Pith. sign in

REVIEW 2 major objections 6 minor 76 references

CtrlVTON lets users control garment size, style, and placement in virtual try-on via segmentation masks, following those layouts far more faithfully than the strongest proprietary editors while matching them on garment fidelity.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 03:35 UTC pith:EGYJCHXR

load-bearing objection Solid systems paper: new instance-level VIP-Seg plus editing-based VTO with real mask control that beats proprietary editors on layout while matching fidelity; the quantitative SOTA is on natural silhouettes, but the core claim still holds. the 2 major comments →

arxiv 2607.09362 v1 pith:EGYJCHXR submitted 2026-07-10 cs.CV cs.AI

CtrlVTON: Controllable Virtual Try-On via Visual-Instance-Prompt Segmentation

Seungyong Lee , Hyun Jun Jang , Sangoh Kim , Sungjoon Park This is my paper
classification cs.CV cs.AI
keywords Virtual Try-OnControllable Image GenerationVisual-Prompt SegmentationInstance SegmentationDiffusion TransformersImage EditingMask Conditioning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Most virtual try-on systems can place a garment on a person, but they give almost no say over how it is worn—loose or fitted, tucked or open, layered or swapped. This paper argues that gap closes with two linked advances. First, VIP-SAM solves visual-instance-prompt segmentation: given a flatlay of one specific garment, it finds that exact instance on a person even amid similar clothes, occlusion, and deformation. Second, CtrlVTON recasts try-on as full-image editing rather than inpainting, then conditions on those masks so users get pixel-level layout control. The same model unifies full swap, selective swap, layering, and multi-garment outfits. On a new public benchmark it beats proprietary editors on mask adherence while remaining competitive on how faithfully the garment itself looks.

Core claim

The paper claims that virtual try-on becomes spatially controllable when instance-level garment masks are available and try-on is cast as editing over (reference person, reference garment, target) triplets: user-provided masks then serve as reliable pixel-level controls for style, size, and placement, avoiding the mask-shape and identity-leak failures of inpainting while matching top systems on garment fidelity.

What carries the argument

VIP-SAM injects support-garment features into a SAM or SAM2 encoder via early cross-attention adapters so the query is conditioned on the specific instance from the start; CtrlVTON freezes an editing diffusion transformer and trains a LoRA that channel-wise concatenates the three garment masks with their image latents, preserving spatial alignment without growing token count.

Load-bearing premise

Training depends on synthetic reference-person images made by off-the-shelf generators and then filtered by a vision-language model, silhouette checks, and human reviewers; if those synthetics systematically miss real poses, lighting, or silhouettes, both quality and mask following will degrade in real use.

What would settle it

If, on VITON-HD-edit or a held-out set of real user masks, proprietary editors match or beat CtrlVTON on IoU, Hu-moment distance, and Hausdorff distance while staying comparable on garment-fidelity scores—or if re-extracting masks with an independent segmenter collapses the adherence gap—the claim of superior spatial control would fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Users can restyle fit, tuck depth, sleeve length, zipper height, or layering by editing a mask instead of rewriting text.
  • One checkpoint unifies full swap, selective single-item swap, add/layer, and multi-garment composition via task tokens plus color-coded masks.
  • The released VITON-HD-edit benchmark becomes a shared testbed for editing-based and mask-controllable virtual try-on.
  • Visual-instance-prompt segmentation is established as a distinct task from category-level few-shot segmentation, with a fashion-domain dataset and SOTA baselines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Early support-feature injection may transfer to other instance re-identification settings, such as matching a product photo to that item in crowded retail or warehouse video.
  • If synthetic reference-person quality is the main bottleneck, improving the contour filter or the generator stage could raise both fidelity and mask adherence without changing the model architecture.
  • Mask-as-control is a practical retail interface where text prompts cannot express precise fit and layering; the interactive base-then-edit workflow the paper sketches is a natural product path.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper defines Visual-Instance-Prompt Segmentation (VIP-Seg)—segmenting a specific garment instance in a person image given a flatlay/support image—and solves it with VIP-SAM via early cross-attention injection of support features into SAM/SAM2 encoders. It then presents CtrlVTON, which recasts virtual try-on as full-image editing (pref, gref)→p rather than inpainting, and adds pixel-level layout control by channel-wise concatenating garment masks (Mp, Mpref, Mgref) through a LoRA adapter on a frozen DiT editing backbone, with discrete garment-class and task tokens (full_swap / partial_swap / add). Training uses a synthetic-triplet pipeline with VIP-SAM masks, multi-stage QC (VLM, CMF silhouette filter, human review), and releases VITON-HD-edit. Empirically, VIP-SAM is SOTA on a fashion VIP-Seg split and on same-class COCO-20i / PASCAL-5i; CtrlVTON-base is competitive or better on single- and multi-garment VTO; CtrlVTON reports substantially higher mask adherence (IoU 0.961 vs ~0.87) than strong proprietary editors on VITON-HD-edit while matching garment fidelity.

