Palette results on panorama generation. The middle 256x256 image is the original image. The model is repeatedly applied to 50% right uncropping and 50% left uncropping (4 times each) to obtain the final 256x1280 image. For more click here. These examples show that Palette is surprisingly robust, generating realistic and coherent outputs even after 8 repeated applications of uncrop.

Image-to-Image Diffusion Models 

We develope Palette: a unified framework for image-to-image translation based on conditional diffusion models. We evaluate this framework on four challenging image-to-image translation tasks, namely colorization, in-painting, uncropping, and JPEG restoration.

Palette results on infinite zoom out. The most zoomed-in frame is the original image. We then repeatedly apply Palette for uncropping along all 4 sides to generate infinite zoom out effect. For more examples click here.