acceptodds
Withdrawn from ICLR 2027

RealFit: Measurement-Conditioned Virtual Try-On

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The core value of real-world try-on lies in allowing consumers to both a) visualize how a garment looks on themselves, as well as b) see how a garment will fit on them: whether any given size will stretch tightly over body contours, or drape loosely with excess slack. Prevailing Virtual Try-On (VTO) models address the first use-case, but ignore the second: by conditioning solely on 2D visual inputs, they default to synthesizing deceptively well-fitted try-on results regardless of garment dimensions. Data scarcity is a central bottleneck for developing fit-aware VTO systems. Standard public benchmarks consist of catalog imagery lacking both garment dimensions and body measurements. While recent synthetic approaches leverage 3D physics simulation to generate exact measurement labels, they suffer from sim-to-real domain gaps and remain constrained to simplistic garment topologies. In this work, we address the issue of fit through RealFIT, the first framework for scalable, real-world measurement-conditioned virtual try-on across diverse and structurally complex apparel categories. To bridge the data gap, we introduce a multi-tier data hierarchy that bootstraps geometric priors from large-scale web retail metadata and VLM-assisted measurement annotations, anchored by a newly curated real-world dataset with tape-measured metric ground truth. Using this data, we propose a fit-aware virtual try-on model based on flow matching that integrates a dedicated measurement encoder to condition generation on human and garment dimensions. To enhance the model's sensitivity to metric conditions, we also introduce a Localized Metric Contrastive (LMC) loss, which explicitly boosts the conditioning signal by contrasting velocity fields across paired sizing perturbations within targeted spatial boundaries. Furthermore, we establish a systematic evaluation protocol to benchmark fit accuracy. Extensive experiments demonstrate that RealFIT achieves state-of-the-art photorealism and fit accuracy and controllability across varied body geometries and complex garment styles. We plan to release our benchmark dataset, evaluation suite, and code upon acceptance.

Primary area:
generative models
Keywords:
Virtual Try-On, Fit, Generative Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rbcXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RealFIT, a measurement-conditioned virtual try-on framework designed to generate realistic try-on images that reflect physical garment dimensions. The authors construct two datasets, RealFIT-Web and RealFIT-Bench, combining large-scale automatically annotated data with real-world tape-measured garments. The method builds on a flow-matching model with a dedicated measurement encoder and a Localized Metric Contrastive (LMC) loss to improve sensitivity to size changes. The paper also introduces two evaluation metrics, Landmark Fit Score (LFS) and Size Monotonicity (SM), to assess fit accuracy and size controllability.

Critical strengths
  1. The paper addresses an important limitation of existing virtual try-on methods. While recent approaches produce realistic images, they generally do not account for the physical dimensions of garments. Conditioning on actual garment measurements is a meaningful step toward more practical virtual try-on systems.
  2. The dataset is a valuable contribution. In particular, RealFIT-Bench provides real-world garment measurements that are largely missing from existing VTO benchmarks. The combination of large-scale weakly annotated data and a smaller dataset with physical measurements is a reasonable approach to addressing data scarcity.
  3. The experimental results are promising. RealFIT achieves strong performance in both image quality and size controllability. The ablation studies are also informative, particularly the substantial drop in SM when ill-fit augmentation is removed, which supports the importance of explicitly learning from size variations.
Critical weaknesses
  1. The paper aims to achieve centimeter-level virtual try-on, but neither LFS nor SM directly measures physical accuracy. For example, when garment length increases from 60 cm to 70 cm, SM could still reach 100% even if the generated change is much smaller than expected. Since RealFIT-Bench lacks real multi-size ground truth for the same person and garment, the current experiments mainly demonstrate monotonic size control rather than physical accuracy. A more direct evaluation would strengthen the central claim.

  2. In Table 1, FIT is evaluated only on upper-body garments, while RealFIT is evaluated on the full benchmark. These results are not directly comparable. The authors should report RealFIT's performance on the same upper-body subset for a fair comparison.

  3. RealFIT-Web uses existing VTO datasets and retailer size charts, while RealFIT-Bench contains crowdsourced human images and body measurements. The paper does not clearly explain the relevant data licenses, redistribution permissions, or participant consent procedures. The authors should clarify these details and specify which parts of the datasets will be publicly released.

Questions for authors
  1. How do the authors verify that generated size changes correspond to actual physical measurements, rather than just being monotonic? Can this be evaluated using real multi-size ground truth?

