PRX Part 4: Our Data Strategy

Photoroom Details Pragmatic Data Strategy for PRX Model

AI company Photoroom has released a technical deep-dive into the data pipeline powering its PRX image model, detailing a strategy that prioritizes breadth and engineering flexibility over aesthetic perfection during pre-training. In a new article, the team explains how they assemble a diverse corpus from public and internal sources, uniformly re-caption all images with a Vision Language Model (VLM), and stream it for training. This disclosure offers a rare look into the often-opaque data processing work that underpins the performance of modern generative AI models.

The team’s approach is guided by the principle that "pre-training is for breadth; fine-tuning is for taste," focusing the initial stage on teaching the model a wide range of visual concepts. To manage this large-scale data, Photoroom leverages a dual-format system: using the Lance columnar format for efficient querying, filtering, and exploration, and then converting to Mosaic Data Shards (MDS) for high-performance distributed training. The article also outlines several key engineering trade-offs made for the sake of agility.

• On-the-fly Text Encoding: Instead of pre-computing and storing text latents, the team computes them during training using the Qwen3-VL text encoder. This incurred a minimal 3-4% throughput cost but provided the flexibility to change encoders without rewriting the entire dataset.
• Image Format: Images are stored as JPEG at quality 92 rather than a lossless format like PNG. Internal tests showed this had an imperceptible impact on model output quality for pre-training while significantly reducing storage requirements.
• VLM Re-captioning: Every image was re-captioned to create a uniform standard of long, descriptive captions, which they found substantially improves sample quality by turning potential visual "noise" into promptable attributes.

By documenting its data strategy, Photoroom provides a practical blueprint for other organizations building large-scale models. The focus on leveraging existing open-source tools like Lance and Ray Data, coupled with deliberate trade-offs like on-the-fly encoding and JPEG compression, highlights a maturing industry trend towards sustainable and efficient MLOps. This pragmatic approach contrasts with narratives of limitless data and compute, showing how thoughtful data engineering can yield competitive results without requiring a complete reinvention of tooling.

Photoroom's PRX data strategy underscores a critical industry insight: a flexible, well-tooled data pipeline that prioritizes engineering velocity and dataset breadth is more valuable for foundational model training than a theoretically perfect but rigid and slow-to-iterate corpus.