You generate a face, a product or a look you love, and the next prompt gives you someone else. That drift is the problem a LoRA solves. This guide explains what a LoRA is in plain terms, what it can and cannot learn, and how to train one for your own subject without a local GPU.
Quick answer: A LoRA (Low-Rank Adaptation) is a small add-on trained on top of a frozen AI model. It teaches that model one specific thing, such as a person, a product or a visual style, so you can call it back with a trigger word instead of hoping a prompt lands. You train it from a set of your own images (or clips, for video) and load it at generation time.
1.What does LoRA mean in AI?
LoRA stands for Low-Rank Adaptation. The method was published by Microsoft researchers in 2021: instead of retraining every weight in a large model, you freeze the original weights and train two small matrices per layer that nudge the output. In the paper's GPT-3 experiments, that cut the number of trainable parameters by 10,000 times and the GPU memory needed by 3 times, with no added delay at inference (Hu et al., arXiv 2106.09685).
Image and video generators adopted the same idea. A "LoRA" in an image tool is the small file that results from that training: the base model stays untouched, and the LoRA is loaded on top of it when you generate.
2.How does a LoRA work, in practice?
- Pick a base model. The LoRA is trained against one specific model, for example Krea 2 Turbo or WAN 2.2 Video.
- Show it examples. 10 to 30 sharp images of one subject or one style is a typical image dataset; video trainers learn from short clips.
- Tie it to a trigger word. The trigger word becomes the handle you type later ("ohwx person", "brandlook style").
- Load it when you generate. The prompt says what should happen; the LoRA supplies what the subject or style looks like. A strength setting controls how hard it pulls.
Illustration: a style LoRA learns a look from a consistent set, then carries it into new images.
3.What can a LoRA learn?
- A person or character. The same face across new outfits, places and poses. This is the most common use.
- A product. One bottle, shoe or package, reproduced accurately in new scenes for listings and ads.
- A visual style. A brand palette, a film grade, an illustration technique.
- A motion or camera move (video). Video LoRAs learn from clips, so they can capture how something moves, not just how it looks.
What a LoRA does not do well: fix a weak dataset, learn a subject from two photos, or work on a different base model than the one it was trained on.
4.Are LoRAs model specific?
Yes. A LoRA only loads on the base model family it was trained for. An adapter trained on Krea 2 Turbo will not run on FLUX.1 Dev, and a video adapter trained for text-to-video will not load in an image-to-video run of the same model. Choose the base model you will actually generate with before you train.
5.LoRA vs prompting vs reference images
| Approach | Good for | Weak at |
|---|---|---|
| Prompt only | One-off images, exploring ideas | Keeping the same face or product across a series |
| Reference image in the prompt | A handful of consistent shots, no training time | Holding identity or style across dozens of outputs |
| Trained LoRA | Series, catalogues, recurring characters, a house style | Needs a clean dataset and a training run |
If you need the same subject more than a few times, a LoRA usually saves time and credits on every generation after the first.
6.How do you train a LoRA without a GPU?
Local training on your own graphics card is free in money but not in setup time. Open-source trainers ship example configurations sized for 24 GB cards (AI Toolkit on GitHub). If you do not have that hardware, a browser trainer runs the job in the cloud.
In Imagera's LoRA Trainer you:
- Pick a trainer. Image trainers include Krea 2 Turbo, Z-Image Turbo, FLUX.1 Dev, FLUX.2 Klein 9B, Qwen Image 2512, WAN 2.2 Image and Ideogram 4.0; Qwen Edit 2509 learns an edit from before/after pairs; video trainers include LTX-2.3 Character, LTX-2.3 Motion, WAN 2.2 Video and four MiniMax H3 modes.
- Upload the dataset. Each trainer shows its own guidance, for example "15–30 sharp photos of one subject or style" for Krea 2 Turbo.
- Set the trigger word and steps. Every trainer has a 1,000-step minimum, so a first run is not undertrained.
- Check the credit cost shown before you start, then train. Most image trainers finish in about 8 to 20 minutes; video trainers take longer.
When it finishes, the model sits in your library and you generate with it or open it straight in Sandbox. The step-by-step dataset guide is in how to train a LoRA online.
7.How strong should a LoRA be?
Start around 0.7 to 0.8 for a person or product and 0.5 to 0.7 for a style. Too low and the subject drifts; too high and colours burn or the subject leaks into the background. When you stack two or three LoRAs, lower each one, because the strengths add up. Generate a small batch at two strengths and keep the one that stays recognisable without breaking the scene.
8.What does it cost?
Training and generation are paid in credits. The trainer quotes the exact cost of a run before you start; it depends on the trainer, the number of steps and options such as training resolution. Plans start at $19.99/mo for 240 credits a month on the Starter plan, and credit packs start at $39.99 for 350 credits. See pricing for the current ladder.
