From the source
Photoroom describes a 24-hour training speedrun for a text-to-image model using a combination of techniques such as x-prediction in pixel space, perceptual losses, token routing with TREAD, and representation alignment with REPA, and open-sources the code.
