baselines; please double-check the KID computation for that row.","section":"Table 1"},{"comment":"The main table uses 25 sampling steps for all diffusion models, while Table 3 uses each method's original step count (2 for CAT-DM, 50 for StableVITON, 25 for EARSB); this makes cross-table comparisons inconsistent and should be clarified.","section":"Sec. 5.1 and Table 3"},{"comment":"The filtering criterion (c) uses LPIPS reconstruction error from a try-on model; since that try-on model is trained on the same benchmark used for evaluation, the potential selection bias or data leakage in synthetic-pair construction should be discussed.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The experimental story is plausible and the ablations are useful, but the sampler equations in §4.2 contain a serious technical error that must be reconciled with the released code before the paper can be accepted. If the code uses the corrected update, the text can be fixed; if the code follows the text, the reported experiments may reflect a training-inference mismatch. I would like to see the authors confirm which one holds and provide the corrected derivation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the synthetic pair augmentation is a genuinely practical idea: they train a human-to-garment model to generate product-view garments from clothed human photos, filter the results, and use the resulting (human, synthetic garment) pairs to augment virtual try-on training. The ablations show this helps both GAN and diffusion baselines, and the real/synthetic conditioning flag is a nice way to handle the domain gap. Second, the paper's other contribution—the error-map-reweighted Schrödinger bridge (EARSB)—has an inconsistency in the sampling equations that needs resolving before the results can be taken at face value. In Eq. (2) the forward process is x_t = μ_t(x0,x1) + √Σ_t ε, with μ_t a convex combination of x0 and x1. Inverting that correctly requires subtracting the x1-dependent term and dividing by the x0 coefficient. Equation (5) instead sets \\hat x0 = x_t − √Σ_t ε_θ. Even with a perfect noise estimate, that returns μ_t, a blend of the target and the artifact image, not x0. Feeding that biased endpoint into Eq. (6) anchors the reverse trajectory to the artifact. The problem is already present when M=1, so the claim that the model reduces to I2SB in that limit is also suspect. This could be a typo—the implementation may use the correct update—but the paper needs to state the correct inverse and the code needs to match.\n\nWhat the paper does well: the experiments are extensive. They ablate the error map (w.o. M, random M, no classifier guidance), compare against GAN and diffusion baselines on two datasets, report FID/KID/SSIM/LPIPS, include a user study, and show a thoughtful failure case and limitation section. The weak supervision of the error classifier with just a few hours of bounding-box labels is clever, and the precision-recall comparison against unsupervised baselines supports it.\n\nSoft spots: beyond the equation issue, all metrics are single numbers with no error bars or significance tests, which is common but worth flagging. The synthetic data filtering uses reconstruction LPIPS as a criterion, which could select for easy examples; the authors acknowledge related imperfection in the limitations. Those are minor compared with the sampling equations.\n\nThis paper is for people in virtual try-on and anyone using I2SB-style bridges for image refinement. It deserves a serious referee because the empirical setup is careful and the synthetic data idea transfers. But the referee should ask for the corrected sampling derivation and for the released code to confirm the training/sampling match.\n\nMy recommendation: engage with it, but conditional. Ask the authors to fix the equations and provide code before accepting.","headline":"The synthetic-pair augmentation idea is practical and well-ablated, but EARSB's sampling equations as written do not invert the bridge forward process, so the central claim needs correction or code verification.","tokens_in":16529,"tokens_out":4576,"would_cite":true,"duration_ms":44664,"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":"This paper claims that virtual try-on quality can be pushed past prior baselines by augmenting training with ~21,000 synthetic (person, garment) pairs and refining base-model outputs with an error-aware Schrödinger bridge.","keywords":["virtual try-on","synthetic data augmentation","human-to-garment extraction","Schrödinger bridge","error-aware noise scheduling","weakly-supervised error classification","diffusion refinement","image generation"],"falsifier":"Take a toy problem with known endpoints — for example, two simple image classes or Gaussian point clouds where the true conditional distribution of x0 given x1 is computable — and run the masked schedule of Eqs. (3)-(8) with a deliberately imperfect error map. If the masked process still transports samples to the correct x0 distribution within the I2SB objective's tolerance, the bridge claim holds; if the chain shows systematic bias in low-M regions, the error-aware success is a property of this specific