Significance. If the results hold, the work closes a clear practical gap: most VTO systems lack precise control over fit, style (tuck, open/closed, layering), and placement. Unifying swap, selective swap, layering, and multi-garment try-on under one editing model with mask control is valuable for e-commerce and interactive styling. Strengths include a well-motivated instance-level segmentation task, a public editing/controllability benchmark (VITON-HD-edit), multi-family metrics (embedding fidelity, VLM-as-judge, IoU/Hu/Hausdorff), an explicit channel- vs token-wise injection comparison, and a SAM3 re-evaluation that preserves ranking and mitigates self-reference through VIP-SAM. The combination of open baselines, proprietary editors, and released data/code intent makes the contribution usable by the community.

major comments (2)
  1. §5.4 and Table 4: The headline claim that CtrlVTON follows user-provided layouts far more faithfully than proprietary editors (IoU 0.961 vs ~0.87, lower dHu and dH) is quantified on VITON-HD-edit control masks Mp that are natural garment regions from the same VIP-SAM-driven synthetic pipeline used in training (Sec. 4.2, Supp. C). Fine-grained style/fit control (tucking, open/closed, loose/fitted, layering offsets) is shown only qualitatively with author-drawn masks (Figs. 9–10, 18–19). Supp. F.5’s SAM3 re-run rules out pure segmenter self-reference but still uses natural silhouettes. Please add a quantitative mask-adherence evaluation on deliberately deformed or hand-edited masks (or a held-out set of user-style edits), or narrow the abstract/§5.4 claim so that the large gap is scoped to natural layout masks rather than the full advertised interactive-styling regime.
  2. Sec. 4.2 and Supp. C.2–C.3: Training critically depends on synthetic pref from off-the-shelf inpainting/editing models, filtered by VLM checks, CMF silhouette-leakage threshold τ, and three annotators. The paper does not report sensitivity of final VTO or mask-adherence metrics to τ, candidate count, or masking strategy mix (box/loose/tight). Because silhouette leakage is explicitly identified as a failure mode that teaches the model to ignore gref (Fig. 17, Supp. C.3), a short ablation or failure-rate breakdown of the QC funnel would strengthen the claim that the learned editing behavior generalizes beyond the synthetic distribution.
minor comments (6)
  1. Fig. 2 caption and §3: Clarify whether support encoders (ResNet-50 / DINOv2 / DINOv3) are frozen or fine-tuned, and whether adapters are inserted after every block for both ViT and Hiera variants, so the architecture is fully reproducible from the main text.
  2. Table 1: The protocol change (train and evaluate on the same categories rather than held-out classes) is appropriate for VIP-Seg but should be stated more prominently in the table caption so readers do not misread the numbers as standard few-shot novel-class scores.
  3. Eq. (1)–(2) and §4.3: State the LoRA rank, which linear projections receive adapters, and the learning rate / step budget for the LoRA stage relative to the base fine-tune (Supp. D currently focuses on VIP-SAM and base compute).
  4. §5.2 Metrics: Briefly note that unpaired evaluation precludes SSIM/LPIPS and that FID is de-emphasized; this is already argued but a one-sentence pointer in the main metrics paragraph would help non-VTO readers.
  5. Fig. 11 and Supp. B: The inpainting failure examples are persuasive; consider citing them once in the main §4.1 so readers who skip the supplement still see the motivation for the editing formulation.
  6. Typographical consistency: task tokens appear as full_sw ap / partial_sw ap in places (spaces in “swap”); normalize spelling throughout text, figures, and tables.

