SkinDeepRESEARCHSteve Seguin

Live synthetic demo

Read an image. Return an action.

A small trained model looks at the pixels and returns a direction.

Back to finding where to click · Real screenshot recordings

This small trained model reads a 32×32 image and returns UP, RIGHT, DOWN or LEFT. It runs here in your browser. It learned synthetic arrow cards; it is not a general screenshot model.

Loading…

Loading a small set of trained weights. No large language model download.

Try your own arrow image

Use a dark arrow on a light background. Your image stays in this browser. The model always chooses one of four directions, even for images outside its training.

What actually runs?

A neural network with 32,932 trained parameters receives only the canvas pixels. A 32-unit hidden layer feeds four output scores. The highest score becomes the action. It does not receive the renderer's seed or answer, and it does not generate text.

Trained weights · Renderer and inference code

Try the same test

On 400 new filled-arrow cards, the trained model got 393 / 400 correct; four fixed image templates got 130 / 400. An unseen outline style scored 349 / 400 with the trained model. These are synthetic checks, not evidence of real-screen accuracy.

Training, controls and limitations · Sealed experiment · Every saved result