To showcase these capabilities, we've implemented and made available an interactive viewer.
Check out our code at github.com/brepdiff/bre...
Interactive results are also available on our webpage! brepdiff.github.io/interactive/...
8/8

More interestingly, we show that by leveraging DDIM inversion, we can interpolate between B-reps, even when their topologies differ. We believe these capabilities represent a significant step toward enhancing traditional design workflows with generative modeling.
7/8

The explicitness of the masked UV grid representation enables users to freely manipulate geometry without being constrained by topological validity. This supports intuitive and advanced editing operations such as autocompletion and merging B-reps.
6/8

Masked UV grids are directly compatible with diffusion transformers. In the paper, we introduce a modified noise schedule and an asynchronous denoising strategy to more effectively learn the underlying distribution, achieving performance on par with existing multi-stage models.
5/8

We instead propose a single-stage generative model for B-reps. Our key insight is that the full topology can be inferred from the faces alone by representing each face as a masked UV grid defined by structured point samples of its surface.
4/8

The boundary representation (B-rep) is the predominant format for 3D models in CAD applications.
However, its highly irregular and complex structure presents significant challenges for generative modeling, often necessitating intricate multi-stage generation pipelines.
3/8

BrepDiff generates boundary representations (B-reps) with diffusion. Here, we show examples of denoising trajectories along with their post-processed CAD solids.
2/8

📢 I'm excited to share BrepDiff, our single-stage B-rep diffusion model, which enables unconditional generation and editing of B-reps.
This work will be presented at #SIGGRAPH2025 with @mingilikesmangos.bsky.social, @dongsuzhang.bsky.social, and Young Min Kim.
🌐: brepdiff.github.io
1/8