LoRA training requires Premium or above. Maker Mode removes the limit on how many you keep. See Credits & Plans.
What you can train
Model types
Setting up a training run
1
Upload your dataset
Upload individual images or a zip archive. There’s a minimum and maximum image count — the form tells you if you’re outside it.
2
Name it and set a trigger word
The trigger word is what you type in a prompt to invoke the LoRA. Pick something distinctive that won’t collide with ordinary vocabulary.
3
Choose model type and training style
See the tables above. Portrait for faces, Style when you’re teaching a look rather than a subject.
4
Set steps
More steps can mean better quality and cost more credits. Start moderate.
5
Options
Face crop for people, Is Style when teaching an aesthetic, a fine-tune comment, and resume-from-checkpoint to continue a previous run.
6
Start training
The job is queued and appears under Active Jobs.
Dataset quality
This is where results are won or lost — far more than in the settings:- Variety beats volume. Different angles, lighting, and backgrounds teach the subject; twenty near-identical shots teach the background.
- Consistent subject, varied context. The thing you’re training should be the only constant.
- Crop to the subject. Don’t make the model guess what it’s learning.
- Quality in, quality out. Blurry or low-resolution inputs produce a blurry LoRA.
- For style, use images sharing the aesthetic but not the subject — otherwise you train the subject by accident.
Credits
Training costs credits, and cost scales with step count on Flux Turbo. If your balance is short, training won’t start — you’ll be told to add credits and try again. Check your balance before starting a long run.Using your LoRA
Once trained, it appears in Generate for models that support custom weights. Include the trigger word in your prompt to invoke it.Starting and cancelling training jobs is app-only — you can read training status through the API but not start a job. See what you can automate.