Research map / Image generation / Editing and style transfer
GarFast: Realistic and Fast Garment Transfer with a Simplified Parser-Free Approach
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GarFast, a single lightweight ResUnet encoder-decoder garment-transfer model trained parser-free via a two-stage weight-shared cycle with correction, fine-grained domain consistency, skin-variant refinement and skinMix losses.
Problem
Virtual garment try-on needs both high-fidelity results and fast inference, but parser-based methods depend on human parsing and existing parser-free pipelines use dual networks or flawed pseudo-labels that slow training and inference.
Approach
Uses one weight-shared lightweight generator optimized in two stages where the real person image serves first as input and then as supervision, with a correction strategy using identity masks, pixel-level domain discrimination (Dfine), SkinMix-based discriminator regularization (Dpixel), and a perceptual skin-variant refinement loss.
Key results
- On VITON, GarFast reaches SSIM 0.93 and FID 9.39, the best among compared methods in Table 1.
- On VITON-HD, GarFast attains SSIM 0.908 and LPIPS 0.070, while FID (9.64) and KID are slightly behind diffusion baselines such as StableVITON and DCI-VTON.
- GarFast has 10.98M parameters, 22.27G FLOPs and 147 FPS on VITON, versus 153M parameters and 78 FPS for DCTON; inference is under 0.03 s on an A100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add correction strategy (ablation) | GarFast · VITON | SSIM: 0.47 -> 0.89 | Helped |
| add skin-variant refinement (ablation) | GarFast · VITON | SSIM: 0.89 -> 0.90 | Helped |
| add fine-grained domain consistency (ablation) | GarFast · VITON | SSIM: 0.90 -> 0.92 | Helped |
| add SkinMix regularization (ablation) | GarFast · VITON | SSIM: 0.92 -> 0.93 | Helped |
| add GarFast's two-stage framework and optimization strategies | garment transfer · VITON-HD | FID: 8.754 -> 9.640 | Hurt |
| add GarFast's two-stage framework and optimization strategies | garment transfer · DressCode Upper | FID: 11.27 -> 12.01 | Hurt |
Surprising findings
- Without the correction strategy, FID barely changes during training (SSIM 0.47, FID 106.79 in ablation A), showing the two-stage model otherwise fails to converge.
- Diffusion baselines achieve lower FID/KID than GarFast on VITON-HD and DressCode despite GarFast's better SSIM and LPIPS.
Limitations
- Limited dataset size and single source website restrict skin-tone diversity.
- Slightly inferior FID/KID versus diffusion-based methods on high-resolution benchmarks.
Experimental setup
Datasets: VITON VITON-HD DressCode
Benchmarks: VITON VITON-HD DressCode
Metrics: SSIM LPIPS FID KID FPS FLOPs #Params
Baselines: GP-VTON DCI-VTON StableVITON PbE Anydoor LaDI-VTON HR-VITON USC-PFN PF-AFN DCTON
Builds on: cycle consistency (CycleGAN) USC-PFN DCTON PF-AFN ResUnet
Closest papers by meaning
- FreeCloth: Free-form Generation Enhances Challenging Clothed Human Modeling (CVPR 2025 · similarity 0.88)
- GBC-Splat: Generalizable Gaussian-Based Clothed Human Digitalization under Sparse RGB Cameras (CVPR 2025 · similarity 0.88)
- FashionR2R: Texture-preserving Rendered-to-Real Image Translation with Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Garment Recovery with Shape and Deformation Priors (CVPR 2024 · similarity 0.87)
- AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.