ziqihuangg/Collaborative-Diffusion

Employs a "dynamic diffuser" architecture that learns spatial and temporal influence functions to selectively weight contributions from pre-trained uni-modal diffusion models, enabling coordinated multi-modal control (text, segmentation masks, sketches) during the reverse diffusion process. Supports both face generation from multi-modal conditions and editing of real images while preserving identity, with implementations at 256×256 and 512×512 resolutions built on latent diffusion. Compatible with enhancement techniques like FreeU and integrates with PyTorch/Hugging Face transformers ecosystem.

438 stars. No commits in the last 6 months.