  2. Could the authors report RealFIT's results on the same upper-body subset used to evaluate FIT in Table 1?

  3. Could the authors clarify the data licensing and participant consent procedures, as well as which parts of the datasets will be publicly released?

Reviewer_d5VwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles a meaningful problem and proposes an effective-looking framework for improving fit-aware VTO. The synthetic ill-fit augmentation and measurement-specific supervision are promising, and the reported ablations provide evidence that these components improve responsiveness to garment measurements.

However, I believe the current experiments more convincingly establish relative fit controllability than physically calibrated, measurement-accurate virtual try-on.

The most important missing evidence is whether the model generates quantitatively correct garment fit changes when supplied with real-world measurements, rather than merely producing plausible changes in the expected direction.

Critical strengths

Establishing evaluation protocols: Landmark Fit Score and Size Monotonicity is a great initiative needed by the community. Putting in the effort to collect RealFIT-Bench is non trivial, although it would be more valuable if there were data of the same person wearing the same garment in at least 3 different sizes.

Critical weaknesses

1. Evaluation Does Not Establish Physically Calibrated Fit Control

The paper evaluates fit awareness using Landmark Fit Score (LFS) and Size Monotonicity (SM).

LFS measures garment coverage discrepancy against a ground-truth try-on image at the observed garment size. SM evaluates whether generated garment coverage increases monotonically as the input garment measurements are uniformly scaled from −20% to +20%.

Both are reasonable metrics, but they measure different and incomplete aspects of measurement awareness:

  • LFS evaluates fit accuracy at the observed size, but cannot establish whether the model accurately responds to alternative measurements.
  • SM evaluates the ordering of generated garment coverage across different measurement inputs, but does not assess whether the magnitude of the generated fit change corresponds to physical reality.

For example, suppose that when garment length is increased by 10% and 20%, the real garment's hemline moves downward by 5% and 15%, respectively. A model that instead produces changes of 10% and 20% would still achieve perfect SM, despite being quantitatively incorrect.

Furthermore, because SM uniformly scales all garment dimensions, a model could potentially achieve high SM by responding to garment length alone while ignoring bust, waist, or sleeve measurements.

Suggested improvements:

Ideally, the evaluation should include real photographs of the same person wearing the same garment design in three or more physically measured sizes, with pose and camera controlled as closely as possible. Such data would allow the authors to evaluate not only monotonicity but also whether the generated fit changes match the actual changes between sizes.

Even a relatively small, carefully collected benchmark would provide valuable evidence.

If possible, it would be more informative to independently vary individual garment measurements, rather than scaling all dimensions together, to determine whether the model can control distinct aspects of garment fit.

If collecting such data is infeasible, the paper should more clearly acknowledge that SM measures directional consistency, not physical measurement accuracy.

2. Reliability of the Training Measurements and Synthetic Ill-Fit Supervision

The construction of RealFIT-Web raises a more fundamental question about what the model actually learns.

According to Section 3.1, the training data is constructed primarily through two procedures:

  1. A VLM estimates absolute garment measurements from images, first estimating the person's height and then inferring garment dimensions from visual proportions.
  2. Nano Banana 2 synthesizes ill-fitting examples by visually altering garment sizes, after which the VLM re-estimates measurements for the modified images.

Although these procedures can produce useful training examples, neither inherently provides physically calibrated measurement supervision.

In particular, estimating absolute measurements from a single image is inherently ambiguous without reliable external scale information. The resulting VLM annotations may be more reliable at capturing relative garment proportions or qualitative fit differences than actual dimensions in centimeters.

Likewise, the ill-fit augmentation generates visual changes first and then assigns estimated measurements to the results. It does not generate garment deformations from independently specified, physically verified garment dimensions.

Consequently, the training procedure appears particularly well suited to teaching the model a relationship between visually plausible garment-size changes and pseudo-measurement annotations, but it is less clear whether that relationship is accurately calibrated to real-world measurements.

This raises the possibility that RealFIT primarily learns a numerically conditioned version of the qualitative fit-editing behavior exhibited by Nano Banana 2.

Synthetic-data distillation is not inherently problematic. It can be an effective way to specialize a generative model. However, it is important to distinguish such specialization from learning physically accurate garment fit.

Suggested improvements:

Given thatThe authors already have a tape-measured benchmark, RealFIT-Bench, is already available, it can be used to validate the training-data annotation procedure.

