{
  "version": 1,
  "language": "en",
  "title": "Find part correspondences",
  "description": "PoseAnything generates video from pose/reference latents. Part masks and attention matching find corresponding regions for cross-attention.",
  "caption": "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.",
  "alt": "PoseAnything generates video from pose/reference latents and uses part masks and attention matching to find corresponding regions for cross-attention.",
  "sources": [
    "https://arxiv.org/abs/2512.13465",
    "https://ryan-w2024.github.io/project/PoseAnything/"
  ],
  "limitations": [
    "The playback is an example. Animation time is not measured runtime.",
    "This diagram has no fabricated attention value, heatmap, reference photograph or generated result."
  ],
  "targets": {
    "reference-pose": {
      "id": "mechanism-reference-pose",
      "kind": "node",
      "label": "Reference + pose"
    },
    "latents": {
      "id": "mechanism-latents",
      "kind": "node",
      "label": "VAE latents"
    },
    "dit-context": {
      "id": "mechanism-dit-context",
      "kind": "node",
      "label": "DiT context"
    },
    "masks": {
      "id": "mechanism-masks",
      "kind": "node",
      "label": "Part masks"
    },
    "matching": {
      "id": "mechanism-matching",
      "kind": "node",
      "label": "Part matching"
    },
    "part-attention": {
      "id": "mechanism-part-attention",
      "kind": "node",
      "label": "Part-aware attention"
    },
    "encode-latents": {
      "id": "mechanism-encode-latents",
      "kind": "edge",
      "label": "encode-latents"
    },
    "latent-context": {
      "id": "mechanism-latent-context",
      "kind": "edge",
      "label": "latent-context"
    },
    "skeleton-masks": {
      "id": "mechanism-skeleton-masks",
      "kind": "edge",
      "label": "skeleton-masks"
    },
    "mask-matching": {
      "id": "mechanism-mask-matching",
      "kind": "edge",
      "label": "mask-matching"
    },
    "attention-patterns": {
      "id": "mechanism-attention-patterns",
      "kind": "edge",
      "label": "attention-patterns"
    },
    "matched-parts": {
      "id": "mechanism-matched-parts",
      "kind": "edge",
      "label": "matched-parts"
    },
    "context-features": {
      "id": "mechanism-context-features",
      "kind": "edge",
      "label": "context-features"
    }
  },
  "stateLabels": {
    "stage": "Stage",
    "artifact": "Artifact"
  },
  "scenarios": [
    {
      "id": "part-correspondence",
      "title": "Find correspondences in different frames",
      "description": "This diagram shows dependencies in the coherence mechanism. It does not show measured attention values.",
      "initial": {
        "stage": "Available",
        "artifact": "Reference image and skeletal sequence available"
      },
      "unselected": [],
      "steps": [
        {
          "title": "Calculate latents and masks",
          "description": "The framework calculates input latents and part masks.",
          "duration": 3000,
          "active": [
            "latents",
            "masks"
          ],
          "paths": [
            "encode-latents",
            "skeleton-masks"
          ],
          "state": {
            "stage": "Calculated latents and masks",
            "artifact": "Reference/pose latents and dilated part masks"
          }
        },
        {
          "title": "Make transformer context",
          "description": "The transformer receives reference and pose tokens with attention patterns.",
          "duration": 3000,
          "active": [
            "dit-context"
          ],
          "paths": [
            "latent-context"
          ],
          "state": {
            "stage": "DiT context available",
            "artifact": "Reference and pose tokens with attention patterns"
          }
        },
        {
          "title": "Find part correspondences",
          "description": "The framework finds part correspondences from masks and attention patterns.",
          "duration": 3000,
          "active": [
            "matching"
          ],
          "paths": [
            "mask-matching",
            "attention-patterns"
          ],
          "state": {
            "stage": "Part match available",
            "artifact": "Corresponding regions in different frames"
          }
        },
        {
          "title": "Use part-aware attention",
          "description": "Cross-attention uses features from corresponding body parts.",
          "duration": 3000,
          "active": [
            "part-attention"
          ],
          "paths": [
            "matched-parts",
            "context-features"
          ],
          "state": {
            "stage": "Part-aware update",
            "artifact": "Features for corresponding body parts"
          }
        }
      ]
    }
  ]
}