The team behind the open-source PRX text-to-image model has published a detailed logbook of their training experiments, and it’s a rare look at what actually works versus what just sounds good in a paper. Moving past the architectural decisions of their first post, they’ve now set up a clean, reproducible baseline for their 1.2-billion-parameter model and started testing popular training efficiency techniques one by one. The goal is brutally simple: find what improves convergence or raw training speed.
Their benchmarking scoreboard relies on a mix of metrics including FID, CMMD, and DINO-MMD, alongside raw samples-per-second throughput. It’s a practical, no-nonsense setup that intentionally avoids auxiliary objectives or architectural hacks in the baseline, so each intervention’s impact is crystal clear. The post groups their explorations into four buckets: Representation Alignment, Training Objectives, Token Routing and Sparsification, and Data.
The team, which credits community Discord exchanges for shaping the work, isn’t just testing these methods in isolation. They’re layering them to see which tricks remain useful when combined—a critical step most ablation studies skip. While the specific results from each bucket are still coming, the setup suggests a coming shakeout of trendy training tweaks.
A public ‘speedrun’ is also promised, where they’ll stress-test the best combined configuration under tight constraints. The full training recipe, including code for these experiments, will be released in the next post. For practitioners tired of hype cycles, this is a refreshingly empirical, ground-up approach to finding efficiency gains that actually move the needle.