schedule rather than of Schrödinger bridge theory.","tokens_in":15390,"feed_emoji":"👕","tokens_out":9168,"duration_ms":75136,"temperature":0.7,"pith_summary":"This paper tries to establish that two low-cost ingredients can push virtual try-on quality past prior GAN- and diffusion-based methods: synthetic training pairs, made by a human-to-garment model that turns photos of clothed people into canonical product-view garment images, and a refinement stage that repairs only the flawed parts of a base model's output. The refinement stage, called EARSB, uses a weakly-supervised error classifier to build a heatmap of artifacts and then feeds that heatmap into a Schrödinger bridge's noise schedule, so noise is added precisely where errors are. On VITON-HD and DressCode-Upper the paper reports better FID, KID, and SSIM than seven prior baselines, and a user study prefers its outputs in 59% of pairwise comparisons. If the paper is right, existing try-on models can be improved without collecting new paired product images or retraining them from scratch.","feed_headline":"Synthetic pairs plus targeted refinement lift virtual try-on quality","feed_subtitle":"Refinement adds noise only where artifacts appear; synthetic pairs add 21K training samples.","key_machinery":"Two mechanisms carry the argument. The human-to-garment model is a ClothFlow-style UNet with flow-like feature warping that maps the segmented clothing region of a person photo onto the canonical product view, trained with an L1 reconstruction plus adversarial loss on existing paired data such as VITON-HD. The refinement mechanism, EARSB, extends the I2SB Schrödinger bridge: the noise in the forward process is replaced by εr = M·ε, where M = WSC(x1, C) is the sigmoid error heatmap from a weakly-supervised two-encoder classifier trained with a few hours of bounding-box labels; training minimizes the mean squared error between the predicted reweighted noise and εr, and sampling re-injects noise and guidance only where M is high. The error map thus acts as a per-pixel noise scheduler: pixels judged correct are copied through nearly untouched, and pixels judged artifact-ridden are re-synthesized.","core_discovery":"On its own terms, the paper's central claim is that virtual try-on errors are best fixed by knowing where they are. The human-to-garment model — a flow-based UNet trained on existing paired data — extracts a canonical garment view from a single clothed-person image, and after filtering the resulting pairs by background cleanliness, frontal pose, and LPIPS reconstruction error, roughly 21,000 synthetic (human, garment) pairs from DeepFashion2 and UPT improve multiple base try-on models when mixed into training at a 15% batch ratio with a real/synthetic conditioning flag. The second claim is that the image-to-image Schrödinger bridge of I2SB becomes a sharper refiner when the noise is reweighted pointwise by the error map, εr = M·ε, so clean regions are copied through nearly noiseless while artifact regions receive more noise, with the same error map masking the classifier guidance term during sampling. The full pipeline, EARSB with synthetic augmentation, reports the best FID, KID, and SSIM on both benchmarks, and the ablations trace the gain to the error map itself: uniform noise, random maps, inpainting of flagged regions, and removed guidance all do worse.","pith_inferences":["The error-masked process εr = M·ε is asserted rather than derived; a natural test is whether the masked forward-backward process still converges to the true endpoints on a toy distribution, and if not, the gains may depend on the specific error-map statistics rather than a general bridge principle.","The same refine-where-flagged recipe should transfer to other image-to-image tasks with cheap error localizers, such as face restoration or text rendering in edited images, though the paper only demonstrates try-on.","The synthetic-pair gain may scale further with 3D-aware garment extraction; the paper's own limitation note reports residual texture deformations because the generator operates in image space without 3D information."],"forward_implications":["Any existing virtual try-on model, GAN- or diffusion-based, can be wrapped by EARSB and improved without modifying the base model's weights.","Synthetic pairs substitute for scarce copyrighted product photos: roughly 21,000 filtered pairs sourced from unlabeled human images suffice to improve several published try-on models.","The error-aware schedule keeps image quality at fewer sampling steps, so refinement can be run cheaply when the initial image is mostly good.","Because the error classifier is trained on a specific base model's artifacts, retargeting the refiner to a new base model costs only a few hours of bounding-box labeling.","The real/synthetic conditioning flag matters: mixing synthetic pairs without the label hurts performance, while labeled mixing beats pretrain-then-finetune."],"supporting_citations":[{"why":"Supplies the I2SB Schrödinger bridge that EARSB reweights with the error map, including the training objective and sampling equations.","marker":"[24]"},{"why":"Provides the flow-based UNet and garment-warping modules used