Repurposing Stable Diffusion Attention for Training-Free Unsupervised Interactive Segmentation

Figure 1. We introduce M2N2, an unsupervised training-free point prompt based segmentation framework. We enhance the semantic information present in the self-attention of Stable Diffusion 2 by using a Markov process to generate semantically rich Markov-maps. We then perform a truncated nearest neighbor of each point's Markov-map to obtain a final segmentation.