MIGC

Why we included this project

Most text-to-image models struggle when a single prompt asks for several distinct objects: they blend attributes, drop items, or scatter things randomly. MIGC, a CVPR 2024 Highlight paper, addresses this by letting you draw a box or mask for each instance, so you can tell the model exactly where each object goes and what it should look like. It includes attribute and quantity controls, a benchmark for evaluating multi-instance generation, and pretrained weights for Stable Diffusion 1.4. A WebUI and Colab demo let you try it before integrating it into your own pipeline. The MIGC++ follow-up, published in TPAMI, adds an iterative editing mode that refines results while keeping identities consistent. One caveat: the code is released for non-commercial research use only, so check that against your project before committing.