For example, report:

  • Absolute measurement errors between VLM-estimated measurements and tape-measured ground truth, broken down by measurement type.
  • Where real paired-size examples are available, errors in the differences between estimated measurements and their physically measured counterparts.
  • Separate statistics for VLM-estimated measurements and measurements obtained directly from retailer size charts.

This would help determine whether the training annotations contain reliable absolute metric information or primarily relative fit information.

3. Baseline Comparisons Do Not Fully Test the Value of Measurement-Specific Conditioning

The paper demonstrates improvements over conventional VTO baselines, including IDM-VTON and FIT, as well as multimodal image editors such as Qwen-Image-Edit and Nano Banana 2.

However, I have a concern about the comparison with multimodal image editors.

The authors provide numerical measurements directly as text prompts. While this is a reasonable baseline, it may not represent the strongest way to use general-purpose image editors for measurement-aware garment editing.

Indeed, the synthetic training data demonstrates that Nano Banana 2 can generate visually plausible alterations in garment fit when instructed to change garment size qualitatively.

This suggests a natural alternative: translate numerical body and garment measurements into explicit, semantically meaningful fit instructions before passing them to the image editor.

For example, rather than providing a garment chest measurement of 110 cm and a body chest measurement of 100 cm, an LLM or a deterministic rule-based system could derive the garment ease and instruct the editor to produce a relatively loose chest fit while preserving other garment properties.

This baseline is especially relevant because the paper's training measurements are themselves predominantly estimated from visual appearance.

Suggested improvement:

Compare RealFIT against a two-stage baseline:

  1. Convert body and garment measurements into relative fit descriptions or region-specific editing instructions using an LLM or deterministic measurement-to-fit rules.
  2. Provide those instructions, together with the person and garment images, to Nano Banana 2 or Qwen-Image-Edit.

This would help establish whether RealFIT offers meaningful advantages over a general-purpose editor equipped with a measurement-aware prompt interface.

A related experiment would be to train a VTO model on the same synthetic data using relative measurement features or discrete fit categories rather than estimated absolute measurements.

Such comparisons would directly test whether absolute numerical conditioning contributes beyond the relative fit information already present in the synthetic examples.

4. Ablations Support Component Utility but Do Not Establish the Necessity of Specific Designs

The ablation study provides useful evidence that the individual components contribute to the system's performance. However, several conclusions appear stronger than what the experiments directly establish.

4.1 Ill-Fit Augmentation

Removing ill-fit augmentation substantially reduces Size Monotonicity, suggesting that the augmented data improves measurement sensitivity.

However, this comparison does not distinguish between two possible explanations:

  • The model benefits from greater training-data diversity and broader coverage of garment fits.
  • The model specifically benefits from having paired fit/ill-fit examples of the same person and garment.

A useful ablation would compare the proposed paired augmentation with a training set containing similarly diverse ill-fit examples without explicitly using the pairwise relationship, while controlling for the number and distribution of training examples.

This would help establish whether the benefit comes primarily from dataset diversity or counterfactual supervision.

4.2 Fourier Measurement Encoding

The comparison against numerical text prompts demonstrates that directly supplying measurement values through a text encoder is less effective than the proposed dedicated numerical encoder.

However, it does not establish that Fourier features are necessary or preferable to simpler numerical representations.

A straightforward baseline would be a standard MLP receiving z-score-normalized numerical measurements, using training-set means and standard deviations.

This would directly test whether Fourier encoding contributes beyond having an explicit numerical conditioning pathway.

4.3 Localized Metric Contrastive Loss

The ablation demonstrates that LMC improves upon ordinary flow-matching training.

However, the motivation for LMC is that garment-size changes occur primarily near garment boundaries and that the ordinary reconstruction objective is dominated by unchanged regions.

A much simpler alternative would be spatially weighted flow matching, assigning higher loss weights to regions near garment boundaries.

For example, garment masks could be used to identify boundary regions, and the ordinary flow-matching objective could be reweighted accordingly.

Another natural alternative is direct pairwise regression of the predicted velocity difference against the ground-truth velocity difference, using the same size-perturbed training pairs.

Comparing these alternatives would clarify whether LMC's improvement comes from its particular contrastive formulation or simply from placing additional supervision on regions affected by measurement changes.

