Boundary representation (B-Rep) generation is a fundamental task in computer-aided design, yet the direct synthesis of high-fidelity and structurally valid B-Reps remains a major challenge. Existing deep generative methods suffer from two forms of brittleness: representation brittleness, caused by padding noise and feature contamination in the latent space, and generation brittleness, stemming from sequential error propagation and a train-inference mismatch due to non-differentiable validity enforcement. We propose HiFi-BRep, a novel framework that addresses these limitations through two synergistic contributions. First, a topology-aware encoder constructs a high-fidelity latent representation by eliminating padding via learnable queries and preventing feature contamination with topology-guided attention. Second, a single-stage decoder jointly predicts geometry and topology in parallel, embedding core manifold constraints as a differentiable learning objective. This design ensures mutual guidance between geometry and topology while avoiding cascaded errors. Extensive experiments show that HiFi-BRep significantly outperforms state-of-the-art methods in both structural validity and geometric fidelity, providing a robust solution for high-quality B-Rep synthesis. Code and models are publicly available at https://github.com/1nnoh/HiFi-BRep.
Alex: Welcome to another episode of ResearchPod.
Sam: Today we're looking at a new framework called HiFi-BRep, which tackles a specific problem in computer-aided design — the kind of 3D modeling software used to design everything from car parts to aircraft components. The challenge is getting AI to generate those shapes reliably, without producing files that crash the manufacturing software.
Alex: So the issue isn't just making shapes that look right — they have to be structurally valid at a deeper level?
Sam: Exactly. Think of it like the difference between a drawing of a box and an actual box you could fold and seal. Older AI methods often produce shapes that look fine on screen, but when you hand them to the manufacturing software, it finds gaps, missing surfaces, or edges that don't connect properly — and the whole thing falls apart.
Alex: Why do older methods keep running into that problem?
Sam: Two reasons. First, the way they store shape information is fragile — a bit like trying to keep a careful record on a notepad that keeps getting smudged. Second, they build shapes in multiple stages, one step feeding into the next, so if something goes wrong early on, the error gets passed down and amplified. And crucially, the system only checks whether the shape is valid at the very end — by which point it's too late to fix anything.
Alex: So HiFi-BRep tries to solve both of those at once?
Sam: That's the idea. On the storage side, instead of using blank placeholder slots to handle shapes of different sizes — which introduces noise — it uses what you might call "smart query slots" that actively gather only the information they need. And it applies a strict rule about which parts of the shape are even allowed to communicate with each other during that process.
Alex: Can you make that concrete?
Sam: Sure. Imagine you're assembling a 3D object out of flat panels and curved edges — like folding a cardboard box, but more complex. Each panel needs to meet its neighbors perfectly. The system keeps information about faces and edges in two separate channels, and only allows them to interact when they're actually adjacent. A face panel can only communicate with the edges that physically touch it — not with some distant edge on the other side of the model. That keeps the data clean.
Alex: And on the generation side — how does it avoid the cascading error problem?
Sam: Instead of building the shape in multiple sequential steps, it predicts the geometry and the structural connections all at once, in a single pass. And baked into that process is a rule drawn directly from solid geometry: every edge in a valid 3D solid must belong to exactly two faces — no more, no fewer. The system is trained to enforce that rule as part of its core objective, rather than checking for it as an afterthought.
Alex: So it's learning the laws of solid geometry, not just learning to imitate shapes.
Sam: That's a good way to put it. And because it's doing everything in one pass rather than chaining steps together, it also runs considerably faster than comparable methods, according to the paper.
Alex: Does the paper say anything about where the system still falls short?
Sam: It does, and the authors are fairly candid about the limits. The current version assumes shapes are closed and watertight — like a sealed container with no holes. It doesn't yet handle open-boundary shapes, large assemblies made of many separate parts, or more complex geometric configurations.
Alex: What causes failures even within those boundaries?
Sam: The system handles the bulk of the structure, but it still hands off the final surface trimming to a traditional mathematical engine — an external CAD kernel. If there's a disagreement at that trimming stage, or if the geometry produces a very thin, unstable patch, the kernel will reject the face. So the AI does the heavy lifting, but it remains dependent on conventional tools for that last step.
Alex: And what does the paper suggest as the path forward?
Sam: The authors point toward two main directions: more flexible decoding that can handle variable-length shapes without fixed budgets, and techniques to suppress the junction errors that currently get passed on to the external kernel. The goal is to close that gap so the system can handle more complex, real-world design cases.
Alex: So the current version is a meaningful step, but there's still a clear distance between this and a fully autonomous design pipeline.
Sam: That's a fair summary. The core contribution is showing that building structural validity into the generation process itself — rather than checking for it afterwards — produces more reliable results. Whether that approach scales to the full complexity of industrial design is the open question the authors leave for future work.
Alex: Thanks for walking us through it. And thanks to everyone listening to ResearchPod.