This benchmark needs a browser with Canvas and JavaScript support.
3Recording view ready19.5-second tour follows. H restores controls.
1920 × 1080 design • F fullscreen • Space play / pause • R restart tour • H controls • 1-7 scenes
Final overview frame. Choose fullscreen and start the tour to record.

Landscape recording studio

Every scene is exactly 16:9. The entire card fits inside your browser without scrolling. On a non-16:9 monitor, black letterboxing preserves the design.

Screen recording

1. Open this page in a desktop browser. 2. Click Fullscreen. 3. Start your screen recorder. 4. Click Start 19.5s tour. A three-second countdown lets the controls clear before playback. Trim that countdown from the recording. Each scene holds for 6.5 seconds, then the tour finishes on the completed overview. Enable Loop to repeat.

H shows or hides controls; Space pauses or resumes; R restarts the tour; 1 to 6 and arrow keys select scenes; S freezes the completed current scene; F toggles fullscreen. Escape exits fullscreen and restores controls. Moving the mouse does not bring controls into your recording. Save PNG exports a complete frame without the controls, even during animation.

The animation reveals bars and decorative light effects. Numerical labels stay at their recorded values. This is not live inference, and nothing here contacts an inference server.

Share, download or render locally

This page is a recorded-results viewer, not a remote model demo. Viewing, animating, and exporting it do not require a DGX Spark, a vLLM install, or model weights.

Scene links support ?scene=overview, ?scene=sweep, ?scene=long, ?scene=cliff, ?scene=scale, ?scene=revision and ?scene=native. Add &clean=1 for a controls-free view. ?autoplay=1 starts the tour after a three-second countdown; add &loop=1 to repeat it. Fullscreen still requires a button or keypress.

Open the downloaded HTML directly for offline playback. For automated PNG export, use scripts/render_benchmark.py in the recipe. Links copied from the hosted page point to the deployment on which you opened it, including your own fork.