SIBGRAPI 2026 · Tutorial
Slides and notebooks for Beyond Watertight Geometry: Neural Unsigned Distance Fields as a General Surface Representation. Two slide decks and two notebooks that run on a plain CPU.
A wild TUTORIAL appeared!
The Level-Set Pokémon
Block 1 slides. Carries a bulb of nested level sets on its back. Hatched watertight, it learned to drop the sign: SDFs, UDFs, gradients, projection and Marching Cubes, from the ground up.
The Extraction Pokémon
Block 2 slides. Its tail burns as a low-poly mesh. It sorts the methods by where they intervene: points, images, alternative representations and meshing, from CAP-UDF to MeshUDF and DCUDF.
The Tensor Pokémon
Block 1 notebook. Its shell is a grid of tensors. Ten labs that follow the Block 1 slides, from |s| to an end-to-end neural UDF in 3D, plus a bonus Stanford Bunny.
The Gradient Pokémon
Block 2 notebook. Stores charge in the ∇ on its belly and releases it through autograd. Simplified labs for CAP-UDF, NeuralUDF and DUDF, each one right after its slides.
No install needed: press RUN ONLINE on a notebook to open it in Google Colab. To run locally instead (CPU, a few minutes per LAB), use Python 3.10 or newer:
python -m venv .venv
source .venv/bin/activate
pip install torch numpy scipy matplotlib \
scikit-image plotly jupyter
jupyter labThe slides announce each LAB when it is time to switch to the notebook, and each LAB
ends with the slide to go back to. Run LAB 0 first, and keep
FAST_MODE = True to follow along live.
Block 1 goes from SDFs to neural UDFs, gradients, projection and Marching Cubes. Block 2 covers points, images, alternative representations and meshing.