T2I Model Training Optimization Guide

Photoroom has published experimental research results for efficiently training its open-source Text-to-Image (T2I) model, PRX. This research focuses on optimization in the training process, not the model architecture.

The key points are as follows:

• Experiment Purpose: Verify techniques that substantially contribute to speeding up model convergence, learning better representations, and stabilizing the training process.
• Baseline Setup: Based on the 1.2B parameter PRX model, an environment using only pure Flow Matching without additional objective functions is set as the baseline.
• Verification Method: Various recently proposed training techniques are reproduced under a consistent setup, and the performance changes when each technique is applied individually or in combination are analyzed through Ablation study.

Photoroom plans to release the full Training Recipe, including the experiment code, in the future.