Disclaimer for Visualization Results

Since we don’t store the mouse’s initial position—only the delta of dx/dy—the initial position and alignment can appear inconsistent in the videos.
Note that in FPS games, the game engine automatically recenters (locks) the cursor, whereas our visualizer does not, for intuitive understanding.
We observed that when the model predicts a "press" event without a subsequent "release" event, the visualizer remains in the pressed state, which occurs intermittently.

G-IDM Pseudo-Labeling Results on a YouTube Video

On Minecraft, Stardew Valley, and other 3D games featuring inventory and various GUI-based interactions (such as inventory management and item trading), G-IDM labels both 3D exploration scenes and GUI screens solely from raw video, without any extra input filtering, demonstrating robust performance across diverse scenarios.
On Counter-Strike 2, we observed that G-IDM recognizes spectating mode and pauses action predictions on spectating mode.
In addition, G-IDM robustly labels gameplay and UI states in 2D games like Brotato, showing its generalization beyond 3D environments.

Counter-Strike 2

https://youtu.be/3JZ_DZz523g

Brotato

https://youtu.be/2299slgGGPA

Stardew Valley

https://youtu.be/VxX3Lw-ifW0

Minecraft

https://youtu.be/Ir0VMS762IY

Slime Rancher

https://youtu.be/btQ0qOZNATQ

Barony

https://youtu.be/0vo5SMO1RfU