for both the human-to-garment generator and EARSB's denoiser.","marker":"[15]"},{"why":"The GAN try-on model whose VITON-HD outputs EARSB refines; its results are the main GAN baseline.","marker":"[37]"},{"why":"Paired benchmark dataset that trains the human-to-garment model and evaluates try-on quality; HR-VTON is also a baseline and initial-image source.","marker":"[21]"},{"why":"Base model that generates DressCode-Upper initial images for EARSB and serves as a GAN baseline.","marker":"[31]"},{"why":"The diffusion try-on model whose outputs EARSB(SD) refines.","marker":"[7]"},{"why":"Diffusion baseline that the H2G-UH augmentation is shown to improve, and a user-study comparison target.","marker":"[19]"},{"why":"Source of unpaired upper-body human images filtered into the H2G-UH synthetic pairs.","marker":"[25]"},{"why":"Source of unpaired human images filtered into the H2G-FH synthetic pairs.","marker":"[36]"},{"why":"Justifies the classifier guidance term in the sampling update by estimating the score with the denoised image.","marker":"[8]"}],"fun_headline_variants":["Error-aware noise scheduling sharpens virtual try-on","Synthetic pairs and targeted noise lift try-on quality","Fix only try-on artifacts with error-aware noise","21K synthetic pairs refine virtual try-on output","Targeted refinement beats uniform noise in try-on"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that multiplying the bridge noise pointwise by an arbitrary error map, εr = M·ε, still yields a valid Schrödinger bridge that can be trained with the standard I2SB objective; the paper does not derive that the masked noise preserves the required forward-backward marginals or that the reverse-time sampler remains consistent.","fun_headline_variants_meta":{"raw":{"variants":["Error-aware noise scheduling sharpens virtual try-on","Synthetic pairs and targeted noise lift try-on quality","Fix only try-on artifacts with error-aware noise","21K synthetic pairs refine virtual try-on output","Targeted refinement beats uniform noise in try-on"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2684,"prompt_tokens":1046,"completion_tokens":1638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":1580}},"tokens_in":662,"tokens_out":1638,"duration_ms":10965,"temperature":1.0,"reasoning_tokens":1580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:26:46.280390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a toy problem with known endpoints — for example, two simple image classes or Gaussian point clouds where the true conditional distribution of x0 given x1 is computable — and run the masked schedule of Eqs. (3)-(8) with a deliberately imperfect error map. If the masked process still transports samples to the correct x0 distribution within the I2SB objective's tolerance, the bridge claim holds; if the chain shows systematic bias in low-M regions, the error-aware success is a property of this specific schedule rather than of Schrödinger bridge theory.","supporting_citations":[{"cited_title":"I2sb: image-to-image schr¨odinger bridge","cited_arxiv_id":null,"evidence_quote":"Supplies the I2SB Schrödinger bridge that EARSB reweights with the error map, including the training objective and sampling equations."},{"cited_title":"Clothflow: A flow-based model for clothed person generation","cited_arxiv_id":null,"evidence_quote":"Provides the flow-based UNet and garment-warping modules used for both the human-to-garment generator and EARSB's denoiser."},{"cited_title":"GP- VTON: Towards general purpose virtual try-on via collabo- rative local-flow global-parsing learning","cited_arxiv_id":null,"evidence_quote":"The GAN try-on model whose VITON-HD outputs EARSB refines; its results are the main GAN baseline."},{"cited_title":"High-resolution virtual try-on with misalignment and occlusion-handled conditions","cited_arxiv_id":null,"evidence_quote":"Paired benchmark dataset that trains the human-to-garment model and evaluates try-on quality; HR-VTON is also a baseline and initial-image source."},{"cited_title":"Towards squeezing-averse virtual try-on via sequential deformation","cited_arxiv_id":null,"evidence_quote":"Base model that generates DressCode-Upper initial images for EARSB and serves as a GAN baseline."},{"cited_title":"StableVITON: Learning semantic corre- spondence with latent diffusion model for virtual try-on","cited_arxiv_id":null,"evidence_quote":"Diffusion baseline that the H2G-UH augmentation is shown to improve, and a user-study comparison target."},{"cited_title":"Deepfashion: Powering robust clothes recognition and retrieval with rich annotations","cited_arxiv_id":null,"evidence_quote":"Source of unpaired upper-body human images filtered into the H2G-UH synthetic pairs."},{"cited_title":"Towards scalable unpaired virtual try-on via patch-routed spatially- adaptive gan","cited_arxiv_id":null,"evidence_quote":"Source of unpaired human images filtered into the H2G-FH synthetic pairs."},{"cited_title":"Diffusion posterior sampling for general noisy inverse problems","cited_arxiv_id":null,"evidence_quote":"Justifies the classifier guidance term in the sampling update by estimating the score with the denoised image."}],"review_version":1}