John Femiani

John Femiani

Armstrong Professor of Computer Science at Miami University in the Department of Computer Science and Sowtware Engineering

SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation

Published in International Conference on Machine Learning (ICML), 2025

SketchDNN is a generative model for synthesizing CAD sketches that jointly models both continuous parameters and discrete class labels through a unified continuous-discrete diffusion process. The core innovation is Gaussian-Softmax diffusion, where logits perturbed with Gaussian noise are projected onto the probability simplex via a softmax transformation, facilitating blended class modeling for high-fidelity CAD sketch generation.

[ICML 2025 Poster by Sathvik](/images/SketchDNN%20Poster%20(2).pdf) Examples of the diffusion process-- random primitives are shown on the left, evolving towards realistic 2D CAD drawing exames on the right

Main Points

  • Mixing categorical (e.g. primitive type) with continuous (position, shape) information is hard, we present a novel solution.
  • Without care, categories do not evolve at the right speed, preventing realistic diffusion results.
  • We use a novel and simple representation that actually works to mix categorical and continuous variables
  • We had to use a different noise sechedule and a slightly modified inference process
  • Contrast this with autoregressive approached – diffusion works in many passes, allowing it to respect the many symmetries and alignments (tangent features, points that should conincide).

Recommended citation: Sathvik Chereddy, John Femiani, "SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation." International Conference on Machine Learning (ICML), 2025.