Four ways to lock a character's identity, ranked by cost and drift
Without a face that stays the same across a year of posts there is no character, and without a character there is no subscription to sell. There are four methods that work, they have very different costs, and each fails in its own recognisable way.
What you get from this page
- The four methods, with the specific drift signature each one produces
- Why the cheapest method is the wrong choice for anything you will sell
- A quality gate that catches identity drift before publication
- When to combine methods instead of choosing between them
On this page
A subscriber does not consciously audit your character's face. They notice something is wrong, and the feeling is that the content is cheap. That reaction is what identity drift costs you, and it is why consistency is a commercial problem rather than an aesthetic one.
There are four methods in general use. They are not equivalent, and the differences show up in exactly the places that matter: close-ups, unusual angles, and any image a subscriber has paid specifically to receive.
The four methods
| Method | Setup cost | Per-image cost | Drift | Best for |
|---|---|---|---|---|
| Prompt and seed locking | Minutes | Lowest | High | Exploration and early concepting only |
| Reference image conditioning | Under an hour | Low | Moderate | Small sets, quick turnarounds, custom requests |
| Face-swap post-processing | Under an hour | Moderate | Low on faces, none elsewhere | Fixing an otherwise good image |
| Trained character LoRA | Hours, plus a training run | Lowest at volume | Lowest | Anything you intend to sell |
Relative rankings, not measurements. Cost and drift are rated against each other based on how each method works; there are deliberately no percentages here because we have not published a keeper-rate series per method.
1. Prompt and seed locking
Write a detailed appearance description, fix the seed, and reuse both. It costs nothing and it is where everyone starts.
- What it does well
- Establishes a rough visual direction quickly. Useful for deciding what a character should look like before you commit to training anything.
- Drift signature
- Face structure shifts whenever anything else in the prompt changes: pose, lighting, clothing, camera distance. The character is recognisably a type rather than a person, and side-by-side comparison of two images makes it obvious.
- Why it fails commercially
- A subscriber sees your images sequentially in a feed, which is the worst possible viewing condition for this method. Continuity errors accumulate into the impression that the account is low effort.
2. Reference image conditioning
Feed a reference image of the character alongside the prompt so the model conditions on that appearance. IP-adapter style approaches and their descendants all sit here.
- What it does well
- A large improvement over prompting for near-zero setup. Excellent for one-off requests where training is not worth it, and for testing whether a character concept holds up before committing to a training run.
- Drift signature
- Identity holds at medium distance and degrades in close-up. Fine facial structure, the specific geometry that makes a face that face, averages toward the base model. Expect the character to look subtly different in a portrait crop than in a full-body shot.
- The practical constraint
- The output is only as consistent as the reference, so reference selection becomes its own quality problem. A reference with an unusual expression or lighting condition propagates that into everything.
3. Face-swap post-processing
Generate the image, then replace the face with a consistent one as a separate step.
- What it does well
- Rescues images that are compositionally right but wrong on identity. As a repair tool it earns its place in the pipeline.
- Drift signature
- The face is consistent and everything else is not. Body proportions, skin tone continuity at the jaw, hands and hair all wander freely, and the boundary between swapped and original regions is the first thing a careful viewer notices.
- The bigger objection
- It fixes a symptom. If your underlying pipeline cannot hold identity, swapping faces onto its output gives you consistent faces on inconsistent bodies, which reads as uncanny rather than as fixed.
4. Trained character LoRA
Train a small adapter on a curated set of images of your character, then load it with a trigger word for every generation. This is the method that solves the problem.
- What it does well
- Holds identity across poses, expressions, lighting, camera distance and clothing, because the model has learned the character rather than being nudged toward it. It is also the only method whose consistency improves with effort in a predictable way.
- Drift signature
- Low, and where it appears it is usually diagnosable: overfitting to the training set's poses and backgrounds, or an under-trained adapter that loses the character at high guidance. Both are fixable by retraining, which is not true of the other three methods' failure modes.
- The real cost
- Not the compute; a character LoRA is a cheap training run on rented hardware. The cost is the dataset: assembling a clean, varied, correctly captioned set is a few hours of careful work, and the quality of that work determines the quality of the result more than any hyperparameter does.
