Photoroom describes the data pipeline behind its PRX text-to-image model, assembling training data from public and internal sources while prioritizing diversity over perfection. The team uses Lance for dataset exploration and curation, then re-captions all images with vision-language models such as Qwen3-VL-8B to produce long, detailed captions that improve model quality. Processed data is converted into Mosaic Data Shards for distributed training, with filtering, deduplication, and aspect-ratio bucketing applied to optimize compute efficiency.
