Gemma 4 camera resolution benchmark

This opt-in check loads about 5.55 GB of Gemma 4 E2B decoder and BF16 vision packs. It runs one model at a time, then releases it. No model loads until you press the button or use ?autorun=1. The default uses local packs; use the published packs to run this on your own Mac or other device, reusing cached weights when available. Stop other model tabs first.

It compares 70, 40, and 16 requested image tokens on the same 320 × 240 frames. Add ?budgets=70,40,35,16 to include the 35-token setting too. Generated frames are only a runtime diagnostic, not an image quality test. To evaluate labeled images, pass ?images=URL for a JSON array of {"id":"...","url":"...","expectedPerson":true,"expectedSmile":false} rows. Image URLs must allow browser CORS access.