tkb77/oekaki-sketch-diffusion

Oekaki sketch diffusion: a tiny class-conditional line-drawing generator

A 5.0 M-parameter DDPM that draws 64×64 line drawings of five categories: circle, square, triangle, star and smiley face. In the Oekaki VLA project, it imagines the picture that the robot then draws (tkb77/smolvla-oekaki-goal).

Code: https://github.com/tkb77/oekaki-vla (oekaki/generator/)

Model

• U-Net DDPM, 200 timesteps, base width 64.
• Classifier-free guidance (scale 3.0); weights are the EMA weights.
• Sampling: DDIM, 50 steps; the intermediate images can be returned for animation.
• Trained for 100k steps, batch 256, on 60,000 Quick, Draw! drawings per category: recognized drawings, at most 8 strokes, rendered with the project's common line renderer. A 10 % split by key_id is held out.
• Adapted from lcccluck/sketch-model (Apache-2.0).

Quality

• No empty images in 1000 samples.
• A separate judge CNN (99.1 % test accuracy on real drawings) recognises 99.9 % of the generated drawings.

Files

• last.pt: a PyTorch checkpoint, a pickle. It holds the EMA weights, the class names and the sampling and clean-up settings. Load it only with the code repository: oekaki.generator.imagine.Imaginer.
• config.json: the training configuration.

from oekaki.config import resolve_ckpt
from oekaki.generator.imagine import Imaginer

im = Imaginer(resolve_ckpt("tkb77/oekaki-sketch-diffusion", "last.pt"))
out = im.generate("star", seed=0)  # out["image"]: 64x64 uint8, out["segments"]: strokes for the robot

License

Apache-2.0. Trained on the Quick, Draw! dataset (Google Creative Lab, CC BY 4.0, https://github.com/googlecreativelab/quickdraw-dataset).