Photoroom outlines the data pipeline for its PRX model, which assembles training data from public and internal sources, re-captions images with a VLM, and converts the result into a streamable corpus.
- The team prioritizes dataset diversity over per-image aesthetics during pre-training to teach visual structure rather than just polish.
- Captions are generated by a VLM to ensure long, accurate descriptions that allow the model to treat text and logos as controllable attributes.
- Lance is used for feature engineering and curation due to its support for predicate pushdown and vector search, while Mosaic Data Shards (MDS) handles streaming during training.
- Text latents are computed on the fly using Qwen3-VL instead of pre-computing them, resulting in a 3–4% throughput cost but allowing flexible encoder changes.
- Images are stored as JPEG at quality 92, which was measured to be imperceptible compared to PNG and produced indistinguishable model generations.
This approach provides a lightweight, scalable starting point for pre-training a 7B model while maintaining flexibility in data format and text encoding.