PRX Part 4: Our Data Strategy

Source: HuggingFace Date: 2026-07-06 URL: https://huggingface.co/blog/Photoroom/prx-part4-data

Summary

Fourth installment of Photoroom’s technical series on PRX, their 7B-parameter text-to-image diffusion model, covering the data pipeline: mixed public/internal images re-captioned with Qwen3-VL-8B, converted to Mosaic Data Shards for streaming, with detailed captions doing more for quality (FID, CMMD, DINO-MMD) than aggressive filtering. Pragmatic engineering choices — on-the-fly text embeddings, JPEG-92 over PNG — kept the pipeline simple without sacrificing measured quality.

Implications

• Model capability clocks (open): a concrete, publicly documented recipe (VLM re-captioning + streamable shard format + light filtering) for training a mid-size diffusion model outside the frontier labs — useful ground truth for how open/mid-tier training pipelines are actually built in 2026.
• Local-model hardware fit: 7B is squarely in the range that fits consumer/prosumer hardware for inference; the data-pipeline detail (caption quality over volume, on-the-fly embeddings) is a reusable pattern for anyone fine-tuning smaller models locally rather than training from scratch.