Training Studio: Watch It Train, Generate As It Goes
TL;DR:
You turn it on yourself. One switch in your preferences. Switch back whenever.
Watch it train. Progress, logs, and epochs showing up as they finish, instead of a notification when it's all over.
Generate with an epoch before the run is done. That one is new. You used to have to wait for the whole thing.
Compare epochs side by side to work out which one is actually the good one.
Nothing is saved until you press start, and nothing goes public unless you publish it.
We rebuilt the trainer. It's called the Training Studio, it's in beta today, and you can switch it on yourself.
Two things drove it: make setting up a run simpler, and let you actually follow one once it's going.
Here's the whole thing in four and a half minutes.
Following your run
The old way was set it up, walk away, find out at the end.
Now the page fills in while it works. The step count climbs, logs come through, and each epoch turns up with its sample images the moment it finishes. Your trainings list does the same, so you can see where everything is without opening anything.
And this is the part that didn't exist before: you can generate with an epoch while the run is still going. Epoch four lands while epoch seven is still training, and you can hit Generate on it right there. It opens the generator with that LoRA already loaded. If it's working, you know early. If it isn't, you know early.
When it's finished, Compare epochs puts every sample prompt against every checkpoint in one grid. Read across a row and watch one image evolve from the first epoch to the last, instead of opening ten tabs and squinting.
Setting one up
You say what you're making, not what to configure. Image, video or audio, then character, style, concept or effect. It suggests a base model from that, and you can take the suggestion or pick anything else.
Four ways to bring data in. Upload, pull from your own generations, import a zip, or reuse a dataset from a previous run.
Labeling is automatic and free, in the format the model actually wants. Tags for some, captions for others. You don't get asked.
Steps in plain terms. "each image seen ~100x (good for a character)" instead of a number you have to translate.
Up to five runs on one dataset, priced and started together.
You see the price before you commit, and a failed run is refunded automatically.
Nothing exists until you start. No half-built model page sitting in your profile.
Turning it on
It's off by default. Open your preferences, find the Features section, and flip the switch called Training Studio. It has a Beta badge on it.
After that, Train a LoRA and Training both take you there. Don't like it? Switch it off. The old one hasn't gone anywhere.
It's a beta
Runs stay for 30 days, so download or publish anything you want to keep. Krea 2 can't train its text encoder, and the field tells you so. Trigger words do nothing on a few of the bigger models.
Turn it on. Train something. Tell us what broke.