How Photoroom Built a 7B‑Scale Pre‑Training Corpus from Public and Private Data
They mixed public and internal datasets, re‑captioned images with a vision‑language model, prioritized breadth over perfection, and kept filtering light to preserve noisy but learnable signals—key choices that let them train PRX efficiently. The pipeline turns raw images into a streamable corpus, using long, accurate captions to turn visual noise into controllable attributes.