Virtual Try-on (VTON) has become a core capability for online retail, where realistic try-on results provide reliable fit guidance, reduce returns, and benefit both consumers and merchants. Diffusion-based VTON methods achieve photorealistic synthesis, yet often rely on intricate architectures such as auxiliary reference networks and suffer from slow sampling, making the trade-off between fidelity and efficiency a persistent challenge. We approach VTON as a structured image editing problem that demands strong conditional generation under three key requirements: subject preservation, faithful texture transfer, and seamless harmonization. Under this perspective, our training framework is generic and transfers to broader image editing tasks. Moreover, the paired data produced by VTON constitutes a rich supervisory resource for training general-purpose editors. We present PROMO, a promptable virtual try-on framework built upon a Flow Matching DiT backbone with latent multi-modal conditional concatenation. By leveraging conditioning efficiency and self-reference mechanisms, our approach substantially reduces inference overhead. On standard benchmarks, PROMO surpasses both prior VTON methods and general image editing models in visual fidelity while delivering a competitive balance between quality and speed. These results demonstrate that flow-matching transformers, coupled with latent multi-modal conditioning and self-reference acceleration, offer an effective and training-efficient solution for high-quality virtual try-on.
Figures & tables
Figure 2: Overview of scenarios handled by our PROMO framework. The left panel illustrates three common scenarios that customers encounter during online shopping: 1) model image available without a specific garment image, 2) garment image available without a model, and 3) both model and garment images available. Our system addresses the missing information in each scenario by generating the necessary conditional inputs required by our model. Notably, for scenarios without model images, our system features pure image-reference capability, allowing text prompts to be omitted while maintaining a robust baseline performance.
Figure 3: Compared to directly using the original DensePose [ 11 ] , our method better estimates plausible body shapes under loose clothing, effectively preventing information leakage.
Figure 4: We downsample the extracted human parsing to match the resolution of the latent space. Compared to the standard supervision in (a), we adopt a weighted loss design in (b).
Figure 5: Our method for efficient spatial condition injection. We directly paste the pose condition onto the agnostic image, then perform downsampling, eventually to reduce the 2N tokens to N/4 tokens, achieving 87.5% token reduction.
Figure 6: Overview of the Style Prompt System . A finetuned VLM processes an image of a person wearing the source garment, extracting person and garment attributes into a JSON format. This structured JSON is then converted into a natural language prompt for our pipeline.
Figure 7: Attention mechanism: zt and Tstyle have global visibility, while Ci only attends to itself.
Figure 8: Temporal Self-Reference mechanism.
Figure 9: Visualization of comparison between ours and baseline models on DressCode dataset. Best viewed when zoomed in.
Figure 10: Visualization of comparison on virtual try-on between our PROMO between state-of-the-art image-editing model on DressCode dataset. Best viewed when zoomed in.
Methods
DressCode
VITON-HD
Paired
Unpaired
Paired
Unpaired
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
FID ↓
KID ↓
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
FID ↓
KID ↓
LaDI-VTON [ 28 ]
0.7564
0.3800
5.4700
1.9270
7.3475
2.4690
0.8721
0.1531
6.8500
1.3766
9.6131
2.0207
CatVTON [ 6 ]
0.8944
0.1600
6.5372
3.9591
8.4567
4.4897
0.8673
0.1882
9.4412
4.7389
12.5716
5.4555
OOTDiffusion [ 47 ]
0.8883
0.0800
3.6623
0.8550
7.0463
2.7910
0.7919
0.1905
32.8944
20.0843
39.9626
26.0686
Any2AnyTryon [ 12 ]
0.9107
0.1208
3.0828
1.0565
5.5404
1.5258
0.8705
0.1569
7.1238
2.1797
10.3120
2.7086
Table 1: Quantitative results on VITON-HD and DressCode datasets. We compare the metrics under both paired (model’s clothing is the same as the given cloth image) and unpaired settings (model’s clothing differs) with other methods.
Methods
DressCode
VITON-HD
Paired
Unpaired
Paired
Unpaired
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
FID ↓
KID ↓
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
FID ↓
KID ↓
Qwen-Image-Edit-2509 [ 44 ]
0.8370
0.1831
13.5464
1.9944
13.9163
2.7315
0.7790
0.2174
20.6403
0.5018
18.7954
1.1485
Seedream 4.0 [ 36 ]
0.8217
0.1816
16.8140
7.2200
25.4943
6.1900
0.7600
0.2352
26.9076
9.3880
26.8181
8.4790
Nanobanana [ 10 ]
0.8647
0.1186
10.7050
1.0860
10.6720
0.7160
0.8100
0.1520
19.1688
1.1299
20.4191
0.4411
PROMO (Ours)
0.8919
0.0885
8.9268
0.4674
10.1217
0.5763
0.8439
0.1365
15.6658
2.0513
18.0356
1.5828
Table 2: Quantitative results on VITON-HD and DressCode datasets. Comparison between our model with image-editing methods.
Method
Texture ↑
Body Shape ↑
Style ↑
Color ↑
Overall ↑
Huiwa
94.42
88.85
94.80
99.04
78.85
Kling
87.12
93.46
79.87
96.53
60.19
Douyin
96.73
79.04
85.19
95.77
61.54
PROMO (Ours)
93.65
94.62
96.92
97.88
84.42
Table 3: Excellence rates (%) of different methods on in-the-wild dataset. A result is considered excellent if more than half annotators rate it as such. Overall indicates the percentage of images that are excellent in all four dimensions. Bold denotes the best result and underline denotes the second best.
Paired
Unpaired
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
FID ↓
KID ↓
- w/o parsing area loss
0.8896
0.0872
3.2847
0.5076
4.6390
0.9535
- w/o style prompt
0.8896
0.0926
3.7178
0.8926
5.3516
0.6180
- w/o 3D-RoPE
0.8702
0.1304
6.7293
1.7160
7.8175
2.2801
PROMO (Ours)
0.8913
0.0887
3.3103
0.4902
5.3516
0.4992
Table 4: Ablation study on DressCode dataset. The best and second-best results are highlighted in bold and underlined , respectively.
Figure 11: Visualization of ablation study on parsing area loss. Best viewed when zoomed in.
Figure 12: Visualization of ablation study on 3D-RoPE. Without 3D-RoPE, the model fails to correctly distinguish subject transfer and spatial alignment relationships between conditions.
Inference time (s)
Paired
Unpaired
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
FID ↓
KID ↓
- w/o T.S.R
22.2378
0.8913
0.0887
3.3102
0.7975
4.7391
0.5313
- w/o S.T.M
11.1030
0.8916
0.0873
3.4878
0.5126
4.8546
0.4907
PROMO (Ours)
9.1775
0.8913
0.0887
3.3103
0.4902
4.7393
0.4992
Table 5: Quantitative ablation results on DressCode, including quality metrics and inference time. T.S.R is for Temporal Self-Reference and S.T.M is for Spatial Token Merging. Inference time is averaged over 20 runs on an H800 GPU.
State Key Lab of CAD & CG, Zhejiang University and Style3D Research, China · Style3D Research, China · State Key Lab of CAD & CG, Zhejiang University, China