GlazyBench: A Benchmark for Ceramic Glaze Property Prediction and Image Generation
Developing new ceramic glazes is traditionally a slow, expensive process of trial and error, often leaving independent artists without a reliable way to predict how a recipe will look after firing. GlazyBench is a new research initiative that provides the first large-scale, standardized dataset to help bridge this gap using artificial intelligence. By collecting and cleaning thousands of real-world glaze recipes, the project aims to train AI models to predict the final appearance of a glaze—such as its color, transparency, and surface texture—directly from its chemical ingredients, and then generate visual representations of those results.
Building a Standardized Dataset
The researchers created GlazyBench by processing 23,148 real-world glaze recipes from the Glazy platform. Because community-contributed data can be inconsistent, the team implemented a rigorous cleaning process. They removed duplicates, handled outliers, and standardized firing parameters like temperature (cone ratings) and kiln atmospheres. To ensure the data is useful for AI, they organized the chemical information into Unity Molecular Formula (UMF) representations and used machine learning models to verify the accuracy of color labels, resulting in a high-quality, reliable foundation for training models.
Predicting Glaze Properties
The benchmark evaluates how well AI can predict the outcome of a glaze recipe before it is ever fired. The researchers established baselines using traditional machine learning models, such as Random Forest and CatBoost. These models analyze the chemical composition and firing conditions to predict four key properties: transparency, surface finish (e.g., glossy or matte), color family, and specific RGB color values. The results show that while tree-based ensemble models like CatBoost perform well at identifying surface and transparency trends, predicting precise color categories remains a significant challenge, highlighting the complexity of ceramic chemistry.
Generating Visual Representations
Beyond predicting numerical properties, GlazyBench introduces a two-step image generation pipeline. First, the system predicts the physical characteristics of the glaze. Second, it uses these characteristics as conditions to generate a 128x128 pixel image of the glaze’s appearance. To support this, the team manually curated a dataset of high-quality images and used segmentation and quality-patch extraction to ensure the training data was consistent. This pipeline allows for a complete validation process, moving from raw chemical ingredients to a final visual preview.
Research Challenges and Future Directions
The experiments conducted in this study demonstrate that while AI-assisted material design is a promising field, it is also highly complex. The current results show that there is still a significant gap between model predictions and the nuanced reality of ceramic firing. The researchers note that the severe class imbalance in surface types—where some finishes are much more common than others—and the difficulty of capturing accurate color data from non-standardized photos are major hurdles. By providing this standardized benchmark, the authors hope to encourage further research into more robust models that can handle the variability of real-world studio environments.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!