Visualize VGGT Inference Process via DPT Logit Lens

This web demo explores how VGGT's geometry emerges across layers using a "logit lens"-like style ablation. Inspired by the paper Understanding Multi-View Transformers (Stary and Gaubil et al., 2025).
This

• visualize point clouds and camera poses encoded in the intermediate hidden layers (by applying DPT head with swapping its inputs with earlier layers and applyng camera head)
• canonicalizes poses, aligns scales, and subsamples colored point clouds
• visualizes each result in Gradio, and combined (interpolated) motions via Viser.

The point clouds and camera poses are refined step by step.

Setup

pip install -r requirements.txt
pip install git+https://github.com/facebookresearch/vggt.git

Run and open the page by web browsers.

python run_g2.py

---

---

Other Examples

Dec-02-2025.10-07-10-2.mov

Dec-02-2025.10-12-54-2.mov