Training Data

Overview

This documentation establishes principles for gathering high-quality training data for 3D printing models. Adhering to these guidelines ensures accurate, reliable data that improves model performance and minimizes risks and inefficiencies during the data collection process.

Data should represent diverse real-world scenarios while minimizing unnecessary costs, risks, and resource usage.

The goal is to collect training data that captures partially printed objects with:

• Obvious areas of successful printing
• Obvious areas of failure
• Obscure or subtle failure areas

During data collection, every effort should be made to reduce:

1. Damage to the Manufacturing Environment
2. Equipment (e.g., 3D printers, print beds, extruder nozzles, tools).
3. Surrounding items (e.g., tables, shelving, chairs, storage areas).
4. The physical workspace (e.g., walls, flooring, windows, ventilation systems).
5. Safety Risks to Personnel
6. Protect researchers and nearby staff from hazards such as moving parts, heat, fumes, or sharp debris.
7. Material and Resource Waste
8. Minimize unnecessary use of 3D printing filament and other consumables.
9. Avoid excessive wear on machinery.
10. Time Loss
11. Optimize processes to avoid unnecessary delays during printing or data recording.

The best way to create an effective 3D model is to sculpt a model designed to fail and then inject malicious G-Code into the G-Code.

Below is a list of strategies for producing failures. Popular digital models are modified using these strategies to antagonize print failures.

Gap & Shift

This method is performed by slicing the model at different points. This can vary, but the following models are cut in horizontal slices between $0\% \leq x \leq 50\%$ horizontally, and anywhere vertically.

Stable Diffusion

This method is performed by feeding a failure frame into Stable Diffusion, a generative artificial intelligence model. Prompt engineering is used to generate new images that depict various stages of print failures.