Making ML-powered web games with Transformers.js — MobileViT inference in browser
AI Impact Summary
This project demonstrates the feasibility of deploying ML-powered web games directly in the browser using Transformers.js and a MobileViT model. The core technical challenge is running the vision transformer model in a browser environment, achieved through a combination of ONNX conversion and Web Workers for offloading inference. This approach enables real-time prediction speeds exceeding 60 predictions per second, mirroring the performance of the original Quick, Draw! game.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info