Code

Image models interpret your prompts through a language model. TokenCollider runs phrases through that same model and finds the six orthogonal directions they vary along most, using principal component analysis (PCA). Three of those directions become position in 3D, and three become color. You can see which phrases sit near each other, and how that changes from one layer of the model to the next. Pick any point in that space and export it as a file the image model samples from like a prompt.

Free and MIT licensed. Runs locally with Python and PyTorch.