reachomk/gen2seg-mae-h

gen2seg: Generative Models Enable Generalizable Instance Segmentation

This is the official model release for the ImageNet-1k pretrained Masked Autoencoder (MAE-H) variant of our `gen2seg` generative instance segmenter. It is the same checkpoint we used to generate figures in the paper.

The sementations in the image above are different from those in the paper teaser because they are generated with this MAE-H model instead of our SD model. If you are looking for our Stable Diffusion 2 (SD) variant, you can find that at https://huggingface.co/reachomk/gen2seg-sd

Paper: https://arxiv.org/abs/2505.15263

Please see our website https://reachomk.github.io/gen2seg for demos and additional qualitative samples.

You can run this model at our GitHub: https://github.com/UCDVision/gen2seg or our Huggingface Space: https://huggingface.co/spaces/reachomk/gen2seg