字生 / In your browser
A character, from noise.
Pick a character. Watch the model work from fresh noise.
starts here.
Generate a new sample, or inspect a recorded model run.
What’s running?
The original 880,385-parameter conditional U-Net and the repository’s 1,000-step DDPM update, with fresh Gaussian noise at each reverse step. The timeline records 101 actual states, including the starting noise. The first steps can still look noisy; recognizable structure often arrives later.
The faster DDIM options use the same weights but take a different, deterministic sampling path. Browser and PyTorch random-number generators differ, so the same numeric seed does not produce identical pixels across implementations.
These are synthetic printed-glyph samples from a 12-character pilot, not handwriting or pen-stroke paths. Recognition accuracy and unseen-font performance have not been measured.