MIGC
MIGC and MIGC++ are advanced deep learning frameworks for multi-instance generation in text-to-image synthesis. Developed as a CVPR 2024 Highlight and later accepted by TPAMI 2024, these tools provide precise control over the layout and attributes of multiple objects within a single generated image. The initial MIGC model enables users to define instance locations and attributes for complex scene construction using stable diffusion. MIGC++ extends these capabilities with iterative editing modes and the ability to simultaneously utilize bounding boxes and segmentation masks for stricter spatial control. The software suite supports attribute modification, quantity control, and animation creation. It includes pretrained weights for Stable Diffusion v1.4, with migration support for v1.5, and offers command-line inference scripts alongside a graphical WebUI interface. Researchers and developers can access the COCO-MIG benchmark to evaluate performance, replicate results using the provided installation guide, or te