To summarize, following further analysis would help:

  1. Validate their VLM-derived measurements against tape-measured ground truth.
  2. Evaluate the magnitude of fit changes using real paired-size examples, rather than relying on monotonicity alone.
  3. Compare against measurement-to-fit prompt rewriting for general-purpose image editors.
  4. Strengthen the ablations with simpler alternatives to Fourier encoding and LMC.
Questions for authors

Questions have been raised in the above sections.

2 comments
AuthorsOpenReview ↗

Response to Reviewer d5Vw

We thank Reviewer d5Vw for recognizing the value of RealFIT-Bench, LFS, and SM. Below we address each point concisely, noting where requested analyses are already reported in the paper and adding the suggested VLM prompt-rewriting baseline.

1. Physical Calibration & Evaluation Metrics (Weakness 1)

As explicitly stated in Sec. 1 and Sec. 4.4, LFS measures physical size accuracy against ground truth while SM measures monotonic size consistency; they are meant to be read together. Importantly, RealFIT-Bench does not just test one "standard" fit—over 64% of its unaugmented real photos are naturally tight or loose fits (1 to >2 sizes off; Sec. 3.2, Fig. 2c, App. C.4). Since our mask-free input I_p wears a different garment of a different length (Sec. 3.1) and the layflat I_g has no metric scale, getting 0.9476 LFS on 1,395 unseen test pairs is only possible if the model predicts physically accurate fit magnitudes from the input cm measurements across tight, normal, and oversized garments.

"Ideally, the evaluation should include real photographs of the same person wearing the same garment design in three or more physically measured sizes"

We agree, although it is logistically infeasible to collect data of the same person wearing the same garment in multiple sizes at a meaningful scale while maintaining the same pose and camera setting. We also argue that such data is not necessary for evaluation, since RealFIT predicts each try-on independently without referencing another size—combining LFS across diverse real fit ratios (above) with SM across size sweeps already tests both physical accuracy and relative size control.

"If possible, it would be more informative to independently vary individual garment measurements, rather than scaling all dimensions together, to determine whether the model can control distinct aspects of garment fit."

We already did this in Sec. 5.5 and App. A.3 (Fig. 9), where we vary length, bust, and sleeve length one at a time (±10%) while keeping all other measurements fixed, as well as varying only subject height H (Fig. 7). Each dimension moves only its corresponding garment boundary.

2. Reliability of Training Measurements & Synthetic Ill-Fit Supervision (Weakness 2)

"Given that the authors already have a tape-measured benchmark, RealFIT-Bench, it can be used to validate the training-data annotation procedure. For example, report: Absolute measurement errors between VLM-estimated measurements and tape-measured ground truth, broken down by measurement type."

We already report this exact validation in Sec. 5.5 ("LLM Annotation Accuracy"). Benchmarked against manual tape measurements on RealFIT-Bench, our two-call self-verified VLM pipeline (App. C.1) achieves near-zero mean bias and low median absolute errors across every dimension: length −0.89 cm (MedAE 3.0 cm), bust +2.10 cm (4.7 cm), waist +0.56 cm (2.5 cm), and sleeve −0.11 cm (2.25 cm)—well within manual tape-measuring variance.

"This raises the possibility that RealFIT primarily learns a numerically conditioned version of the qualitative fit-editing behavior exhibited by Nano Banana 2."

First, RealFIT is not trained solely on VLM or synthetic labels: pretraining incorporates published retailer size charts (Sec. 3.1), and the model is subsequently finetuned on RealFIT-Bench with real tape-measured garments and subjects (Sec. 3.2, Sec. 5.1), anchoring the model to physical ground truth. Second, on unseen, unaugmented real test photos in RealFIT-Bench, RealFIT outperforms Nano Banana 2 by a wide margin in both physical fit accuracy (0.9476 vs. 0.8876 LFS) and size monotonicity (90.54% vs. 70.33% SM, Table 1), showing that it learns calibrated metric fit rather than merely distilling Nano Banana 2.

AuthorsOpenReview ↗

Response to Reviewer d5Vw (continue)

3. Comparison with Multimodal Image Editors (Weakness 3)

"translate numerical body and garment measurements into explicit, semantically meaningful fit instructions before passing them to the image editor. For example, rather than providing a garment chest measurement of 110 cm and a body chest measurement of 100 cm..."

