PRX Part 4: Our Data Strategy – Building a Diverse Pre-training Corpus
Photoroom
Photoroom's PRX team details their data strategy: assembling a diverse pre-training dataset from public and internal sources, re-captioning all images with a VLM, and storing data in Lance for curation and MDS for streaming. They emphasize breadth over aesthetics, use JPEG at quality 92, and compute text latents on the fly with Qwen3-VL.
Photoroom's PRX team shares their data strategy for pre-training a 7B text-to-image model. They assemble training data from a mix of public and internal datasets, prioritizing diversity and breadth over per-image perfection. All images are re-captioned using a VLM for consistent, long captions. Data is stored in Lance format for feature engineering and exploration, then converted to Mosaic Data Shards (MDS) for distributed training. Images are encoded as JPEG at quality 92, which showed negligible quality loss compared to PNG. Text latents are computed on the fly using Qwen3-VL during training, incurring only a 3-4% throughput cost. The team built a UI to browse the dataset using full-text search and vector similarity, enabling qualitative assessment and early issue detection.
- Сокращения
- VLM = Vision-Language Model — визуально-языковая модель
- MDS = Mosaic Data Shards — мозаичные шарды данных
- JPEG = Joint Photographic Experts Group — объединённая группа экспертов по фотографии
- PNG = Portable Network Graphics — портативная сетевая графика
- PSNR = Peak Signal-to-Noise Ratio — пиковое отношение сигнала к шуму
- LPIPS = Learned Perceptual Image Patch Similarity — обученное перцептивное сходство патчей изображения
- UI = User Interface — пользовательский интерфейс
Source: Hugging Face blog —
original