SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation
TLDR
SketchDNN uses a joint continuous-discrete diffusion model to generate CAD sketches, achieving state-of-the-art FID and NLL on SketchGraphs.
Reasoning
The paper introduces a novel Gaussian-Softmax diffusion that handles heterogeneous primitive parameters and permutation invariance, with strong quantitative results. However, it focuses solely on sketch generation without addressing full CAD models or real-world applications beyond the dataset.
Read-first score
Read-first score 44.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.
Field roles
Rank sensitivity
Stability: volatile; rank range: 34.
Keyword Scores
Deep Analysis
Innovations
- Joint continuous-discrete diffusion process for CAD sketch generation
- Gaussian-Softmax diffusion: logits perturbed with Gaussian noise and projected onto probability simplex via softmax to create blended class labels for discrete variables
- Addressing heterogeneity of primitive parameterizations and permutation invariance of primitives in CAD sketches
Methodology
SketchDNN is a generative model that jointly models continuous parameters and discrete class labels of CAD sketches using a unified continuous-discrete diffusion process. The core Gaussian-Softmax diffusion perturbs logits with Gaussian noise and applies softmax to obtain blended discrete labels. The model is evaluated on the SketchGraphs dataset using FID and NLL metrics.
Key Results
SketchDNN reduces FID from 16.04 to 7.80 and NLL from 84.8 to 81.33, setting a new state-of-the-art on the SketchGraphs dataset.