.png){kind=icon}
.png){kind=icon}
Return
MAGDiff: a synergistic multi-attribute-guided diffusion framework for personalized fashion garment generation
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-026-04611-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Personalized fashion garment generation aims to synthesize customized designs by integrating multiple conditioning attributes such as text, sketch, texture and color. However, existing methods often struggle with blurred textures, color distortion and modality interference when fusing these diverse inputs. Here, we propose MAGDiff, a synergistic multi-attribute-guided diffusion framework for personalized fashion garment generation. MAGDiff jointly models text, structure and style information within a unified diffusion process for fine-grained and controllable garment synthesis. Specifically, a synergistic dual-stream style integration mechanism is proposed to combine global style features with latent texture details, ensuring accurate preservation of both fine-grained texture and color distribution. To alleviate attribute interference between structure and style induced by direct feature concatenation, we further design a synergistic attribute-conditioned modulation mechanism, which regulates the interaction between structure and style features across multiple scales through learnable gated fusion, enabling adaptively to balance their contributions while maintaining structural consistency and fine-grained style. Moreover, a complementary decoupled attention strategy separates textual and stylistic guidance into parallel attention branches, reducing cross-modal redundancy caused by overlapping conditions and enhancing generation controllability. Extensive experiments on two large-scale fashion datasets (CM-Fashion-s and Polyvore-s) demonstrate that MAGDiff outperforms representative methods across multiple metrics: it achieves an FID of 15.48 (CM-Fashion-s) and 12.89 (Polyvore-s), LPIPS of 0.265 and 0.281 and superior color distribution scores (CDH of 10.31 and 10.09). MAGDiff provides a robust and flexible solution for multi-attribute-guided personalized garments generation, advancing the controllability and visual fidelity of AI-driven fashion design. Code: https://github.com/Preciousrs/MAGDiff .
Keywords:
Multi-attribute guidance
Fashion garment generation
Diffusion model
Personalized design
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K