First, note that RealFIT and all the baselines do not take body chest or waist measurements as input, only wearer height H and the photo I_p (Sec. 4.1), so deterministic rules cannot compute garment ease; a VLM must infer body proportions visually from I_p. Second, qualitative prompts worked in our training augmentation (App. C.3) only because Nano Banana 2 was making coarse, uncalibrated edits ("1–2 sizes smaller/larger") to an already-worn ground-truth image I_gt, whose resulting dimensions we then had to measure post-hoc. At test time, qualitative words like "loose" are too coarse to specify exact cm boundaries from a layflat garment or resolve 5% size steps.

To verify this directly, we ran the suggested two-stage baseline on RealFIT-Bench—using a VLM to translate (I_p, H, and garment cm measurements) into explicit and detailed fit instructions for Nano Banana 2. While explicit fit descriptions improve Nano Banana 2's directional monotonicity over raw numerical prompts (SM: 70.33% → 79.73%), pointwise fit accuracy slightly degrades (LFS: 0.8876 → 0.8831), and both remain well below RealFIT (0.9476 LFS / 90.54% SM)—confirming that qualitative prompt rewriting cannot replace continuous metric conditioning.

4. Ablation Design Choices (Weakness 4)

"However, this comparison does not distinguish between two possible explanations: The model benefits from greater training-data diversity... (vs.) having paired fit/ill-fit examples of the same person and garment."

Table 1 already separates these two effects. Going from Ours w/o Ill-Fit Aug. (54.72% SM) to Ours w/o LMC Loss (84.86% SM) adds D_aug under standard single-sample flow matching—where each example is sampled independently without using the pair—isolating the gain from data diversity. Adding Ours (with LMC) on the exact same data then uses the pairwise relationship, further boosting SM from 84.86% to 90.54% (and LFS from 0.9406 to 0.9476).

"However, it does not establish that Fourier features are necessary or preferable to simpler numerical representations (such as a standard MLP)."

Sinusoidal/Fourier feature lifting before an MLP (App. B) is standard practice for continuous scalar conditioning across diffusion and generative models (e.g., DiT timestep/scalar embeddings) as of 2026; it should not require re-ablating from scratch. Our Ours w/ Text Meas. ablation (Table 1) already validates the actual design question: dedicated continuous numerical conditioning vs. text tokens.

"A much simpler alternative would be spatially weighted flow matching, assigning higher loss weights to regions near garment boundaries... Another natural alternative is direct pairwise regression of the predicted velocity difference against the ground-truth velocity difference..."

There are countless alternative loss formulations one could hypothesize, and retraining a 9B flow model across arbitrary design sub-variants is neither computationally feasible nor a reasonable bar for ablation studies. Our controlled ablation (Ours vs. Ours w/o LMC Loss, Table 1) already establishes LMC's clear benefit (+5.7% SM on RealFIT-Bench). In fact, we tested both suggested variants early in development and discarded them because they failed: spatially reweighting single-sample flow matching lacks cross-size contrast so the model still ignores m, and direct MSE regression on velocity differences was unstable and caused boundary artifacts under noisy VLM size labels (Sec. 4.3).

Reviewer_pcKjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

RealFIT studies measurement-conditioned virtual try-on using person and garment images, wearer height, and garment measurements. The paper introduces RealFIT-Web and RealFIT-Bench, with synthetic size-edited pairs added for training. The model builds on FLUX.2 with measurement encoding, mask-free conditioning, and an LMC loss to improve its response to size changes. The authors also propose two metrics for garment coverage accuracy and size monotonicity, and report improvements over existing methods on their benchmarks.

Critical strengths

1,The authors introduce RealFIT-Bench to tackle the lack of measurement-annotated data for fit-aware virtual try-on. Built via crowdsourcing with 168 participants and 987 garments, the dataset pairs flat-lay images and manual tape measurements with real on-body photographs.. This effort addresses a long-standing evaluation gap. 2,The proposed LFS and SM metrics are better aligned with fit-aware try-on than traditional metrics such as FID and SSIM,making the evaluation more informative.

Critical weaknesses

1,The model only takes height as an explicit body measurement and relies on the input photo for other proportions. Without shoulder width or body chest and waist measurements, the results may show plausible size changes without accurately reflecting fit. This leaves the reliability of its real-world fit predictions insufficiently demonstrated. 2,Cross-size supervision relies on image editing and VLM-based scale estimation; however, the paper has not sufficiently isolated or quantified these biases, nor has it directly demonstrated that fine-tuning can eliminate them.