Circularity Check

1 steps flagged

No derivation-by-construction circularity; only a mild, author-mitigated train/eval segmenter self-reference that does not force the SOTA claims.

specific steps
  1. other [Sec. 4.2 (masks Mp/Mpref); Sec. 5.2 metrics (a)(c); Supp. F.1, F.4–F.5]
    "Mp is the mask of gref in p, obtained by querying VIP-SAM (Sec. 3) with gref as the visual prompt. ... For each generated try-on image we crop the garment region using its VIP-SAM mask ... we re-extract the garment mask Mgen via VIP-SAM ... To rule this out, we re-run inference-time mask extraction and metric computation on VITON-HD-edit using an independent segmenter, SAM3 ... the ranking is unchanged: CtrlVTON remains far ahead"

    VIP-SAM is used both to construct the training control masks and to extract evaluation crops/Mgen, creating a mild self-referential evaluation loop. This is not definitional circularity (IoU is not forced if the model fails to follow Mp), and the authors break the loop with SAM3; residual risk is distributional (natural VIP-SAM silhouettes), not a result-by-construction.

full rationale

CtrlVTON is an empirical systems paper (architecture + data pipeline + benchmarks), not a first-principles derivation. There is no equation that defines a quantity in terms of itself and then reports it as a prediction; no fitted parameter is renamed as an out-of-sample forecast; no uniqueness theorem or ansatz is imported from overlapping-author prior work to force the result. VIP-SAM and CtrlVTON are trained and scored against external public benchmarks (VITON-HD, OmniTry Bench, DressCode-MR, Garments2Look, COCO-20i, PASCAL-5i) and proprietary editors, with standard flow-matching / segmentation objectives. The only self-reference is that VIP-SAM supplies both training masks (Mp, Mpref, Mgref) and evaluation crops/masks (garment fidelity and mask-adherence). That is a methodological risk of self-referential evaluation, not a reduction of the claimed layout-following gap to the inputs by construction: if the editor ignored Mp, re-extracted Mgen would not match. The authors explicitly re-run Tab. 4 metrics with independent text-prompted SAM3 (Supp. F.5); rankings and absolute gaps are essentially unchanged, so the dual use is not load-bearing. Remaining concerns (natural vs. hand-edited mask distribution) are generalization/correctness issues, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard diffusion/flow-matching machinery, the assumption that synthetic pref triplets are distributionally adequate after filtering, and the architectural choice of early feature injection plus channel-wise mask LoRA. No new physical entities are postulated; free parameters are ordinary training and filtering hyperparameters.

free parameters (3)
  • CMF silhouette-leakage threshold τ and discard rule
    Hand-chosen contour-match fraction used to reject synthetic pref candidates whose silhouette is too similar to the original garment; directly affects training distribution diversity (Supp. C.3).
  • LoRA rank / placement and flow-matching training schedule
    Adapter capacity and optimization hyperparameters for the mask-conditioning extension; not derived from first principles.
  • Number of synthetic candidates per source pair (4–5) and three-stage QC weights
    Pipeline design choices that determine which pref images enter the final training set.
axioms (4)
  • domain assumption Synthetic pref images generated by off-the-shelf inpainting/editing models, after VLM + CMF + human filtering, are sufficiently close to the real (person, different-garment) distribution for the editing model to learn identity-preserving garment transfer.
    Load-bearing for the entire editing formulation (Sec. 4.1–4.2); never independently validated against real multi-outfit captures of the same person.
  • ad hoc to paper Early cross-attention injection of support features into SAM/SAM2 encoders is sufficient for instance-level discrimination under same-category distractors and non-rigid deformation.
    Core architectural hypothesis of VIP-SAM (Sec. 3, Fig. 2); supported empirically but not derived.
  • standard math Channel-wise concatenation of VAE-encoded masks with image latents preserves spatial alignment without increasing token count or attention cost.
    Standard latent-space conditioning trick; empirically compared to token-wise alternative (Supp. E).
  • domain assumption Flow-matching DiT velocity field with discrete garment-class and task tokens can express full_swap / partial_swap / add operations.
    Assumed by the base-model design (Sec. 4.3); validated by TFC scores but not proven complete.
invented entities (2)
  • VIP-Seg task definition independent evidence
    purpose: Formalizes instance-level visual-prompt segmentation required for garment-specific masks under same-category distractors.
    New task statement distinct from category-level VRP-Seg; independent evidence is the released fashion benchmark and repurposed COCO/PASCAL numbers.
  • VITON-HD-edit benchmark independent evidence
    purpose: Public testbed for editing-based VTO, mask-controllable VTO, and VIP-Seg.
    Constructed by applying the authors’ pipeline to the VITON-HD test set; enables external comparison.

