How it works
1
Pick a trainer
Choose the base model you want to train on from the trainers table below. You select it with
trainer_id.2
Start training
Send your image URLs, a trigger word and your Hugging Face details to Start Training. You get back a training
id.3
Check the status
Poll Training Status with that
id every minute or two until it reads training_completed.4
Generate with your LoRA
When the status response’s
data.inference.status is ready, it gives you the exact text-to-image request to run your LoRA. See Use your trained LoRA.Trainers
Start Training has a ready-to-send request for each trainer.
Before you start
A Hugging Face account and write token
A Hugging Face account and write token
Your trained LoRA is uploaded to a repo on your Hugging Face account, so every trainer needs:
hf_username: your Hugging Face username.hf_token: an access token starting withhf_, created at huggingface.co/settings/tokens with write access.
Training images
Training images
- Send 1 to 100 image URLs in
images. 15 to 30 good images is a sensible target. - Each URL must download directly, without a login. Google Drive, Hugging Face and S3 links work as long as they are public. Every URL is checked before training starts, and the error names any image that failed.
- Use clear, varied images of one subject: different angles, lighting and backgrounds.
Wallet balance
Wallet balance
Training is billed per hour of GPU time, currently $5.14 per hour. The first hour is charged when the GPU starts on your run, and each further hour as it begins. You need at least one hour’s balance to start a run. Stopping a run ends the billing.
Training steps
training_steps controls how long the trainer learns from your images. A good starting point is 100 to 150 steps per image: 10 images × 150 = 1,500 steps. Fewer images need more steps per image; more images need fewer.
Runs are capped at 2,000 steps. A higher value is trained at 2,000.
Training status
The status endpoint reports one of three values inmessage and data.status:
Each trainer runs one job at a time, so a new run can wait in the queue behind others. It still reads
training_started while it waits, and billing only starts once the GPU picks it up. A run that has not started within 6 hours fails.Use your trained LoRA
For Flux LoRA, Z-Image Turbo and Stable Diffusion runs, ModelsLab loads your LoRA for inference automatically after training. Keep polling the status endpoint and watchdata.inference:
When it is
ready, data.inference.api holds the full request to send. It calls the LoRA text-to-image endpoint with the base model as model_id and your LoRA as lora_model:
json
data.inference.playground_url opens the same model in the ModelsLab playground.
The LoRA file itself is always in your Hugging Face repo at
https://huggingface.co/{hf_username}/{repo}, where repo is the training_id you sent, or the training id if the trainer does not take one.
Endpoints
Start Training
Queue a LoRA training run on any trainer.
Training Status
Check progress and get the request that runs your LoRA.
Stop Training
Stop a queued or running run and end its billing.

