Photoroom Reveals Data Strategy Behind the PRX Model

Covers the key data pipeline strategies behind the performance of Photoroom's PRX model.

Data Collection and Composition Principles

• Diversity-Focused Pre-training: In the early stages, since learning visual concepts and composition matters more than aesthetic polish, the team avoided excessive filtering and instead focused on securing coverage and diversity in the data.
• Hybrid Data Sources: A mix of public datasets and internal datasets was used, leveraging existing quality-filtered datasets to improve efficiency.
• Long Captioning Strategy: Instead of short captions, Long Captions that describe every element of an image in detail were used, guiding the model to learn even noise within images (logos, text, etc.) as controllable attributes.

Data Format and Engineering

• Mosaic Streaming & MDS: The Mosaic Data Shards (MDS) format was used for flexibility and performance in distributed training.
• Use of Lance: To compensate for the rigid structure of MDS, the columnar data format Lance was used during the feature engineering and data curation stages for efficient data manipulation.