Guide
Why the outfit and the room change too
You fixed the face and the jacket still changed colour. Clothing and location drift for a different reason than faces do, and the fix that holds a face does not hold them.
The asymmetry
A face is one thing; an outfit is a list
A face is a single object with fixed proportions. Whatever method you use to hold one — a reference image, a trained embedding, fine-tuned weights — it is holding one shape that does not legitimately change between pictures.
An outfit is not one thing. It is a jacket, a collar, a specific blue, a zip, the way a sleeve is pushed up, a pocket. Each of those is separately describable, separately generated, and separately able to come out differently. You are not asking the model to reproduce one object. You are asking it to reproduce a dozen, all at once, every time.
That is why wardrobe drift feels worse than face drift even when it is objectively smaller. The face is wrong or it is not. The jacket is eighty percent right in a different way each time, and the errors do not repeat, so there is nothing to correct.
Why it is worse than it looks
A room has a geometry that must agree with itself
Backgrounds add a problem clothing does not have. A location has to be internally consistent across shots taken from different places in it. If a window is on the left in a wide shot, it has to be behind the character in the reverse angle, and the model has no model of the room to consult — it has a description and whatever the reference image happened to show.
So you get rooms that are individually plausible and collectively impossible. The sofa changes sides. A doorway appears on a wall that was blank. The light comes from a window that is not in the frame and was not in the frame last time either. None of these is noticeable in one image, and all of them are obvious the moment two images are meant to be the same place.
It is the same failure as the face, one level up: the model is generating something that satisfies a description rather than recalling a thing that exists.
What does not work
Describing the outfit harder
The instinct is the same one that fails for faces, and it fails the same way. “A faded olive canvas jacket with brass buttons and a corduroy collar” is a better prompt than “a jacket”, and it still describes a category. The buttons will be brass and there will be a different number of them.
Worse, long wardrobe descriptions compete with everything else in the prompt. Spend thirty words on a jacket and the model has that much less attention for the pose, the framing and the face you were trying to hold in the first place. People often solve wardrobe drift and reintroduce face drift in the same edit, then conclude the tool is random.
There is one case where description is enough: when the garment is genuinely generic. “A plain white t-shirt” has few degrees of freedom, so it comes back close enough every time. The more distinctive the item, the less description does for you — which is the opposite of the intuition.
What does work
Treat the environment as a thing you lock, not describe
The same move that fixes faces fixes this: stop carrying the identity in words and carry it in something the generation conditions on. A reference image of the location, held and reused, does for the room what a character reference does for the face.
Two practical consequences follow. The first is that a location is worth establishing once, deliberately, as its own asset — generated or photographed, then kept — rather than re-described per shot. The second is that wardrobe is usually better treated as part of the character than as part of the prompt: if the reference images show the outfit, the outfit comes along with the face, free.
That has a cost worth knowing about. A character whose references all show one outfit becomes hard to put in a different one, because you are now fighting the reference rather than using it. If a character needs to change clothes across a story, that is a second set of references, not a prompt edit.
Unlimited Studio keeps environments as first-class assets for this reason — locked the same way characters are, and reusable across shots — rather than leaving a location to be re-described every time and re-invented every time.
A shortcut worth knowing
Crop tighter than you think you need to
Most background inconsistency is invisible if the background is not in the shot. A tight frame removes the furniture that would have moved, and a shallow depth of field blurs what is left into something no viewer can check.
This is not a trick, it is what the craft does anyway. Close coverage is how scenes are shot when the location is a set with three walls, and the reason is the same: the parts that are not there cannot contradict anything. Save the wide shot for the one frame where the location matters, get it right once, and cut to faces for the rest.
Keep reading
The rest of the series
- Why AI video clips are so shortThree separate reasons for the ceiling, and why longer films are assembled rather than generated.
- Putting two characters in one sceneIdentity leakage: why one face borrows the other, and why prompt wording is not the fix.
- Why the model ignores half your promptAttention is finite and spread unevenly. Why the end of a long prompt is the part that gets dropped.
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.