How to train a LoRA on your own character

A LoRA is the answer when reference images stop being enough — when you need one face, one product or one style to survive angles, lighting and prompts that a reference photo cannot carry. On OpenModels training runs on the same GPU that serves everything else, which is why it is a Creator-plan feature, why it is metered on GPU time rather than sold as a flat run, and why it is quoted before it starts.

Where training actually is

This is the part that trips people up, so it goes first. There is a Character Soul entry on the image page, and it is a coming-soon placeholder — the training panel behind it is not open to accounts. The route that genuinely works today is the Agent section under /llms, where the platform's tools are driven in conversation, and training is four of those tools.

In practice that means you ask for it. "What LoRA options are there?" lists the base models, the categories and the step ranges. "Stage these twenty photos as a character dataset" builds the set. "Train it, dry run first" gives you a price. Then you say go. The Agent section is Creator and up in Beta, and the training tools carry the same Creator gate on the server, checked before anything with a side effect happens.

Which plan: LoRA training needs the Creator plan or higher. The gate is enforced on the server before the trainer is even started, so a lower plan is told no rather than being left with a half-started run. The charge comes out of the ✦ gold balance your plan grants or a credit bundle buys, not out of the refilling ⚡ green wallet.

What you can train it on

Base models and what each accepts
Base model	Kind	LoRA types	Dataset	Default steps
Z-Image Turbo	Image	Character, style	5 photos minimum	3,000
Krea 2 Turbo	Image	Character, style	5 photos minimum	3,000
LTX 2.3	Video	Character, action, style	Photos for characters, clips for action and style	3,000
MiniMax H3	Video and audio	Character, action, style	3 clips minimum	1,500
MiniMax H3 Ref2V	Reference to video	Character, action, style	3 clips minimum	2,000

The pairing to know is that a Krea 2 Turbo LoRA runs on Krea 2 Turbo — the plain text-to-image checkpoint — and not on Krea 2 Identity, which already carries its own identity adapter. Training the strongest identity model on the platform and then having nowhere to generate with it is a real trap, and it is why the plain checkpoint exists as a separate row in the picker.

Steps run from 250 to 6,000. Low counts are for testing that the dataset and trigger word behave; the defaults above are what the recipes actually recommend. You also get up to six sample prompts, which the trainer renders as it goes so you can see the adapter forming rather than waiting to find out at the end.

Training a character LoRA

Collect the dataset. At least five photos for an image LoRA; more is better and twenty is a sensible target. Vary angle, distance and lighting, keep the subject unobstructed, and avoid ten near-identical frames from one burst — the adapter learns the repetition rather than the subject.
Ask the Agent what the options are. The options tool is free and returns the base models, the training categories, the step ranges and the current billing rate. It is the right way to pick a base model, because the recipes differ per model.
Stage the set. The dataset tool downloads your photos into a draft set and tells you whether it is big enough to train. A set is either photos or clips, never both. Clips can also be cut directly out of a video you have already analysed, at the moments you name.
Pick a trigger word and get a quote. The trigger is the word that invokes the LoRA later — letters and numbers, at least three characters, and something that is not already a common word in your prompts. Then ask for a dry run: it returns the exact credits, the estimated GPU minutes and the ETA without starting anything.
Run it, then wait properly. Training claims the GPU, so a run may sit waiting while in-flight render jobs drain. That wait is normal and bounded — if the card cannot be claimed within the timeout the run is refunded and says so. Poll the status tool rather than watching the page.

What it costs, and why it is metered

Training is billed at 1,000 credits per GPU-hour, estimated from your base model, step count and dataset, charged up front at that estimate. It used to be a flat fee per run, and a flat fee is indefensible when a 250-step test and a 6,000-step run that owns the card overnight are the same price. Metering on GPU time is the version that is fair in both directions.

For scale: the default 3,000-step Z-Image Turbo run estimates a bit under two GPU-hours, so it quotes in the region of 1,500 to 2,000 credits. That is more than a Creator plan's monthly credit grant, so a real training run generally means buying a credit bundle as well as holding the plan. Saying that plainly is more useful than a per-hour rate you have to do arithmetic on.

What it costs: The quote is the number you are charged — the same estimator runs for the dry run and for the real submit. A run that never starts is refunded in full, because what is being sold is GPU time and none was used. A run that starts and then fails is not automatically refunded in full, for the same reason.

One thing that is deliberately not billed: the first run on a base model whose weights are not yet on disk spends time downloading them — Krea 2's are 26.3 GB and take around half an hour — with the GPU idle. That transfer is included in the ETA you are shown and charged at nothing, because billing it would overcharge every run after the first.

Getting a LoRA that is worth the GPU time

Dataset beats steps. Twenty varied photos at 3,000 steps beats six photos at 6,000 steps, and costs less.
Vary the light. An adapter trained entirely on soft indoor light will insist on soft indoor light later.
Caption consistently. The Describe tool captions a whole set in one request, which is the fastest way to get descriptions that agree with each other.
Test small first. A 250-step run is cheap and tells you whether the trigger word and dataset are behaving before you commit to a full one.
Pick a trigger word nothing else uses. A trigger that is also an ordinary English word will fire on prompts you did not mean it to.

Which plan do I need?

Creator or higher. The gate is checked on the server before the trainer starts, and it is the same gate whether you come at it through the Agent tools or any other route.

Why can I see Character Soul on the image page but not use it?

That panel is parked — it renders a coming-soon placeholder rather than the trainer. Training that actually runs today goes through the Agent section under /llms.

How much does one run cost?

1,000 credits per GPU-hour, estimated from your settings. A default 3,000-step Z-Image Turbo run lands in the region of 1,500 to 2,000 credits. Ask for a dry run before committing — it returns the exact number for your set without starting anything.

How many images do I need?

Five is the minimum the dataset tool will accept for an image LoRA, and three clips for a video one. Five will train something; twenty varied photos will train something you actually want to use.

What happens if my run never gets the GPU?

It is refunded in full and told you so. Training holds the card, so a run can wait while in-flight render jobs finish, but that wait is bounded rather than indefinite.

Where does the finished LoRA appear?

As an option on the base model you trained it against — a Krea 2 Turbo character LoRA generates on Krea 2 Turbo, invoked by your trigger word. It belongs to your account rather than the public catalogue.