DreamMix: Decoupling Object Attributes for Enhanced Editability in Customized Image Inpainting
Published in International Journal of Computer Vision (IJCV) · CCF-A · First Author, 2025
DreamMix addresses the tension between identity preservation and text-driven attribute editing in subject-driven image inpainting. It introduces an Attribute Decoupling Mechanism, Textual Attribute Substitution, and a disentangled local-global inpainting framework.
The method reaches a CLIP-T score of 0.289 on SDXL and 0.290 on FLUX, with only 4–6 reference images needed for adaptation. The project has received 100+ GitHub stars.
Recommended citation: YiCheng Yang et al. “DreamMix: Decoupling Object Attributes for Enhanced Editability in Customized Image Inpainting.” International Journal of Computer Vision.
Download Paper