Tsingularity/dift

Extracts dense semantic correspondence features from diffusion model intermediate representations (U-Net upsampling blocks at configurable timesteps), supporting both Stable Diffusion and guided-diffusion models. Enables downstream tasks including point matching, edit propagation across images, homography estimation, and video object segmentation through cosine similarity matching of per-pixel feature vectors. Provides tunable parameters for memory/accuracy tradeoffs (image resolution, ensemble size, feature extraction layer) and includes evaluation scripts for SPair-71k, HPatches, and DAVIS benchmarks.

754 stars. No commits in the last 6 months.