Spaces:

jsu27

/

decomp-diffusion

like 2

Running

Hi, this is a demo for our work in unsupervised image decomposition with diffusion models (project page ). We can discover global and object-level factors from images, where factors are represented by diffusion models, and can combine them to generate novel images. Some examples are shown on this page, such as decomposing an image of 4 geometric solids into its constituent objects, or combining facial features across 2 human face images. We would like GPU access for faster inference and the ability to upsample results with a large model. Thanks!