Research map / Image generation / Editing and style transfer
FashionTailor: Controllable Clothing Editing for Human Images with Appearance Preserving
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 FashionTailor, a triple-flow diffusion framework for instruction-based part-level clothing structure editing on human images, trained on a six-million-triplet StructureFashion dataset.
Problem
Editing the local garment structure (e.g., neckline, sleeve, length) of real full-body human images via simple text instructions while preserving clothing appearance and identity, which existing GAN, inversion, adapter and instruction-editing methods handle poorly.
Approach
A triple-flow design (appearance flow via ClothingNet with a semantic-refined module, instruction-based structure flow, and denoising UNet) fused by cross-blend attention that injects fine-grained appearance features from ClothingNet's self-attention into the denoising UNet, initialized from IP2P and Stable Diffusion v1.5.
Key results
- FashionTailor reaches FID 24.70, KID 0.515, LPIPS 0.048, SSIM 0.916, PSNR 24.444 on StructureFashion, beating IP2P (FID 25.37, LPIPS 0.057, SSIM 0.904) and all other baselines.
- Ablation shows ClothingNet (B2) and cross-blend attention with semantic-refined module (B7) progressively improve KID from 0.600 (B1) to 0.515 (B7).
- Self-enhanced blend attention gave the best generation quality among the four CBA variants; cross-enhanced variant had highest CLIP-S but worst KID/LPIPS/SSIM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace IP2P with FashionTailor | CSE editing · StructureFashion | KID: 0.600 -> 0.515 | Helped |
| add ClothingNet with multi-scale feature injection (ablation) | FashionTailor · StructureFashion | KID: 0.600 -> 0.555 | Helped |
| add ClothingNet with multi-scale feature injection (ablation) | FashionTailor · StructureFashion | LPIPS: 0.569 -> 0.515 | Helped |
| add cross-blend attention (ablation) | FashionTailor · StructureFashion | CLIP-S: 37.411 -> 38.451 | Helped |
| use cross-enhanced blend attention (ablation) | FashionTailor · StructureFashion | KID: 0.523 -> 0.534 | Hurt |
| use cross-enhanced blend attention (ablation) | FashionTailor · StructureFashion | CLIP-S: 38.526 -> 38.672 | Helped |
| use self-enhanced blend attention (ablation) | FashionTailor · StructureFashion | KID: 0.530 -> 0.523 | Helped |
| add the semantic-refined module to ClothingNet (ablation) | FashionTailor · StructureFashion | LPIPS: 0.509 -> 0.486 | Helped |
| add the semantic-refined module to ClothingNet (ablation) | FashionTailor · StructureFashion | SSIM: 0.913 -> 0.916 | Helped |
Surprising findings
- Cross-enhanced blend attention achieves the highest CLIP scores yet performs worst on KID, LPIPS and SSIM, which the authors attribute to over-editing.
Limitations
- Due to computational resource limits, not all available training data was used.
- Evaluation is restricted to the StructureFashion dataset plus qualitative real-world examples, with no human evaluation reported.
Experimental setup
Datasets: StructureFashion
Benchmarks: StructureFashion
Metrics: FID KID LPIPS SSIM PSNR CLIP-S D-CLIP
Baselines: MasaCtrl NTI+FPE InfEdit IP-Adapter MagicClothing MagicBrush InstructPix2Pix
Builds on: InstructPix2Pix Stable Diffusion MasaCtrl Prompt-to-Prompt
Closest papers by meaning
- TexFit: Text-Driven Fashion Image Editing with Diffusion Models (AAAI 2024 · similarity 0.90)
- AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- ChatGarment: Garment Estimation, Generation and Editing via Large Language Models (CVPR 2025 · similarity 0.87)
- Controllable Human Image Generation with Personalized Multi-Garments (CVPR 2025 · similarity 0.87)
- AIpparel: A Multimodal Foundation Model for Digital Garments (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.