Ceramix

Ceramix is a synthetic dataset of textured ceramic vessels with 625 samples designed for computer vision and machine learning research. The dataset contains images of procedurally generated ceramic vessels with controlled variations in geometry and decorative surface designs, providing a structured collection for studying visual representations. The motivation behind Ceramix is to support research that requires controlled visual data. Real-world image datasets often contain uncontrolled variations in lighting, backgrounds, viewpoints, and object placement, making it difficult to isolate the influence of object shape and surface appearance. Ceramix provides a consistent and reproducible dataset where these visual factors can be studied independently. Sample Distribution: Though more models were generated per vessel type during the creation process, only 5 models were hand-picked to achieve a balanced dataset. 5 vessel types. 5 models per vessel type. 5 design categories. 5 variation per design category. 625 in total. Generation method: N instances per vessel type are generated as blender models with Python script. The 125(5 x 5) unique design patterns are applied as skin over the generated models. The design applied blender models are rendered as high quality images of size 1024 x 1024. 5 images per (vessel type, unique design pattern) are hand-picked, rejecting the few absurd ones. Check the GitHub for the scripts used for generation and feel free to improve them. File naming convention: <design_variant><design_type><vessel_type>_<image_index>.png For example, 3_scribble_bowl_1.png represents the first selected rendered bowl generated using the third design variant from the scribble design category. Potential Applications: Visual similarity analysis Image embedding evaluation Computer vision benchmarking Synthetic data generation Evaluation of foundation models (e.g., CLIP, DINO, Vision Transformers) Reproduce or Extend: The complete dataset generation pipeline, documentation, and source code are available in the GitHub repository, allowing the dataset to be reproduced and extended.