TL;DR: We reworked how on-site training is priced and tuned. Smarter defaults out of the box, pay only for the steps you run, new sample controls to play with, and a Train Further button to pick up where a training left off.
Why we're doing this
Training has grown up a lot, and the old setup hadn't kept pace. Pricing was tied to epochs, defaults were a one-size-fits-all guess, and once a training finished, that was it. So we cleaned it up. Nothing flashy, just a handful of changes that make training cheaper to dial in and give you more to tinker with.
What's new
Smarter defaults. New trainings start at around 2000 steps over 10 epochs, tuned with help from the folks who train the most. Better results out of the box, less guessing before your first run.
Pay per step, not per epoch. Cost now tracks the steps you actually run. Trim your steps and your cost comes down with them, so you can size a run to your budget instead of paying for a fixed block.
CFG Scale and Strength on samples. Two new knobs on your training samples. Push them around to see how your model behaves at your preferred settings.
Train further
Finished a training and wish you'd gone a little longer? Now you can. The new Train Further button picks up an existing training and keeps going, so you can add steps without starting from scratch.
Where to find it
It's all in the usual training flow. Start a new training to see the new defaults and per-step pricing, or open a finished one to Train Further.