Update supplied-video FlowHMR mesh and tracking demo

README.md

@@ -38,21 +38,27 @@ Official source: [FlowHMR](https://github.com/flowhmr/flowhmr), revision
 
 ## Visual Results
 
-### Supplied-video demo — Latest (PHC+ GRPO)
+### Supplied-video demo — Latest
 
 <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>
 
 [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)
+[Render provenance](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest/resolve/main/media/demo_latest_mesh.json) ·
+[Other checkpoint on the same video](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Base#visual-results)
 
 **Checkpoint:** [Motius-FlowHMR-Latest](https://huggingface.co/ZeyuLing/Motius-FlowHMR-Latest),
-the official Latest release after PHC+ GRPO post-training.
+the official Latest release after PHC+ GRPO post-training. This video is an actual Latest prediction.
 
 - 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).
 
+Base and Latest use the exact same cached tracking boxes, SAM features and camera
+estimates (shared frontend SHA256: `c26b81670919b5b78b3661a022e3e7db294078cfbb464825fdd8246ea5b8419b`).
+Both use the same seed and sampling settings. A single camera framing fits both
+predicted sequences, so the display scale and view are identical.
+
 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.

media/demo_latest_mesh.json

@@ -15,18 +15,28 @@
     "height": 585,
     "frames": 750,
     "vertex_count": 6890,
-    "half_height": 1.1323156712833715,
+    "half_height": 1.1555242160887533,
     "camera_target": [
-      -0.005707634257960004,
-      0.8669544873149713,
-      0.1286162645327557
+      0.01125761379322961,
+      0.8669598787713986,
+      0.1328583852640172
     ],
     "camera_position": [
-      1.4343384479539978,
-      1.730982136642146,
-      -5.631568064315076
+      1.4513036960051875,
+      1.7309875280985734,
+      -5.6273259435838145
     ]
   },
   "source_sha256": "8f9e9c1193163940eafb512e9a0621a31fe316406e7a69aa3a357f6deceb17f1",
-  "video_sha256": "ddfd29ce41206457c9e5150b72bf9d1e7eae52f0725eb049442b7bf21404b60d"
+  "video_sha256": "76b0c753a950b2b6fd43f9dab38f41371a9aa9eb1d5d33077cf589cc245a0960",
+  "inference": {
+    "camera_translation": "conditioning only; no metric world alignment",
+    "cfg_scale": 1.0,
+    "euler_steps": 20,
+    "seed": 0,
+    "shared_frontend_sha256": "c26b81670919b5b78b3661a022e3e7db294078cfbb464825fdd8246ea5b8419b",
+    "source_video_sha256": "8f9e9c1193163940eafb512e9a0621a31fe316406e7a69aa3a357f6deceb17f1",
+    "vggt_frame_interval": 100
+  },
+  "camera_reference_sha256": "29177e7879f9753cf84ce32714b5107f2e3125ea49db8292d3a9f5fdcfab7f1d"
 }

media/demo_latest_mesh.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ddfd29ce41206457c9e5150b72bf9d1e7eae52f0725eb049442b7bf21404b60d
-size 7564846
+oid sha256:76b0c753a950b2b6fd43f9dab38f41371a9aa9eb1d5d33077cf589cc245a0960
+size 7538359