A browser-only WebGPU playground for training tiny implicit neural representations on images, videos, and webcam frames. The app learns a mapping from pixel coordinates to RGB values, then renders the network output live beside the current target.

Every Frame an INR trains a small coordinate-based neural network directly in the browser with TensorFlow.js and the WebGPU backend. Pick a target image, video, or webcam stream, press Start, and watch the INR reconstruct the frame as training progresses.

The controls are meant for quick experiments:

• Switch between bundled images, videos, and webcam input.
• Change output resolution without leaving the page.
• Compare model sizes from Small to XL.
• Tune Fourier or Gabor coordinate features, activation, RMSNorm, and learning rates.
• Track epoch, loss, PSNR, learning rate, and preview speed while training runs.

Use any static server from the repository root:

python server.py --host 127.0.0.1 --port 8000

Then open:

http://127.0.0.1:8000/

Chrome or Edge with WebGPU support is recommended. If WebGPU is unavailable, the app will show that in the device indicator.

This is intentionally a small static app:

• index.html contains the app shell and controls.
• style.css contains the responsive UI styling.
• script.js contains media loading, TensorFlow.js model construction, training, rendering, and control state.
• vendor/tfjs/ contains the checked-in TensorFlow.js browser bundle.
• media/ contains the bundled targets used by the picker.

No build step is required.