Guide
Using your own photo as a reference
A reference carries things a sentence cannot: the exact distance between someone's eyes, the particular line of a jaw. It is also the step people get least out of, usually for two reasons that are easy to fix.
What it carries
Identity, not style and not pose
A face is a set of proportions — the spacing of the features, the shape of the jaw, where the hairline sits. Those are measurements, and no description reaches them. You can write a paragraph about someone and still get a different person every time, which is the whole problem references exist to solve.
What a reference does not reliably carry is everything else. Supplying a photograph of someone smiling on a beach will not make every later image sunny or cheerful; those are properties of that photograph, not of the person, and the conditioning is aimed at the face. Expect identity to transfer and expect to direct the rest with words.
How many
More angles, not more photographs
The instinct is to supply as many pictures as possible. It is the wrong axis. Twenty photographs taken from the front describe one view of a head twenty times; the model still has no idea what the profile looks like, and will invent one the moment the subject turns.
A head is a three-dimensional object and what helps is coverage of it: front, both three-quarters, both profiles, slightly above, slightly below. Eight well-separated angles describe a head properly, which is why this platform asks for that set rather than for a folder.
Consistency across the set matters as much as the count. If the hair is different in each one, hair becomes a thing the model treats as variable. If one photograph is heavily edited, the editing gets read as a feature of the face.
What makes a good one
Boring photographs make the best references
Even, soft light. Hard shadow hides half the geometry, and whatever is hidden gets invented. Overcast daylight or a large window is close to ideal.
A neutral expression. A broad smile deforms most of the lower face. The model may take that deformation as the face itself and carry it into images where the person is meant to be doing something else.
Nothing in the way. Sunglasses, a hand on the chin, hair across one eye, a heavy filter. Each removes information and the gap gets filled with an average.
Enough resolution. The face should occupy a decent part of the frame. A face eighty pixels across in a group photograph carries very little of the geometry that makes it that person.
The common failure
It handed your own picture back
An edit returns something almost identical to the source. This is the most-reported problem with reference work and it is almost always the instruction rather than the picture.
An edit model weighs what you asked against what it was given. A short or vague instruction — make it better
, more cinematic
— does not describe a change, so the output that best satisfies it is the input. Nothing went wrong; too little was asked.
The fix is to name what should differ and what should stay: same person, same face, change the background to a night street, relight with cool light from the left, keep the jacket
. Long enough to be a change, specific enough to be followed.
One other cause worth ruling out: on an adult-capable account, a request routed to a model with its own filter switched on can return a sanitised version of the input instead of an error. If the result looks like a tamer copy of the reference rather than an unchanged one, that is a different failure with a different cause.
Before you upload
Whose face it is matters more than any of this
Everything above assumes the photograph is yours to use. Likeness is governed by the consent of the person in the picture, not by who owns the file or who took it, and that applies with considerably more force to intimate depiction. The guide on generating a real person covers what is required and what is never permitted.
Asked often
Short answers
- How many reference images do I need?
- For a face, more angles beat more photos. Eight views of one person taken from different directions describe a head in a way twenty near-identical front-facing shots cannot, because what the model needs is the shape from several sides rather than repeated confirmation of one side.
- Why did the model return my reference image almost unchanged?
- Because the instruction did not ask for enough. An edit model weighs your prompt against the picture it was given, and a short or vague instruction loses - the safest output that satisfies it is the original. Saying specifically what should differ, and what should stay, is what moves it.
- What makes a good reference photo?
- Even lighting, the face unobstructed, a neutral expression and reasonable resolution. Hard shadows, sunglasses, heavy makeup and extreme expressions all get read as features of the person rather than of the photograph, and then they follow into every image you generate afterwards.
Keep reading
The rest of the series
- Why a prompt gets refusedThree different refusals wearing the same error. Which one you have, and which can be reworded.
- Does the language of your prompt matter?The script matters more than the language, and the gap is larger than anyone tells you.
- Why AI changes the face every timeThe mechanism behind the problem: no memory, and a prompt that names a category rather than a person.
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.