Local Minimum Pictures presents

How Wrong Am I?

a film by Learning Rate · twenty-five episodes, every Sunday

Same method as Episode 25, trained on a drawing.

The emblem is drawn by a diffusion model: Episode 25’s small network (two hidden layers of 128), trained offline on 10,000 points sampled from the emblem (seed 1919), 30,000 Adam steps in PyTorch on a CPU, about 2 minutes. Here it runs on your GPU: 100,000 grains, 100 denoising steps, then again from a new seed. 97% of the grains land inside the emblem; 8% of the static did.