Update supplied-video FlowHMR mesh and tracking demo

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+media/demo_input.mp4 filter=lfs diff=lfs merge=lfs -text
+media/demo_latest_mesh.mp4 filter=lfs diff=lfs merge=lfs -text
+media/demo_latest_mesh_preview.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -9,7 +9,7 @@ tags:
 - flow-matching
 ---
 
-<h1 align="center">FlowHMR Model Card</h1>
+<h1 align="center">FlowHMR Latest Model Card</h1>
 
 <p align="center"><strong>491-D SMPL-H motion capture from monocular RGB video, with native flow-matching training.</strong></p>
 
@@ -38,25 +38,33 @@ Official source: [FlowHMR](https://github.com/flowhmr/flowhmr), revision
 
 ## Visual Results
 
-<!-- MOTIUS_TASK_DEMOS:START -->
+### Supplied-video demo — Latest (PHC+ GRPO)
 
-### Task Demos
+<video src="https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest/resolve/main/media/demo_latest_mesh.mp4" poster="https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest/resolve/main/media/demo_latest_mesh_preview.png" controls playsinline preload="metadata" width="100%"></video>
 
-| Task | Input / condition | Rendered output | More |
-| --- | --- | --- | --- |
-| Monocular Motion Capture | Monocular RGB video | <video src="https://github.com/user-attachments/assets/791ec4c4-638d-478c-b79d-7ddb09d3582b" controls></video> | [MP4](https://github.com/user-attachments/assets/791ec4c4-638d-478c-b79d-7ddb09d3582b) |
+[Input video](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest/resolve/main/media/demo_input.mp4) · [Download demo](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest/resolve/main/media/demo_latest_mesh.mp4) ·
+[Render provenance](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest/resolve/main/media/demo_latest_mesh.json)
 
-Every public `infer_*` API is represented by a GitHub-native H.264 video player. **All cases** opens the optional interactive comparison.
+**Checkpoint:** [Motius-FlowHMR-Latest](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest),
+the official Latest release after PHC+ GRPO post-training.
 
-<!-- MOTIUS_TASK_DEMOS:END -->
+- Official checkpoint SHA256: `b7eb06b119feaf5ad7287bafdb2fff38d42dfc22ea9d256f8cdd2ada26dfc53a`.
+- Motius inference safetensors SHA256: `f4ae4168a6f576ffe08a3a19b245e0fd65c0d304b8e93bc2c6ab0a46a6f86912`.
+- 750 frames, 25 seconds, 30 fps; seed 0, 20 Euler steps, CFG 1.
+- VGGT camera sampling every 100 frames (9 keyframes interpolated to 750).
 
+The left panel displays the supplied video with the Motius frontend's tracking
+bbox. The right panel reconstructs SMPL-H vertices from the saved predicted
+rotations, shape and translation, using Motius's bundled implementation.
+Three.js provides smooth normals, depth-tested surfaces, lighting and floor
+shadows. The fixed camera fits the complete sequence; horizontal root motion
+is centered for display, while vertical motion is preserved. This is a world
+view, not a mesh reprojection into the source camera. Original audio is retained.
+No external FlowHMR checkout is needed for inference or mesh reconstruction.
 
-| Base | Latest |
-| --- | --- |
-| <video src="https://github.com/user-attachments/assets/791ec4c4-638d-478c-b79d-7ddb09d3582b" controls></video> | <video src="https://github.com/user-attachments/assets/ca588fd7-52c8-49a8-af11-e0def808a715" controls></video> |
+Earlier parity-validation clips: [Base, 30 frames](https://github.com/user-attachments/assets/791ec4c4-638d-478c-b79d-7ddb09d3582b) ·
+[Latest, 30 frames](https://github.com/user-attachments/assets/ca588fd7-52c8-49a8-af11-e0def808a715). These use a different input clip.
 
-Real-video previews are generated with the public Motius pipeline, seed 0,
-20 Euler steps, CFG 1 and the official first-frame floor alignment.
 
 ## Model Overview
 

media/demo_input.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f9e9c1193163940eafb512e9a0621a31fe316406e7a69aa3a357f6deceb17f1
+size 6645208

media/demo_latest_mesh.json

@@ -0,0 +1,32 @@
+{
+  "renderer": "motius-threejs-smooth-mesh",
+  "checkpoint_label": "FlowHMR Latest",
+  "checkpoint_sha256": "f4ae4168a6f576ffe08a3a19b245e0fd65c0d304b8e93bc2c6ab0a46a6f86912",
+  "source_revision": "f12e6a2d46a63d771a66dbb6b5598a1be65b9eea",
+  "frames": 750,
+  "fps": 30.0,
+  "width": 1600,
+  "height": 800,
+  "vertices": 6890,
+  "faces": 13776,
+  "view": "root-centered world; fixed camera; original Y retained",
+  "camera": {
+    "width": 500,
+    "height": 585,
+    "frames": 750,
+    "vertex_count": 6890,
+    "half_height": 1.1323156712833715,
+    "camera_target": [
+      -0.005707634257960004,
+      0.8669544873149713,
+      0.1286162645327557
+    ],
+    "camera_position": [
+      1.4343384479539978,
+      1.730982136642146,
+      -5.631568064315076
+    ]
+  },
+  "source_sha256": "8f9e9c1193163940eafb512e9a0621a31fe316406e7a69aa3a357f6deceb17f1",
+  "video_sha256": "ddfd29ce41206457c9e5150b72bf9d1e7eae52f0725eb049442b7bf21404b60d"
+}

media/demo_latest_mesh.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddfd29ce41206457c9e5150b72bf9d1e7eae52f0725eb049442b7bf21404b60d
+size 7564846