Research map / Image generation / Editing and style transfer
HieraFashDiff: Hierarchical Fashion Design with Multi-stage 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 HieraFashDiff, a two-stage (ideation then iteration) hierarchical diffusion model fine-tuned from Stable Diffusion that generates fashion design drafts from high-level concepts and performs iterative local attribute editing, plus a 5.2k-image hierarchically captioned fashion dataset.
Problem
Existing fashion generation and editing models ignore the practical design workflow (ideation followed by iteration), and lack support for both concept-level generation and fine-grained attribute-level editing in one framework.
Approach
Decompose the diffusion denoising trajectory into an ideation stage conditioned on high-level design concepts and sequential iteration sub-stages conditioned on accumulating low-level apparel attributes (ordered by affected area), with pose conditioning and body-part masks to keep non-target regions unchanged during editing.
Key results
- Draft generation FID 10.27, Coverage 0.76, CLIP-S 30.61 vs best baseline Attend-and-Excite FID 15.33, Coverage 0.63, CLIP-S 28.46 on HieraFashion.
- Iterative editing achieves FID 10.74 at A1 rising to 12.59 at A5, with CLIP-S 31.39 to 33.01, outperforming BrushNet (FID 12.92 to 15.06) across all five attributes.
- Removing hierarchical prompts (Flat variant) yields the worst FID (14.29 for draft generation) and CLIP-S (28.17).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| hierarchical multi-stage descriptions (ideation + iteration) vs flat combined description (ablation) | HieraFashDiff draft generation · HieraFashion draft generation | FID: 14.29 -> 10.27 | Helped |
| hierarchical multi-stage descriptions vs flat combined description (ablation) | HieraFashDiff draft generation · HieraFashion draft generation | CLIP-S: 28.17 -> 30.61 | Helped |
| proposed attribute ordering (by affected area) vs random ordering (ablation) | HieraFashDiff iterative editing · HieraFashion iterative editing (A1-A5) | FID: e.g., A1 10.74 vs 14.26; A5 12.59 vs 16.95 | Helped |
| proposed attribute ordering vs random ordering (ablation) | HieraFashDiff draft generation · HieraFashion draft generation | CLIP-S: 28.98 -> 30.61 | Helped |
| HieraFashDiff vs strongest baselines on draft generation | draft generation · HieraFashion draft generation | FID: 15.33 -> 10.27 (vs Attend-and-Excite) | Helped |
| body part masks added to TexFit (TexFit-M) vs predicted editing regions | TexFit · HieraFashion iterative editing (A1) | FID: 13.98 -> 13.27 | Helped |
| HieraFashDiff vs baselines on iterative editing across all five attributes | iterative local editing · HieraFashion iterative editing (A1-A5) | FID: A1 10.74 vs 12.92 (BrushNet best baseline) | Helped |
| pose map conditioning and body part masks removed from editing (ablation) | HieraFashDiff local editing · HieraFashion local editing ablation | FID and CLIP-S | Helped |
Limitations
- Editing quality degrades over successive iterations (FID rises from 10.74 to 12.59).
- MGD could not be retrained because its training code is not public, so comparison is partial.
- Evaluation uses only FID, Coverage and CLIP-S on a single dataset of 1,200 test examples, with no human or LLM-based judging.
Experimental setup
Datasets: HieraFashion Dress Code Multimodal FACAD DeepFashion-MM
Benchmarks: HieraFashion
Metrics: FID Coverage CLIP-S
Baselines: Cogview Stable Diffusion Attend-and-Excite FICE SD-Inpaint BrushNet TexFit MGD
Builds on: Stable Diffusion Blended Latent Diffusion Attend-and-Excite BrushNet
Closest papers by meaning
- TexFit: Text-Driven Fashion Image Editing with Diffusion Models (AAAI 2024 · similarity 0.88)
- Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction (CVPR 2025 · similarity 0.86)
- FashionTailor: Controllable Clothing Editing for Human Images with Appearance Preserving (AAAI 2025 · similarity 0.86)
- AIpparel: A Multimodal Foundation Model for Digital Garments (CVPR 2025 · similarity 0.85)
- Design2GarmentCode: Turning Design Concepts to Tangible Garments Through Program Synthesis (CVPR 2025 · similarity 0.85)
- StyleTailor: Towards Personalized Fashion Styling via Hierarchical Negative Feedback (AAAI 2026 · similarity 0.85)
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.