Teach the model your look

LoRA Trainer sits in the same Select Generation Type dropdown as every other generation type.

The image workspace has one dropdown for every generation type it offers, and training is one of the entries in it. Pick LoRA Trainer and the panel below switches from a generation form to a training form.

The same dropdown also holds LoRA Image to Image, which is where a LoRA trained here gets used afterward. Training and applying the result stay inside one workspace, just under different entries.

Two trainers are available: Z-Image LoRA Trainer at 1.25 credits per step, and Qwen Image LoRA Trainer at 1.00 credit per step. They train against different base architectures and are priced differently; neither is documented as producing better results than the other.

Switching trainers resets the learning rate field back to its default, since a rate tuned for one is not guaranteed to suit the other.

Two trainers, priced per step: Z-Image at 1.25 credits, Qwen at 1.00 credit.

The name that identifies this LoRA once training finishes.

The name comes right after the model choice: 3 to 40 characters, letters, numbers, spaces, underscores and hyphens only. A live preview under the field shows the standardized version it becomes.

This is the label you will be scanning for once you have trained more than one, so something that records the subject or style saves a search later.

Three input methods are available. Upload Images & Captions takes 4 to 30 image files, dropped in across as many batches as needed and each optionally paired with a same-named .txt caption, then packages everything into a ZIP in the browser once you click Package & Upload. Upload ZIP File takes a ZIP you already built, up to 1,000MB. External ZIP URL takes a direct, publicly reachable link to one instead.

Whichever method is used, the training set is what the LoRA actually learns from, ahead of every other field on this form. Consistency on the trait being taught and variety on everything else matters more than any parameter below.

Three ways in: individual images with optional captions, a ready-made ZIP, or a hosted ZIP link.

The trigger word. This is what you type in a prompt to invoke the LoRA.

The trigger word is 2 to 30 characters and is how you call the LoRA once training finishes. It is shown alongside the LoRA name wherever the trained model turns up later, so including it in a prompt is what activates what you taught.

Choose something the base model has no existing associations with. A rare or invented token works. A common word means you are fighting everything the model already believes about it.

LoRA rank is 4, 8, 16 or 32, with 16 recommended for most subjects. Higher ranks capture more detail at the cost of a slower run and a larger file; 32 is worth trying only if 16 is visibly missing something.

Learning rate ranges from 0.00001 to 0.01 and defaults to 0.0001, which suits most subjects and styles. Pushing it above 0.001 tends to distort output, and dropping it below 0.00005 makes training slow with little to show for it. Leave it at the default unless a previous run already gave a reason to move it.

LoRA rank and learning rate, the two parameters that shape how closely training fits the images.

Training steps, 500 to 10,000. Cost scales directly with the count, shown live above the button.

Training steps run from 500 to 10,000, defaulting to 1,000. Cost is the step count multiplied by the per-step rate, so this slider decides what the run costs as directly as any field on the form.

More steps means a closer fit to the training images, not automatically a better one. Past a point the LoRA starts reproducing the training images rather than generalising from them, which shows up as an inability to place the subject in a setting it never saw. Starting near the minimum and raising it on a second run costs less than discovering the first one overfit.