pith-pipeline@v1.1.0-grok45 · 29757 in / 3419 out tokens · 50355 ms · 2026-07-13T03:35:45.751695+00:00 · methodology

0 comments
read the original abstract

Virtual try-on (VTO) has made significant progress in realistically transferring garments onto a target person. Yet most systems give the user little control over how a garment should be worn -- its size (loose or fitted), style (e.g., tucked in or untucked, open or closed), and spatial placement on the body. We address this gap with two complementary contributions. First, we define and solve Visual-Instance-Prompt Segmentation via VIP-SAM: given a flatlay image of a garment, segment that specific instance in a photograph of a person wearing it. This is an instance-level task, distinct from the typically studied category-level segmentation. Second, we introduce CtrlVTON, a controllable VTO framework that recasts try-on as an image editing problem and adds segmentation masks as pixel-level control over garment layout, including style, size, and spatial placement on the body. VIP-SAM and CtrlVTON each achieve state-of-the-art results on their respective tasks. In particular, CtrlVTON generates images that follow user-provided layouts far more faithfully than the strongest proprietary editing systems while matching them on garment fidelity.

Figures

Figures reproduced from arXiv: 2607.09362 by Hyun Jun Jang, Sangoh Kim, Seungyong Lee, Sungjoon Park.

Figure 1
Figure 1. Figure 1: (a) CtrlVTON-base is a baseline image-editing model that enables semantic control via task tokens (full_swap / partial_swap / add) over multiple garment classes. (b) CtrlVTON enables fine-grained spatial control through hand-drawn masks (yellow), supporting both single- and multi-garment try-on. Abstract. Virtual try-on (VTO) has made significant progress in realistically transferring garments onto a targe… view at source ↗
Figure 2
Figure 2. Figure 2: Segmentation model architectures. (a) VRP-SAM trains extra modules to extract a visual prompt from support and query images; this prompt is fed to SAM’s mask decoder in place of the usual spatial prompts (point, box, mask) (b, c). VIP-SAM with the SAM (ViT) encoder in (b) and the SAM2 (Hiera) encoder in (c). In each variant, adapters inject support-image features into the encoder after every block. The out… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of training samples. We illustrate both single-garment and multi-garment data as well as the notation used throughout this work. Top: target person p, synthetic reference person pref, and the reference garment(s) gref (or g (1..K) ref in the multi-garment case). Bottom: the corresponding masks Mp, Mpref , and Mgref provide spatial control. In the multi-garment setting, each garment mask M (k) gref… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of CtrlVTON. For the base model, we fine-tune an image-editing DiT backbone on triplets (p, pref, gref). The controllability extension (red-dashed region) takes as input the masks Mp, Mpref , and Mgref , which are channel-wise concatenated with the tokens of the corresponding images. It is implemented as a LoRA adapter on top of the frozen base model. Quality control. For each source pair, we synt… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of visual-reference segmentation methods. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the mask-adherence evaluation. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on single-garment VTO. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Effect of the task token on CtrlVTON-base results. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of mask-controllable try-on against proprietary editing models. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Fine-grained control via mask conditioning. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Failure modes of inpainting-based VTO. Each example shows the input person, the masked person fed to the inpainting model, the reference garment, and the inpainted output. Red: an undersized mask leaves residual pixels of the original garment (blue jacket sleeve); the inpainting model treats them as context and renders the new garment incorrectly. Green: enlarging the mask to avoid the above issue erases … view at source ↗
Figure 12
Figure 12. Figure 12: Examples illustrating the contour match fraction (CMF) filter and its advantage over IoU. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative results of task-token control on OmniTry-Bench. [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: CtrlVTON generalizes across segmentation models. [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional single-garment VTO results. Comparisons across diverse garment types, person images, and reference representations, complementing [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Additional multi-garment VTO results. Each row shows multiple reference garments tried on a sin￾gle person, comparing CtrlVTON-base (which composes all garments in a single forward pass) with multi-garment baselines. FastFit and BootComp natively accept multiple garments, while FitDit and OmniTry are run sequentially, applying one garment at a time [PITH_FULL_IMAGE:figures/full_fig_p029_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Effect of data curation on try-on quality. [PITH_FULL_IMAGE:figures/full_fig_p030_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Additional mask-controllable try-on results. [PITH_FULL_IMAGE:figures/full_fig_p031_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Visualization of mask-following metrics. [PITH_FULL_IMAGE:figures/full_fig_p032_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Representative failure cases of our two models. (a) CtrlVTON-base [PITH_FULL_IMAGE:figures/full_fig_p033_20.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 12 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Black Forest Labs: FLUX.https://github.com/black-forest-labs/flux(2024)

