docs(README): note Motif > V-JEPA 2.1 in point-tracking demo
Browse filesCaption: GIF is a top=Motif / bottom=V-JEPA 2.1 comparison grid; add that Motif tracks the subject more stably.
README.md
CHANGED
|
@@ -24,7 +24,7 @@ performance across image and video benchmarks — and leads on DAVIS video track
|
|
| 24 |
<p align="center">
|
| 25 |
<img src="assets/haaland_full_attn_blk20.gif" width="480" alt="Self-attention visualization on a video clip"/>
|
| 26 |
</p>
|
| 27 |
-
<p align="center"><em>Point tracking on a video clip — a query point propagated across frames by patch-feature cosine similarity.</em></p>
|
| 28 |
|
| 29 |
- **Architecture**: ViT-7B (embed 4096 / depth 40 / heads 32), patch 16, 3D axial RoPE
|
| 30 |
(`base=100`), SwiGLU FFN, LayerScale, per-head QK-norm, gated attention, 4 register tokens.
|
|
|
|
| 24 |
<p align="center">
|
| 25 |
<img src="assets/haaland_full_attn_blk20.gif" width="480" alt="Self-attention visualization on a video clip"/>
|
| 26 |
</p>
|
| 27 |
+
<p align="center"><em>Point tracking on a video clip (top: Motif, bottom: V-JEPA 2.1) — a query point propagated across frames by patch-feature cosine similarity. Motif stays locked on the subject noticeably better than V-JEPA 2.1.</em></p>
|
| 28 |
|
| 29 |
- **Architecture**: ViT-7B (embed 4096 / depth 40 / heads 32), patch 16, 3D axial RoPE
|
| 30 |
(`base=100`), SwiGLU FFN, LayerScale, per-head QK-norm, gated attention, 4 register tokens.
|