Zach Dive, a prominent figure in the tech community, recently announced an interactive demonstration of "BrepDiff," a novel single-stage diffusion model for generating Boundary Representations (B-reps), via a tweet on July 7, 2025. The announcement, which included a link to the demo at brepdiff.github.io/interactive/interactive.html, highlights a significant advancement in 3D modeling. Developed by a collaborative team from Seoul National University and MIT, BrepDiff is slated for presentation at the prestigious ACM SIGGRAPH Conference Papers (SIGGRAPH 2025).
Traditional methods for generating B-reps, widely used in CAD software, often rely on complex multi-stage processes that limit user interaction and prevent real-time edits due to non-deterministic dependencies. BrepDiff addresses this by introducing a simplified, single-stage approach. This innovation allows for intuitive and direct manipulation of geometry, offering capabilities such as shape completion, merging, and interpolation, which were previously challenging.
At its core, BrepDiff utilizes a masked UV grid representation, which transforms each B-rep face into a grid of 3D point samples and visibility masks. This representation is then processed by a diffusion transformer. The model employs an asynchronous and shifted noise schedule to enhance training signals, enabling it to accurately capture the distribution of UV grids. While the topology is not explicitly encoded, a valid solid B-rep is reconstructed through a robust post-processing step.
The explicit nature of the masked UV-grid representation empowers users to freely manipulate surface geometry without being constrained by topological validity. The interactive demo, as highlighted by Zach Dive's tweet, vividly illustrates these capabilities, demonstrating how users can perform complex editing sessions, including interpolating between objects, merging shapes, and auto-completing designs. This flexibility is a key differentiator from existing solutions.
BrepDiff's performance is reported to be on par with state-of-the-art cascaded models, yet it offers significantly more complex and diverse manipulations of both geometry and topology. Its upcoming presentation at SIGGRAPH 2025 underscores its recognition within the computer graphics community. This development is poised to offer CAD designers and engineers a more flexible and efficient tool for creating and modifying intricate 3D models, potentially streamlining design workflows across various industries.