    2024

  3. [3]

    Black Forest Labs, Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Müller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space (2025),http...

    Pith/arXiv arXiv 2025

  4. [4]

    arXiv preprint arXiv:2509.20427 (2025)

    ByteDance Seed Team: Seedream 4.0: Toward next-generation multimodal image generation. arXiv preprint arXiv:2509.20427 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    arXiv preprint arXiv:2511.16719 (2025)

    Carion, N., Gustafson, L., Hu, Y.T., et al.: SAM 3: Segment anything with concepts. arXiv preprint arXiv:2511.16719 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    In: ICCV (2021)

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)

    2021

  7. [7]

    arXiv preprint arXiv:2604.19748 (2026)

    Chen, M., Chen, Z., Du, Y., Gao, Z., Hu, T., Lan, J., Lin, C., Shen, Y., Wang, X., Wang, Z., et al.: Tstars-Tryon 1.0: Robust and realistic virtual try-on for diverse fashion items. arXiv preprint arXiv:2604.19748 (2026)

    Pith/arXiv arXiv 2026

  8. [8]

    In: CVPR (2024)

    Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: AnyDoor: Zero-shot object-level image customization. In: CVPR (2024)

    2024

  9. [9]

    In: CVPR (2021)

    Choi, S., Park, S., Kang, M., Choo, J.: VITON-HD: High-resolution virtual try-on via misalignment-aware nor- malization. In: CVPR (2021)

    2021

  10. [10]

    In: ECCV (2024)

    Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for authentic virtual try-on in the wild. In: ECCV (2024)

    2024

  11. [11]

    In: CVPR (2025)

    Choi, Y., Kwak, S., Yu, S., Choi, H., Shin, J.: Controllable human image generation with personalized multi- garments. In: CVPR (2025)

    2025

  12. [12]

    In: ICLR (2025)

    Chong, Z., Dong, X., Li, H., Zhang, S., Zhang, W., Zhang, X., Zhao, H., Jiang, D., Liang, X.: CatVTON: Concatenation is all you need for virtual try-on with diffusion models. In: ICLR (2025)

    2025

  13. [13]

    arXiv preprint arXiv:2508.20586 (2025)

    Chong, Z., Lei, Y., Zhang, S., He, Z., Wang, Z., Zhang, X., Dong, X., Wu, Y., Jiang, D., Liang, X.: FastFit: Accelerating multi-reference virtual try-on via cacheable diffusion models. arXiv preprint arXiv:2508.20586 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    arXiv preprint arXiv:2508.08488 (2025)

    Deria, A., et al.: MuGa-VTON: Multi-garment virtual try-on via diffusion transformers with prompt customiza- tion. arXiv preprint arXiv:2508.08488 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    In: NeurIPS (2025)

    Feng, Y., Zhang, L., Cao, H., Chen, Y., Feng, X., Cao, J., Wu, Y., Wang, B.: OmniTry: Virtual try-on anything without masks. In: NeurIPS (2025)

