Research map / Image generation / Editing and style transfer
TexFit: Text-Driven Fashion Image Editing with Diffusion Models
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 TexFit, a text-only local fashion image editing method that predicts an editing-region mask from the text prompt and conditions a fine-tuned Stable Diffusion v1.4 inpainting-style model on it, plus a new DFMM-Spotlight region-text dataset.
Problem
Local fashion image editing usually needs auxiliary inputs (poses, keypoints, sketches) to locate the edit region, and existing fashion datasets only provide global descriptions, so precise text-only local editing is hard.
Approach
A text-based editing region location module (ERLM), an encoder-decoder conditioned on CLIP text embeddings, predicts the region mask from the prompt; the mask and masked image are then fed as conditions to a latent diffusion model, and the output is composited with the original outside the mask.
Key results
- TexFit achieves FID 10.77 and LPIPS 0.052 on the expanded DFMM-Spotlight test set, versus 12.90/0.063 for SD-Inpaint and 16.34/0.188 for SDEdit.
- TexFit's CLIP-S is 28.10, close to SDEdit's 28.18 and above SD-Inpaint (26.87) and DiffEdit (26.28).
- In a 36-user study over 1000 cases, TexFit received the most votes for image fidelity (563), text matching (607) and identity preservation (841).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TexFit instead of SD-Inpaint | Fashion image editing · DFMM-Spotlight expanded test set | FID: 12.90 -> 10.77 | Helped |
| Use TexFit instead of SD-Inpaint | Fashion image editing · DFMM-Spotlight expanded test set | LPIPS: 0.063 -> 0.052 | Helped |
| Use TexFit instead of SDEdit | Fashion image editing · DFMM-Spotlight expanded test set | CLIP-S: 28.18 -> 28.10 | No clear effect |
| Use TexFit instead of competing methods | Fashion image editing · Human-subject evaluation | ID Preservation selections: 38; 69; 52 -> 841 | Helped |
| Use ERLM-predicted region location instead of DiffEdit region location (ablation) | TexFit fashion image editing · DFMM-Spotlight expanded test set | FID: 12.92 -> 10.77 | Helped |
| Replace the TexFit editing module with SD-Inpaint (ablation) | ERLM region location · DFMM-Spotlight expanded test set | FID: 10.77 -> 58.19 | Hurt |
| Replace the ERLM-predicted mask with the ground-truth mask (ablation) | TexFit · DFMM-Spotlight expanded test set | FID: 10.77 -> 10.51 | Helped |
Surprising findings
- Replacing TexFit's ERLM mask with DiffEdit's automatically generated mask greatly degrades fidelity, showing region localization accuracy is critical.
Limitations
- Evaluation is limited to a single dataset derived from DeepFashion-MultiModal at 512x256 resolution.
- Region masks are restricted to five semantic categories (top, pants, outer, dress, accessories).
- CLIP-S is computed on images with non-edit regions filled white, a non-standard protocol.
Experimental setup
Datasets: DeepFashion-MultiModal DFMM-Spotlight
Benchmarks: DFMM-Spotlight
Metrics: FID LPIPS CLIP-S human preference counts
Baselines: SDEdit SD-Inpaint DiffEdit Multimodal Garment Designer
Builds on: Stable Diffusion latent diffusion models CLIP classifier-free guidance DiffEdit
Closest papers by meaning
- FashionTailor: Controllable Clothing Editing for Human Images with Appearance Preserving (AAAI 2025 · similarity 0.90)
- Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction (CVPR 2025 · similarity 0.89)
- HieraFashDiff: Hierarchical Fashion Design with Multi-stage Diffusion Models (AAAI 2025 · similarity 0.88)
- AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- Cascaded Diffusion Models for Virtual Try-On: Improving Control and Resolution (AAAI 2025 · similarity 0.87)
- AIpparel: A Multimodal Foundation Model for Digital Garments (CVPR 2025 · similarity 0.86)
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.