---
title: HiFi-BRep Generation
emoji: 📐
colorFrom: pink
colorTo: purple
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
short_description: Generate 3D B-Rep CAD structures with HiFi-BRep
python_version: "3.10"
startup_duration_timeout: 10m
---

# HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation

This Space demonstrates unconditional generation of 3D B-Rep CAD structures using the
[HiFi-BRep](https://huggingface.co/papers/2608.16485) diffusion model.

## How it works

1. A DiT (Diffusion Transformer) model generates latent representations via DDIM sampling
2. A VAE decoder transforms latents into B-spline surface/edge control points, bounding boxes,
   and face-edge adjacency predictions
3. The geometry is reconstructed as a 3D point cloud and mesh for visualization

## Model

- **Weights**: [1nnoh/HiFi-BRep](https://huggingface.co/1nnoh/HiFi-BRep)
- **Variant**: DeepCAD 4-30 (4-30 faces range)
- **Code**: [github.com/1nnoh/HiFi-BRep](https://github.com/1nnoh/HiFi-BRep)

## Note

The original reconstruction pipeline uses OpenCascade (pythonocc-core) for full B-Rep
solid construction and STEP file export. Since pythonocc-core is conda-only and not
pip-installable on Hugging Face Spaces, this demo reconstructs the geometry using pure
numpy B-spline evaluation and displays it as a 3D mesh. The diffusion + decoder pipeline
runs identically to the original code.