    2025

  16. [16]

    Google DeepMind model page (2025),https://deepmind.google/models/ gemini/flash/

    Google DeepMind: Gemini 3 Flash. Google DeepMind model page (2025),https://deepmind.google/models/ gemini/flash/

    2025

  17. [17]

    Google blog post (2025),https://blog

    Google DeepMind: Introducing Nano Banana Pro: Gemini 3 Pro Image. Google blog post (2025),https://blog. google/technology/ai/nano-banana-pro/

    2025

  18. [18]

    In: CVPR (2018)

    Güler, R.A., Neverova, N., Kokkinos, I.: DensePose: Dense human pose estimation in the wild. In: CVPR (2018)

    2018

  19. [19]

    In: ICCV (2025)

    Guo, H., Zeng, B., Song, Y., Zhang, W., Zhang, C., Liu, J.: Any2AnyTryon: Leveraging adaptive position em- beddings for versatile virtual clothing tasks. In: ICCV (2025)

    2025

  20. [20]

    In: CVPR (2019)

    Gupta, A., Dollár, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)

    2019

  21. [21]

    In: CVPR (2018)

    Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: VITON: An image-based virtual try-on network. In: CVPR (2018)

    2018

  22. [22]

    In: CVPR (2016)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)

    2016

  23. [23]

    In: NeurIPS (2017)

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)

    2017

  24. [24]

    arXiv preprint arXiv:2603.14153 (2026)

    Hu, J., Cheng, Z., Wong, W., Zou, X.: Garments2look: A multi-reference dataset for high-fidelity outfit-level virtual try-on with clothing and accessories. arXiv preprint arXiv:2603.14153 (2026)

    arXiv 2026

  25. [25]

    IRE Trans

    Hu, M.K.: Visual pattern recognition by moment invariants. IRE Trans. Inf. Theory8(2), 179–187 (1962)

    1962

  26. [26]

    arXiv preprint arXiv:2410.23775 (2024)

    Huang, L., Wang, W., Wu, Z.F., Shi, Y., Dou, H., Liang, C., Feng, Y., Liu, Y., Zhou, J.: In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    IEEE TPAMI15(9), 850–863 (1993)

    Huttenlocher, D.P., Klanderman, G.A., Rucklidge, W.J.: Comparing images using the Hausdorff distance. IEEE TPAMI15(9), 850–863 (1993)

    1993

  28. [28]

    Jiang, B., Hu, X., Luo, D., He, Q., Xu, C., Peng, J., Zhang, J., Wang, C., Wu, Y., Fu, Y.: FitDiT: Advancing the authentic garment details for high-fidelity virtual try-on (2024),https://arxiv.org/abs/2411.10499

    Pith/arXiv arXiv 2024

  29. [29]

    In: ECCV (2024)

    Ju, X., Liu, X., Wang, X., Bian, Y., Shan, Y., Xu, Q.: BrushNet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In: ECCV (2024)

    2024

  30. [30]

    In: ECCV (2024)

    Khirodkar, R., Bagautdinov, T., Martinez, J., Zhaoen, S., James, A., Selednik, P., Anderson, S., Saito, S.: Sapiens: Foundation for human vision models. In: ECCV (2024)

    2024

  31. [31]

    In: CVPR (2024) CtrlVTON 15

    Kim, J., Gu, G., Park, M., Park, S., Choo, J.: StableVITON: Learning semantic correspondence with latent diffusion model for virtual try-on. In: CVPR (2024) CtrlVTON 15

    2024

  32. [32]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kim, J., Jin, H., Park, S., Choo, J.: Promptdresser: Improving the quality and controllability of virtual try-on via generative textual prompt and prompt-aware mask. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16026–16036 (2025)

    2025

  33. [33]

    arXiv preprint arXiv:2602.17636 (2026)

    Kim, J., Shin, Y., Jin, S., Chung, D., Nam, J., Kim, T., Park, J., Kang, H., Kim, S.: Coral: Correspondence alignment for improved virtual try-on. arXiv preprint arXiv:2602.17636 (2026)

    arXiv 2026

  34. [34]

    In: ICCV (2023)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023)

    2023

  35. [35]

