Rehost MediaPipe FaceMesh v2 as native safetensors fx-program

README.md

@@ -0,0 +1,24 @@
+---
+license: apache-2.0
+library_name: sceneworks
+tags:
+- face-landmarks
+- mediapipe
+- facemesh
+---
+
+# MediaPipe FaceMesh v2 (SceneWorks native rehost)
+
+Native safetensors rehost of the MediaPipe FaceMesh v2 face-landmark detector, for SceneWorks' MLX and Candle training backends (frozen, differentiable face-landmark loss for character LoRA training).
+
+## Provenance
+- Source: [py-feat/mp_facemesh_v2](https://huggingface.co/py-feat/mp_facemesh_v2) at revision `39eb85054cf76fe0f57b7e12d6765ae89d89f2b5`, file `face_landmarks_detector_Nx3x256x256_onnx.pth` (an onnx2torch GraphModule of Google's MediaPipe FaceMesh v2 landmark model).
+- License: Apache-2.0, as stated by the source repository.
+- Converted with `crates/media/mlx-gen/tools/convert_mp_facemesh_v2.py` (SceneWorks inference repo, commit `1d87a8eb2`), which lowers the FX graph to an "fx-program" stored in the safetensors metadata (`program`) alongside the 281 parameter tensors.
+
+## Verification
+- Self-check (lowered program re-executed against the torch module): max abs error 0.0.
+- Native executors vs torch on a fixed seeded input: MLX max abs 3.09e-4, Candle max abs 2.86e-4 (output magnitude ~216, ≈1.4e-6 relative).
+- Input: N×3×256×256; output0: N×1×1×1434 (478 landmarks × 3).
+
+`face_landmarks_detector.safetensors` sha256: `94e0dda4bb58733fc7454997868232bdcc2b11531a18d2bbfec5479c75f574f4`

face_landmarks_detector.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94e0dda4bb58733fc7454997868232bdcc2b11531a18d2bbfec5479c75f574f4
+size 4997792