Ethical AI upscaling for games. (formerly "mender")

What does it do?

It upscales an image by a factor of $4$ (1K → 4K, etc.) with a minimal hit to performance. For the renderer implemented in the official example, it produces a $2024\times2024$ image ~12 times faster than targeting $2024\times2024$ directly.

Ethical?

The dataset that was used to train the upscaling neural network was generated with Blender's Geometry Nodes. No artist's work was used in the process.

Roadmap

• Generate procedural data-set with Blender Geometry Nodes.
• Create a slim network architecture.
• Train the network.
• Create ray marching renderer (for SDFs).
• Implement inference on the WebGPU.
• Extract into a separate package, to be used as an upscaling step in other WebGPU projects.
• Create rasterizing renderer (polygonal geometry).

Development

Local setup

• Install Node.js
• Install pnpm
• Run pnpm install to install all dependencies
• Run pnpm dev to run the development server, making the web app available via the web browser.