About features' similarities

[ChimpOnCloud]

Great work! I am curious that in Fig.3, how did you calculate different feature map's similarity(in DiT-XL-2/256x256, the shape of each map is [2, 256, 1152])? In your heatmap, each two maps' similarity are represented into a single number between 0 and 1.

[pratheba]

The features for each layer and timestamp is computed. Cosine similarity is applied on the flattened features per layer between two time steps.