Local LLM writes Three.js demos, watches video, and rewrites the code

My local LLM demoscene generator can now watch its own output and rewrite it!

An open-source project lets a local LLM write Three.js demos, then captures 30 seconds of video at 2fps and sends it back to Qwen 3.8 for a visual improvement pass. The author uses Ninfer on a single 5090, achieving ~210 tok/s decode and 14.66s per rewrite. The model still makes dumb mistakes like using only 1/4 of the screen or walking backwards through a maze; video feedback helps catch those. LM Studio is also supported for regular generation, but video rewriting requires Ninfer. The post doesn't specify Qwen 3.8's parameter count.