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:
- A VLM estimates absolute garment measurements from images, first estimating the person's height and then inferring garment dimensions from visual proportions.
- 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:
- Convert body and garment measurements into relative fit descriptions or region-specific editing instructions using an LLM or deterministic measurement-to-fit rules.
- 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:
- Validate their VLM-derived measurements against tape-measured ground truth.
- Evaluate the magnitude of fit changes using real paired-size examples, rather than relying on monotonicity alone.
- Compare against measurement-to-fit prompt rewriting for general-purpose image editors.
- Strengthen the ablations with simpler alternatives to Fourier encoding and LMC.
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.