Photoroom's PRX data report describes a storage choice that its experiments found adequate for training an image generator: high-quality JPEG instead of lossless PNG. The team trained two otherwise identical PRX models on the same images stored in the two formats and reports similar training metrics and visually comparable generations.
PNG files would have been three to ten times larger in its comparison. That makes compression a consequential storage decision for a large training collection. The finding concerns this training setup, whose source images were already mostly JPEG; Photoroom says it still uses PNG for other models sensitive to image artifacts.
Terms in this piece · Glossary
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