    Labs, B.F.: FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2(2025)

    2025

  36. [36]

    In: ECCV (2022)

    Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion- handled conditions. In: ECCV (2022)

    2022

  37. [37]

    In: ACM SIGGRAPH Asia (2025)

    Lee, S., gi Kwak, J.: Voost: A unified and scalable diffusion transformer for bidirectional virtual try-on and try-off. In: ACM SIGGRAPH Asia (2025)

    2025

  38. [38]

    In: Computer Graphics Forum

    Levin, E., Fried, O.: Differential diffusion: Giving each pixel its strength. In: Computer Graphics Forum. vol. 44. Wiley Online Library (2025)

    2025

  39. [39]

    arXiv preprint arXiv:2601.22725 (2026)

    Li,J.,Chen,T.,Jiang,S.,Wang,W.,Luo,J.,Wu,C.:OpenVTON-Bench:Alarge-scalehigh-resolutionbenchmark for controllable virtual try-on evaluation. arXiv preprint arXiv:2601.22725 (2026)

    Pith/arXiv arXiv 2026

  40. [40]

    In: CVPR (2023)

    Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Open-set grounded text-to-image generation. In: CVPR (2023)

    2023

  41. [41]

    In: ICLR (2024)

    Liu, Y., Zhu, M., Li, H., Chen, H., Wang, X., Shen, C.: Matcher: Segment anything with one shot using all-purpose feature matching. In: ICLR (2024)

    2024

  42. [42]

    arXiv preprint arXiv:2501.06230 (2025)

    Meyer, M., Spruyt, J.: Ben: Using confidence-guided matting for dichotomous image segmentation. arXiv preprint arXiv:2501.06230 (2025)

    arXiv 2025

  43. [43]

    In: ACM MM (2023)

    Morelli, D., Baldrati, A., Cartella, G., Cornia, M., Bertini, M., Cucchiara, R.: LaDI-VTON: Latent diffusion textual-inversion enhanced virtual try-on. In: ACM MM (2023)

    2023

  44. [44]

    In: ICCV (2019)

    Nguyen, K., Todorovic, S.: Feature weighting and boosting for few-shot segmentation. In: ICCV (2019)

    2019

  45. [45]

    OpenAI: The new ChatGPT Images is here.https://openai.com/index/new-chatgpt-images-is-here/(2025)

    2025

  46. [46]

    TMLR (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual feat...

    2024

  47. [47]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  48. [48]

    In: ICLR (2025)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: SAM 2: Segment anything in images and videos. In: ICLR (2025)

    2025

  49. [49]

    In: CVPR (2022)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)

    2022

  50. [50]

    In: BMVC (2017)

    Shaban, A., Bansal, S., Liu, Z., Essa, I., Boots, B.: One-shot learning for semantic segmentation. In: BMVC (2017)

    2017

  51. [51]

    In: CVPR (2024)

    Sheynin, S., Polyak, A., Singer, U., Kirstain, Y., Zohar, A., Ashual, O., Parikh, D., Taigman, Y.: Emu Edit: Precise image editing via recognition and generation tasks. In: CVPR (2024)

    2024

  52. [52]

    In: CVPR (2024)

    Shi, Y., Xue, C., Liew, J.H., Pan, J., Yan, H., Zhang, W., Tan, V.Y.F., Bai, S.: DragDiffusion: Harnessing diffusion models for interactive point-based image editing. In: CVPR (2024)

    2024

  53. [53]

    arXiv preprint arXiv:2508.10104 (2025)

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: DINOv3. arXiv preprint a...

    Pith/arXiv arXiv 2025

  54. [54]

    In: CVPR (2024)

    Sun, Y., Chen, J., Zhang, S., Zhang, X., Chen, Q., Zhang, G., Ding, E., Wang, J., Li, Z.: VRP-SAM: SAM with visual reference prompt. In: CVPR (2024)

    2024

  55. [55]

    In: ECCV (2018)

    Wang, B., Zheng, H., Liang, X., Chen, Y., Lin, L., Yang, M.: Toward characteristic-preserving image-based virtual try-on network. In: ECCV (2018)

    2018

  56. [56]

    In: ICCV (2025)

