Guide
How to get the same character in every image
Five approaches, what each one costs, and where each one breaks. One of them is still the strongest option available and it is not the one we sell.
The decision
What you are really choosing between
Every method that holds a character does the same thing: it moves the identity out of the prompt and into something more durable — an image the model conditions on, a token trained to mean this face, or weights adjusted to know it. The previous guide covers why words alone cannot do it.
What separates them is setup cost, how much they hold, and how they fail. Those three are the whole decision.
Side by side
The five approaches
More detail in the prompt
- Setup cost
- Free
- What it holds
- Almost nothing
Narrows the range for the first few details, then flattens. Past roughly a dozen clauses, later ones are diluted. Describe hard enough and every output starts to look identical, which is a corner of the model rather than your character and does not survive a change of pose.
A fixed seed
- Setup cost
- Free
- What it holds
- The same image, not the same person
Identical seed, prompt, model and settings reproduce an image exactly. Change the prompt to move the character and the denoising path changes from the first step. Useful for reproducing a result, not for putting someone in a new scene.
A trained embedding
- Setup cost
- Minutes, a handful of images
- What it holds
- A likeness, loosely
Teaches the model a new token that means this face. Cheap and portable, and it degrades politely rather than failing outright. It carries less than full fine-tuning and tends to drift under strong styling or an unusual angle.
A LoRA, or full fine-tuning
- Setup cost
- Hours, a curated dataset, per character
- What it holds
- The most, when it is fed well
Adjusts the weights themselves on 15-30 good photographs. Done properly this is still the strongest result available, and the rest of this page does not pretend otherwise. The costs are real: you need the dataset before you start, training time per character, a file to manage for each one, and a retrain when you want a change. It can also overfit, which shows up as your character wearing the same expression and the same lighting in every image.
Reference conditioning
- Setup cost
- None beyond the generation
- What it holds
- The face, as far as the reference shows it
The generation is conditioned on an actual photograph rather than on words, so there is no training step and no file per character. The limit is what the reference contains: condition on one front-facing image and the first profile shot is invented, because the model has never been shown that side of the head.
Be told plainly
When to train a LoRA instead
Train one when you already have a real dataset of the same person — twenty or more varied photographs, different lighting, different angles — and you intend to use that character for a long time. The setup cost is paid once and the result is the most faithful thing available.
Train one, too, when the likeness has to survive heavy stylisation. Pushing a reference-conditioned generation into a strong art style pulls it away from the reference, because the style and the identity are competing for the same pixels. Weights that know the face hold up better under that pressure.
Do not train one to find out whether a character works. That is a dataset and an hour of compute spent before you have seen the character once, and most characters do not survive first contact.
The angle problem
Why one reference image is not enough
Reference conditioning is the cheapest method that genuinely works, and it has one failure that catches everybody. Condition on a front-facing photograph and ask for a three-quarter view, and the model has to invent the parts it has never been shown. Cheekbone, jaw from the side, the back of the head. It invents them plausibly, and differently each time.
The fix is not a better algorithm, it is more reference. Give the system the face from several angles, and condition each generation on whichever one matches the shot being asked for — a profile request conditions on the profile reference rather than guessing from the front.
That is what Unlimited Studio does: eight reference angles per character, picked from automatically by what the prompt implies. It is reference conditioning with the angle gap closed, which keeps the no-training-step property and removes the failure that usually ends the experiment.
Keep reading
The rest of the series
- Keeping a character consistent in AI videoWhy video drifts twice over, and the two techniques that fight it from opposite directions.
- Why the outfit and the room change tooWardrobe and location drift for a different reason than faces do, and need a different fix.
- Why AI video clips are so shortThree separate reasons for the ceiling, and why longer films are assembled rather than generated.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.