BLOG ILLUSTRATION / EXECUTION STORY

Find part correspondences

PoseAnything generates video from pose/reference latents. Part masks and attention matching find corresponding regions for cross-attention.

Find part correspondencesPoseAnything generates video from pose/reference latents. Part masks and attention matching find corresponding regions for cross-attention. Reference + pose image + skeletons VAE latents concat + patchify DiT context reference + pose tokens Part masks segment + dilate Part matching find matches Part-aware attention matched-part features skeleton attention patterns
Author-redrawn PoseAnything mechanism diagram. Part matching uses skeleton-derived masks and attention patterns. Part-aware cross-attention uses the matched regions and transformer context. This diagram can show independent prerequisites together. Playback is not evidence of hardware execution at the same time. Reference images, results, attention heatmaps and numeric color scales stay in the original. Open full-size SVG