Questions for authors

1,Section 3.2 mentions collecting self-measured body dimensions. Which measurements were collected, and what evidence supports using only height and the input photo for accurate fit prediction? 2,When collecting data on the bench, was any data captured of the same person wearing the same garment in different sizes? 3,In Figure 8, the white dress seems to get longer or shorter mainly at the skirt, while the waist section barely changes. Could this be because the size-editing process focuses on moving the hemline? How can we tell whether this reflects real fit changes or a bias in the training data?

1 comment
AuthorsOpenReview ↗

Response to Reviewer pcKj

We thank Reviewer pcKj for highlighting the value of RealFIT-Bench and our fit-aware evaluation metrics (LFS and SM). We address each concern and question concisely below.

1. Body Measurements Collected & Using Only Height + Input Photo (Weakness 1 & Question 1)

"The model only takes height as an explicit body measurement and relies on the input photo for other proportions... Section 3.2 mentions collecting self-measured body dimensions. Which measurements were collected, and what evidence supports using only height and the input photo for accurate fit prediction?"

In RealFIT-Bench, we collected self-measured height, weight, chest, waist, hip, shoulder width, arm length, and inseam for dataset metadata and QA, but deliberately conditioned RealFIT on only height H and the reference photo I_p (Sec. 4.1). First, from a product perspective, requiring shoppers to tape-measure their own shoulder width, chest, and waist introduces high friction and hurts user engagement, whereas nearly every user knows their height. Second, from a modeling perspective, relative body proportions (shoulder, chest, and waist widths) are already visible in I_p once H fixes the global pixel-to-cm scale. In fact, because large-scale VTO pretraining corpora (RealFIT-Web) lack ground-truth body measurements, conditioning on them would require estimating them from I_p and H via a VLM—which provides no additional information beyond the (I_p, H) inputs the model already receives, while only compounding estimation noise. Empirically, across 1,395 real test pairs in RealFIT-Bench (>64% naturally tight or loose), conditioning on H and I_p alone achieves 0.9476 LFS along with 0.8535 SSIM and 0.1088 LPIPS against the ground-truth photos I_gt (Table 1, Fig. 5), accurately matching both boundary extent and full-body silhouette drape.

2. Quantifying VLM/Editing Bias & Effect of Finetuning (Weakness 2)

"Cross-size supervision relies on image editing and VLM-based scale estimation; however, the paper has not sufficiently isolated or quantified these biases, nor has it directly demonstrated that fine-tuning can eliminate them."

First, Sec. 5.5 ("LLM Annotation Accuracy") directly isolates and quantifies VLM measurement error against manual tape measurements on RealFIT-Bench: population mean bias (and MedAE) is −0.89 cm (3.0 cm) for length, +2.10 cm (4.7 cm) for bust, +0.56 cm (2.5 cm) for waist, and −0.11 cm (2.25 cm) for sleeve, confirming near-zero systematic bias. Second, comparing RealFIT on RealFIT-Bench before vs. after finetuning on tape-measured RealFIT-Bench directly demonstrates the benefit of real-data finetuning: the RealFIT-Web pretrained model already achieves 0.9318 LFS, 86.90% SM, and 17.06 FID zero-shot on RealFIT-Bench, and finetuning further improves performance to 0.9476 LFS, 90.54% SM, and 13.69 FID (Table 1).

3. Multi-Size Capture in RealFIT-Bench (Question 2)

"When collecting data on the bench, was any data captured of the same person wearing the same garment in different sizes?"

As noted in Sec. 3.2, each garment in RealFIT-Bench comes from the participant's personal wardrobe and is observed in a single size. However, participants were required to select garments spanning five fit categories from very tight (>2 sizes smaller) to very loose (>2 sizes larger), so over 64% of the unaugmented real photos depict naturally tight or loose fits (Fig. 2c, App. C.4).

4. White Dress in Figure 8: Skirt Length vs. Waist Change (Question 3)

"In Figure 8, the white dress seems to get longer or shorter mainly at the skirt, while the waist section barely changes."

Zooming into row 6 of Fig. 8, the waist actually changes significantly: it is tightly cinched against the torso at −20%, and widens into a loose, straight silhouette with vertical drape folds by +20% (while the sleeves also widen and lengthen; see also App. A.3, Fig. 9 for independent torso-width control with the hemline fixed).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.