Code Execution Guide

Ⅰ. Overview

1. This guide explains how to use the code in the paper to predict polycube structures using the DDPM-polycube algorithm.¹

2. The implementation is mainly in Python and Blender, using Python third-party libraries and Blender's built-in libraries.

Environment/Library/Software	Version	Description
Blender	4.3.2	Built-in library bpy
python	3.9
numpy	1.23.5	Keep the same version to avoid issues
torch	1.10.1+cu102

3. Introduction to Blender Interface and Basic Operations

The image below shows the basic Blender interface. In this workflow, we mainly focus on:

• Yellow rectangle (Viewport): display the model and adjust the current view. In the yellow oval area, drag the axis gizmo to rotate, the magnifying glass to zoom, and the hand to pan.
• Red rectangle (Code Area / Text Editor): switch, browse, and run code. The booklet-like icon switches between code files; the triangle icon executes code.
• Green rectangle (File Area / Outliner): lists all current models. Use the eye icon to show/hide models. Double-click the model name to rename.
• Blue rectangle (File menu): import models, e.g., File -> Import -> FBX (.fbx).

4. Explanation of Subdirectories

Subdirectory	Purpose
diff_32_steps_stl	Stores the reconstructed 32-step diffusion triangle mesh result
pics	Stores images showing the diffusion process
src	Stores source code
testing_models	Stores FBX test models
tmp	Stores various intermediate files output during the process
training_data	Stores the constructed training data
v_cos	Stores point coordinates and connectivity for training
weights	Stores trained neural network weights

Ⅱ. Execution Process

The overall code has two parts: directly running Python files, and running Python-based scripts within Blender.

Preparation:

• Open the Blender file in the src directory (double-click to open).
• Import the model into Blender: File -> Import -> FBX (.fbx), choose a model from the model directory. Check the model name in the Outliner; if not test, rename it to test.

About parameters g1 and g2 (used when running file 2):

• When testing a model, select possible unit combinations based on the model's genus or other information.
• For two units, enter both g1 and g2. g1 represents the -x part of a 2x1 grid, and g2 represents the +x part.
• If the model consists of only one unit, set g1 = -1, and set g2 to the unit number you consider likely.

Example:

• For the “rod” model (genus=1), you might think the -x part is a cube and the +x part is a torus. Run file 2 with g1=0 and g2=3.
• For a genus 0 model, try a single basic cube: g1=-1 and g2=1.

Note: To select the best result, you may need to run multiple times with different reasonable unit combinations (e.g., for genus 1: g1=0,g2=3 or g1=0,g2=6, etc.).

Execution order:

Step	Specific Execution Content
1	Execute File 1 in Blender
2	Execute Python file 2: python 2_Testing.py --g1 <g1_value> --g2 <g2_value>
3	Execute Python file 3
4	Execute File 4 in Blender
5	Execute Python file 5 (optional)

In general, when testing a model:

• Compute genus-related information.
• Enumerate possible basic unit combinations based on the genus.
• Convert combinations to parameters g1 and g2 (context).
• Batch-generate polycube results using the program for these cases.
• Choose the best parameters and corresponding polycube based on the results.

We also provide model training code if you need to train.

Ⅲ. Results

• The polycube recognition results can be viewed in diff_32_steps_stl/step32.k (open with LS-Dyna) or diff_32_steps_stl/step32.stl (open with Blender or other 3D software).
• If you executed files 4 and 5, you can also see images of the diffusion process in the pics folder.

Note: Some code refers to the public sample code of “How Diffusion Models Work”.

References

Footnotes

1. Y Yu, Y Fang, H Tong, J Liu, YJ Zhang. DDPM-Polycube: A Denoising Diffusion Probabilistic Model for Polycube-Based Hexahedral Mesh Generation and Volumetric Spline Construction. arXiv:2503.13541 (2025). ↩