Add model card

README.md

@@ -0,0 +1,30 @@
+---
+pipeline_tag: image-to-3d
+---
+
+# Point4D: Long-range 4D Motion Reconstruction
+
+Point4D is a feed-forward 4D reconstruction model for long-range video sequences. It tracks points through long videos by chaining 3D queries across overlapping chunks, supporting long-horizon 4D motion reconstruction.
+
+- Paper: [Point4D: Long-range 4D Motion Reconstruction](https://huggingface.co/papers/2609.09145)
+- Project page: [https://point-4d.github.io](https://point-4d.github.io)
+- Code: [https://github.com/point-4d/Point4D](https://github.com/point-4d/Point4D)
+
+## Usage
+
+```python
+import numpy as np
+from point4d.loader import load_model
+from point4d.inference import track_long_video, LongTrackConfig
+
+model = load_model(device="cuda")     # uses configs/ and checkpoints/ by default
+
+images = ...                      # (T, H, W, 3) uint8 RGB
+query_uv = np.array([[260, 147]], dtype=np.float32)   # pixels in frame 0
+
+traj, conf = track_long_video(model, images, query_uv,
+                              LongTrackConfig(chunk_size=64, overlap=16))
+# traj: (T, N, 3) in frame 0 camera coordinates,  conf: (T, N)
+```
+
+See the [GitHub README](https://github.com/point-4d/Point4D) for installation and the full interactive demo instructions.