Most guides on training a LoRA character jump straight to learning rates and network rank. That is the wrong order. If your source images are inconsistent, no setting will save the run. If your images are good, a fairly plain set of settings will usually get you a character you can generate over and over.
This guide walks through the whole process in the order you actually do it: gather images, caption them, train, test, fix, then use the result for something.
What a LoRA character is, and when you need one
A LoRA (low-rank adaptation) is a small file trained on top of a larger image model. It does not replace the base model. It nudges it, so that when you use a specific trigger word, the model produces one specific face, body and look instead of a generic one.
You need one when you want the same person to appear across many images and videos. A prompt alone will give you a different face every time, even with a detailed description. A LoRA gives you continuity, which is what makes a character feel like a person rather than a series of strangers.
You do not need to train anything yourself if you would rather use a hosted setup. Training, storing and prompting a character is handled for you in AI character generation, where you train the character once and then generate photos and video of the same person on demand. Train it yourself when you want full control over the dataset and the model, or when you already have a local setup you like.
What a LoRA will not fix
A LoRA learns what is in your images. It will not invent angles you never showed it, and it will not clean up a face that is blurry in every source image. Treat the dataset as the product.
Build the dataset before you touch a trainer
This is where the result is decided.
How many images
For a single character, a small, tight set usually beats a large, messy one. Somewhere in the range of twenty to forty images is a reasonable starting point. If you only have ten good images, train with ten rather than padding the set with bad ones. Every weak image teaches the model something you do not want.
What to vary
You want the model to learn the person, not the circumstances. So vary everything that is not the person:
- Angles: front, three-quarter, profile, slight up and down tilts
- Distance: close-up face, head and shoulders, half body, full body
- Expression: neutral, smiling, talking, eyes closed
- Lighting: daylight, indoor, soft shadow, harder shadow
- Background: plain walls, rooms, outdoors
- Clothing and hair styling, unless the outfit is part of the character
If every image is shot in the same room with the same lamp, the model will bake that room into the character and you will fight it in every prompt afterwards.
What to avoid
Cut anything blurry, heavily filtered, watermarked, or with another person's face in frame. Cut duplicates and near-duplicates from a burst. Crop out extra people rather than including them. If the character is meant to be the face of a business, keep the dataset clean from the start, because it is much harder to unlearn a bad habit than to avoid teaching it.
Get this part right and the rest of your build gets easier, because a consistent character is the foundation for everything described in how to make money with an AI influencer.
File prep
Resize to the resolution your base model expects, keep aspect ratios sensible rather than stretching, and name files in a way you can trace later. You will retrain at some point, and you will want to know which images were in which run.
Caption your images
Captions tell the trainer what is variable and what is the character.
Pick a trigger word
Choose a unique token that the base model does not already associate with something. An invented name or a short made-up string works well. Use it in every caption, always the same way. That token becomes the handle you pull in prompts later.
Caption what changes, not what stays
The rule of thumb: describe everything you want to be able to change, and leave out everything you want baked into the character.
If you caption "short blonde hair" in every image, hair becomes a variable the model can be told to change. If you never mention hair, the model treats it as part of the character. Decide which behaviour you want for each feature before you start, then be consistent.
So caption the pose, the framing, the clothing, the background, the lighting mood. Leave out the face structure and any feature you want locked in.
Keep captions consistent
Use one style across the whole set, either short tag lists or short natural sentences. Mixing the two makes the training signal noisier. Automatic captioning tools are a fine first pass, but read every caption and fix the ones that are wrong. Wrong captions are worse than short ones.
If you are weighing whether to do this yourself or use a managed setup, the trade-off is time against control, and the same question comes up across the platform features you will eventually use to publish and sell the output.
Run the training and read the results
Start with defaults. Change one thing at a time.
The settings that actually move the needle
- Network rank (dim): controls how much capacity the LoRA has. Higher rank captures more detail and also more of the background and artefacts you did not want. For a single character, a low to middling rank is usually enough.
- Learning rate: how fast the model adapts. Too high and the character collapses into a distorted, over-baked look. Too low and you get a faint resemblance.
- Steps and epochs: total exposure to the dataset. More is not better past a point.
- Repeats per image: if your set is small, repeats raise effective steps. Watch that you are not simply memorising the images.
Save intermediate checkpoints
Save the LoRA at several points during training rather than only at the end. Then test each one. It is common for a mid-training checkpoint to be better than the final one, because the final one has drifted into overfitting. Without saved checkpoints you would have to retrain to find that out.
Budget your time and compute
Training runs take time, and repeated runs take more. Decide up front how many attempts you are willing to make before you either accept a result or go back to the dataset. If you are running this on a hosted plan, the same logic applies to credits, and it is worth checking what each plan includes before you start a long series of experiments.
Test before you build anything on it
A LoRA that looks great in one prompt and falls apart in five others is not finished.
A basic test set
Run the same prompts against each checkpoint:
- Plain portrait, neutral background
- Full body, standing, outdoors
- Profile view
- Different lighting, for example a dim indoor scene
- Different clothing from anything in the dataset
- The character doing something, not just posing
Generate several images per prompt with different seeds. You are looking for the same person each time, with the face holding up at a distance and in profile, not just in close-up.
Reading the failures
Overfitting looks like: the same pose keeps appearing, backgrounds from your dataset leak in, clothing you never prompted shows up, or skin and edges look crunchy and over-processed. Fix it by training fewer steps, lowering the learning rate, using a lower weight at generation time, or adding more variety to the dataset.
Underfitting looks like: the face is vaguely right but changes between seeds, or the trigger word barely does anything. Fix it with more steps, a slightly higher rank, or better captions that stop describing the face.
Anatomy problems are often the base model, not your LoRA. Test the same prompt without the LoRA loaded before you blame your training.
Once the character holds up across a test set, you have something you can plan around, which is where the thinking in AI influencer monetization becomes useful.
Turn a trained character into work people pay for
A consistent character is an asset, not a business. The business is the offer.
Decide what you sell
The usual options are the same ones any creator has: a membership with tiers for ongoing sets, pay-per-view posts for individual drops, bundles for a themed batch, and paid requests as a service. Say you want $1,000 a month and you price a membership at $20: that is 50 members, which is a much clearer target than "grow the audience". Work backwards from the number to the offer, not forwards from the content.
If you need help pricing each of those formats, how to monetize content as a creator covers the trade-offs between one-off sales and recurring income.
Keep the character and the audience portable
Store your dataset, your captions, your trigger word and your trained files somewhere you control. Retraining from scratch because you lost the source images is a painful way to learn this. The same applies to the audience you build around the character: check that you can export your followers and your media from wherever you publish. The comparison points to look at when picking that home are laid out in how to choose an OnlyFans alternative, and they apply just as well to an AI character as to a human creator.
Plan the second character before you need it
Once you have trained one LoRA, the second is much faster, because the process is repeatable and you already know what a good dataset looks like. Keep notes on what worked: image count, rank, learning rate, which checkpoint won. That file is worth more than the LoRA itself.