    Wang, X., Sebastian, C., He, W., Ren, L.: Prosam: Enhancing the robustness of sam-based visual reference segmentation with probabilistic prompts. In: ICCV (2025)

    2025

  57. [57]

    In: ICLR (2025)

    Wang, X., Fu, S., Huang, Q., He, W., Jiang, H.: MS-Diffusion: Multi-subject zero-shot image personalization with layout guidance. In: ICLR (2025)

    2025

  58. [58]

    In: CVPR (2023)

    Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR (2023)

    2023

  59. [59]

    In: ICCV (2023) 16 S

    Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: SegGPT: Towards segmenting everything in context. In: ICCV (2023) 16 S. Lee et al

    2023

  60. [60]

    In: CVPR (2024)

    Wang, X., Darrell, T., Rambhatla, S.S., Girdhar, R., Misra, I.: InstanceDiffusion: Instance-level control for image generation. In: CVPR (2024)

    2024

  61. [61]

    IEEE Transactions on Image Processing13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing13(4), 600–612 (2004)

    2004

  62. [62]

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., L...

    Pith/arXiv arXiv 2025

  63. [63]

    In: CVPR (2023)

    Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: SmartBrush: Text and shape guided object inpainting with diffusion model. In: CVPR (2023)

    2023

  64. [64]

    In: CVPR (2023)

    Xie, Z., Huang, Z., Dong, X., Zhao, F., Dong, H., Zhang, X., Zhu, F., Liang, X.: GP-VTON: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In: CVPR (2023)

    2023

  65. [65]

    In: CVPR (2023)

    Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR (2023)

    2023

  66. [66]

    In: CVPR (2020)

    Yang, H., Zhang, R., Guo, X., Liu, W., Zuo, W., Luo, P.: Towards photo-realistic virtual try-on by adaptively generating↔preserving image content. In: CVPR (2020)

    2020

  67. [67]

    arXiv preprint arXiv:2512.24016 (2025)

    Yang, L., Liu, Y., Li, Y., Bai, X., Lu, H.: FitControler: Toward fit-aware virtual try-on. arXiv preprint arXiv:2512.24016 (2025)

    arXiv 2025

  68. [68]

    In: ICCV (2023)

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)

    2023

  69. [69]

    In: ICLR (2024)

    Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Qiao, Y., Gao, P., Li, H.: Personalize segment anything model with one shot. In: ICLR (2024)

    2024

  70. [70]

    In: CVPR (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)

    2018

  71. [71]

    TMLR (2025)

    Zheng, C., Lan, Y., Wang, Y.: Lanpaint: Training-free diffusion inpainting with asymptotically exact and fast conditional sampling. TMLR (2025)

    2025

  72. [72]

    In: CVPR (2025)

    Zhou, Z., Liu, S., Han, X., Liu, H., Ng, K.W., Xie, T., Cong, Y., Li, H., Xu, M., Pérez-Rua, J.M., Patel, A., Xiang, T., Shi, M., He, S.: Learning flow fields in attention for controllable person image generation. In: CVPR (2025)

    2025

  73. [73]

    In: CVPR (2024)

    Zhu, L., Li, Y., Liu, N., Peng, H., Yang, D., Kemelmacher-Shlizerman, I.: M&M VTO: Multi-garment virtual try-on and editing. In: CVPR (2024)

    2024

  74. [74]

    In: CVPR (2023)

    Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Saharia, C., Norouzi, M., Kemelmacher-Shlizerman, I.: TryOn- Diffusion: A tale of two UNets. In: CVPR (2023)

    2023

  75. [75]

    no perceivable issues

    Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In: ECCV (2024) CtrlVTON 17 Supplementary Material A VIP-SAM: Resource Comparison We complement the segmentation accuracy results in Tab. 1 of the main paper with a comparison of the resource footprint of eac...

    2024

  76. [76]

    (8) Lower is better. 26 S. Lee et al. Symmetric Hausdorff distanced H.Hu moments are useful for describing global shape but are insensitive tolocalizedfailures such as a sleeve protruding well beyond the requested mask. Following [27,67], we additionally report the symmetric Hausdorff distance between contours defined by point setsA=∂Mp and B=∂M gen: dH(A...