Combining methods
The methods are not mutually exclusive, and a mature pipeline usually uses three of them for different jobs.
- LoRA as the base, always loaded, carrying the identity.
- Reference conditioning on top when you need a specific outfit or setting the LoRA has not seen.
- Face-swap as repair only, on individual images that are otherwise good. Not as a production step.
- Prompt and seed locking retired entirely once the LoRA exists. It has no remaining role.
A quality gate that catches drift
Whatever method you use, the review step is what protects the library. This is the gate worth running before anything publishes.
Identity review, per image
Compare against a fixed reference sheet, not against the last image you made
Drift is gradual. Comparing sequentially lets it accumulate invisibly; comparing against a fixed anchor does not.
Check the portrait crop specifically
Crop to head and shoulders and look again. Most drift is invisible at full-body scale and obvious in a portrait.
Check continuity of distinguishing features
Moles, freckles, asymmetries, eye colour, hairline. These are what make a face specific, and they are the first things to wander.
Check hands, jewellery and text
Still the reliable artefact locations. A single bad hand in a paid set generates more complaints than five average images.
Confirm the presentation is unambiguously adult
Non-negotiable, and a rejection, not a fix.
Log the reject reason
Reject reasons aggregated over a week tell you exactly what to fix in the pipeline. Without the log you are guessing.
How many images do I need to train a character LoRA?
Community guidance for FLUX character LoRAs generally lands between 15 and 50 images. We have not tested that range systematically, so treat it as the prevailing convention and not as a finding. Variety matters far more than volume. Twenty varied, clean, well-captioned images beat sixty near-duplicates, which will teach the adapter the background and the pose rather than the character.
Can I train a LoRA on images of a real person?
No. Do not do this. It breaches platform policy across the board and, where the person has not consented, carries legal exposure that has nothing to do with whether a platform notices. Generate a synthetic identity and train on that.
My LoRA works at low guidance and loses the face at high guidance. What is wrong?
Usually under-training or too low an adapter weight for the base model you are pairing it with. Both are cheap to test: retrain with more steps, and sweep the LoRA weight across a fixed prompt set before changing anything else.
Is character consistency still a problem with newer base models?
It is a smaller problem than it was, and it has not gone away. Better base models raise the floor of what prompting alone achieves, which mostly means the drift becomes subtler rather than absent. A trained adapter still measurably outperforms prompting on the same base model.
Where these numbers come from
The comparison table in this guide rates methods on cost, effort and drift relative to each other. Those are qualitative rankings drawn from how each method works, not measurements: we have not published a keeper-rate series per method and the table deliberately contains no percentages. The dataset-size range quoted for LoRA training is the commonly cited community figure, labelled as such where it appears.
Sources
- 01
Runpod GPU pricingRates behind the cost ranking in the comparison table
- 02
Fanvue AI content guidelinesPlatform position on likeness and synthetic content, referenced in the face-swap section
Who wrote this
AIOF Editorial DeskResearch & editorial
AIOF is an editorial desk that works from primary sources: platform terms, vendor pricing pages and published company data, read and dated rather than repeated from other write-ups. Where a figure comes from someone else we cite it and grade it. Where we have not measured something ourselves we say so, which today is most of the places you might expect a first-hand number.
- Works from primary platform documents, with the date each page was last checked published on the page itself
- Grades every figure as sourced, third-party estimate, modelled or measured, and labels which
- Publishes no first-party measured benchmarks yet, and says so on the methodology page rather than implying otherwise
- Corrections are made in place with a dated changelog entry, listed on the updates page
Rules and pricing here change often. This page was last touched on . Found something out of date? Tell us and we will fix it in place with a dated note.
Read next
Training a character LoRA
A practical walkthrough of training a character LoRA for a consistent AI persona, covering dataset construction, captioning and the settings that matter.
ProductionTutorial7 minProductionTutorial7 minCost per usable image
A model for the true cost of AI content production, based on keeper rate rather than raw generation cost, with published GPU and API rates.
EconomicsReference5 minEconomicsReference5 minPre-launch checklist
What to have in place before you send the first visitor to an AI creator account, grouped by what it protects you from.
ProductionChecklist3 minProductionChecklist3 min