We use the FFHQ-256 model from Label-Efficient Semantic Segmentation with Diffusion Models along with the few-shot data provided by the authors to generate segmentation-conditioned images.
ffhq.ipynb showcases the full pipeline to generate faces with segmentation guidance. The code is tested with torch==1.13.1 and torchvision==0.14.0.
Generating images using Stable Diffusion guided with segmentation masks is implemented in stable_diffusion.ipynb. We again use the few-shot data (LSUN-Cat) from Baranchuk et al. to guide Stable Diffusion with segmentation masks. The notebook uses the Stable Diffusion v1.2 model. To run you need to first install the LDM environment, as described in the official Stable Diffusion repository.