how to draw representation t-sne in detail?

[Ckck12]

Thank you for your effort! @ChristianInterno

I would like to know about how to draw representation t-sne in detail.

I'm curious about what the nodes and edges represent in the figure in the paper. Did you only use curvature information?

Thanks

[ChristianInterno]

Hi, thanks for reaching out.

Trajectory: Each line, formed by dots connected by edges, represents one single video.
Dots: Each dot is the video's feature curvature at a specific point in time (i.e., a single frame).
Edges: An edge connects consecutive dots for the same video.

Hope this helps!

[Ckck12]

Thank you for your reply, and I'm inspired by your great research.

1. Could you clarify exactly what you fed into t-SNE for Figure 4? For each video/time step, is the embedded point a single scalar curvature value θᵢ, the full per-video curvature vector (θ₁…θ_{T−2}), or a feature that also includes stepwise distance dᵢ or any DINO representation terms?

2. And did you consider not using spatial information because you thought curvature was a good feature for detecting AIGC and generalization? You probably considered utilizing the z vector for training—why did you decide not to use it?