Unsupervised Keypoints from Pretrained Diffusion Models

https://stablekeypoints.github.io/ Image dataset Randomly initialized tokens Optimized tokens Estimated keypoints Localize Figure 1. Teaser -we propose an unsupervised method to learn keypoints based on optimizing text embeddings of latent diffusion models [44]. Our method is motivated by the fact that random text tokens already respond roughly consistently to semantically similar regions. By promoting localization we obtain unsupervised keypoints that outperform the state-of-the-art.