MODEL TRAINING
Teach the model your look
Train a custom FLUX LoRA on your own images, for a style you want to reuse or a subject that has to stay exactly itself. No local GPU, no environment to set up, and the result lands in the same form you generate from.
What a LoRA gives you that a prompt cannot
A prompt can describe a look. It cannot reliably reproduce one. Describing your own visual style in words gets you somewhere near it and then drifts, because every generation re-interprets the description from scratch.
A LoRA moves that knowledge out of the prompt and into the model. Once trained, invoking it with a trigger word applies what it learned consistently, and the prompt goes back to describing the scene rather than the treatment.
The usual reason to train one is repetition. If a look has to appear across dozens of images, or a product has to appear exactly as it is, training pays for itself against the time spent fighting a prompt every single generation.
The training form, step by step
Seven decisions, and the training set is worth more than the other six combined.
Model Training is a generation type like any other in the dropdown.
Select Model Training
Training sits in the same type dropdown as generation, which is worth noticing because it means no separate tool or environment is involved. You pick Model Training and the form becomes a training form.
The output of this process is a LoRA that appears in the LoRA picker of the Text to Image form, so training and using what you trained happen in the same workspace.
Style or subject
A style LoRA learns how things look, meaning the palette, the brushwork, the grain and the lighting habits shared across your images. A subject LoRA learns a specific thing, such as one person, one product or one location.
Pick the wrong one and the training works against you. Feeding twenty photos of the same person into style training teaches the model that your style is that person.
Style or subject. This choice changes what the training optimises for.
Training images are supplied as a ZIP file.
Upload the training set
Training images go up as a single ZIP. The set matters far more than any parameter on this form, since a LoRA can only learn what the images consistently show.
Consistency is the thing to aim for on the trait you are teaching, and variety on everything else. For a subject, vary the angle, lighting and setting while keeping the subject identical.
Set a trigger word
The trigger word is how you call the LoRA once training finishes. Including it in a prompt activates what you taught, and leaving it out means the base model behaves normally.
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.
The trigger word. This is what you type in a prompt to invoke the LoRA.
The name that identifies this LoRA in the picker later.
Name the LoRA
The name is how you find this model in the LoRA picker later. Once you have trained several, generic names become a real problem.
Something that records what it was trained on and at what step count saves time, because the alternative is training again to find out.
Choose the step count
The minimum is 1,000 steps and cost scales with the number, so this is the parameter that decides what a training run costs. More steps means the model fits your images more closely.
Fitting more closely is not automatically better. Past a point the LoRA reproduces your training images rather than generalising from them, which shows up as an inability to place the subject in any new setting.
Training steps, minimum 1,000. Cost scales with the count.
The remaining training options before starting the run.
Review and start training
The remaining options are set before the run starts. Training then proceeds without further input, and the finished LoRA appears in the LoRA picker of the Text to Image form.
From there it applies like any predefined style, with a strength control governing how heavily it sits over the base model.
Settings reference
What the training form asks for and how each choice affects the trained model.
Training images
A ZIP file of the images to train on. The single biggest factor in the result, ahead of every parameter on this form.
ZIP archive
Model type
Style teaches a visual treatment shared across the set. Subject teaches one specific person, product or place.
Style or subject
Trigger word
The token that invokes the LoRA in a prompt. Pick something rare so it carries no existing meaning for the base model.
Free text, ideally uncommon
LoRA name
The label shown in the LoRA picker after training. Descriptive names save real time once several exist.
Free text
Training steps
How long training runs. Cost scales with the count, and very high counts cause the LoRA to reproduce the training set rather than generalise.
Minimum 1,000
Building a training set
The rule that decides most outcomes is simple. Be consistent on the thing you are teaching and varied on everything else. A subject LoRA needs the same subject across many angles, lighting conditions and settings. A style LoRA needs the same treatment across many different subjects.
When a set is varied on the wrong axis, the model learns the wrong thing. Twenty images of one person in one room teaches the room as firmly as the person, and you will find you cannot get that person anywhere else.
Step count trades cost against fit. Too few and the LoRA barely registers. Too many and it memorises rather than generalises, which shows up as output that looks like your training images regardless of what the prompt asked for. Starting near the minimum and raising it on a second run costs less than starting high and discovering the model overfitted.
Related tutorials
Background on the base models and the alternatives to training.
LoRA training questions
What is the difference between style and subject training?
Style teaches how things look across your whole set, such as palette, texture and lighting habits. Subject teaches one specific thing, such as a person or a product. Choosing the wrong mode teaches the model the wrong lesson, so a set of one person under style training makes that person your style.
How do I pick a trigger word?
Choose a rare or invented token the base model has no associations with. Using a common word means competing against everything the model already believes about that word, which dilutes what you trained.
How many training steps should I use?
The minimum is 1,000 and cost scales from there. More steps fit your images more closely, but past a point the LoRA starts reproducing the training set rather than generalising from it. The symptom is a LoRA that cannot place its subject in any setting it did not already see.
Where does the LoRA appear once training finishes?
In the LoRA picker of the Text to Image form, alongside the predefined styles. It applies with a strength control, so you can dial back how heavily it sits over the base model.
Do I need my own GPU to train?
No. Training runs on the platform, so there is no local environment to set up, no dependency management and no hardware requirement beyond a browser.
Can I use a LoRA someone else trained?
Yes. The LoRA picker accepts an external LoRA URL as well as the ones you train yourself and the predefined styles, so you are not limited to your own training runs.
What base model does training use?
Training produces a FLUX LoRA, so the result applies on top of FLUX generation.
Other ways to generate images
Model Training is one of seven generation types in the image workspace.
Video & LipSync
- Video Generator
- Text to Video
- Image to Video
- Start-End Frame to Video
- Draw to Video
- Motion Control
- Video Enhancer
- Video Upscaler
- Video to Video LipSync
- Audio to Video LipSync
- Image to Video LipSync
- Video FaceSwap
- Seedance 2
- Vidu Q3 Pro
- Gemini Omni
- Google Veo 3.1
- Kling 3.0 Pro
- Luma Ray 3.2
- LTX 2.3
- Happy Horse 1.1
- WAN 2.7
- Kling 3.0 Motion
- ByteDance Upscaler
- InfiniteTalk
- InsightFace