How to keep outfits consistent across AI video clips
Outfits drift in AI video for the same reason faces do: the model fills in what the references do not show. To keep clothing consistent, include the outfit in every reference image, add one full-body shot and one close-up of any detail that matters, keep clothing out of the prompt, and treat each outfit as its own version of the character. Garment choice helps too; some clothes are much easier for current models to hold than others.
Why clothing drifts
Most reference sets are built around the face, so the clothing appears only from the shoulders up. When the shot widens, the model has to invent the trousers, the shoes and the back of the jacket, and it invents them differently each time.
Collars, buttons, zips and pocket positions are small details that get re-drawn. Prints and patterns are the worst case, because the model reproduces the impression of a pattern, not the pattern.
Clothes also change with movement. A jacket that is closed in the reference may open when the character turns, and the lining the model invents may not match anything real.
Step-by-step: building outfit references
1. Put the full outfit on in every reference image. Even the face close-ups should show the collar and neckline of the series outfit. A close-up in a different top teaches the model a second outfit.
2. Add a full-body front view and a full-body back view. These two images remove most of the guessing in wide shots. Stand against a plain wall, arms slightly away from the body so the silhouette is clear.
3. Add one detail close-up of whatever must stay exact: a badge, an embroidered name, a distinctive buckle. Fill the frame with it and keep it sharp.
4. If the garment exists physically, a flat-lay photo shows the cut, the seams and the full pattern without folds. Lay it on a plain surface and photograph it straight down in even light. Use it in place of the detail close-up if the set is already full.
5. Stay within the limit. RefVid takes 1 to 5 photos per subject. A balanced five for a character with a fixed outfit is a front face view, a three-quarter view, a full-body front, a full-body back and one detail close-up.
Choosing clothes that hold up
If you are designing the character, you can make consistency easier. Solid colors hold better than prints. Large, simple shapes hold better than small, busy ones. One strong signature item, such as a yellow raincoat or a red scarf, makes the character recognizable even when smaller details vary.
Avoid fine stripes, small checks and text on clothing. Stripes can shimmer or bend in motion, and text on a T-shirt is re-drawn like any other small print. If the brand requires a logo on a shirt, make it large and simple, and plan close-ups where it faces the camera.
Layers multiply the risk. A shirt, an open jacket and a scarf are three things that can each change. For a long series, fewer layers mean fewer regenerations.
Prompting and versioning
6. Leave clothing out of the scene prompt. Writing "wearing her blue coat" adds a second description that may not match the coat in the images. If you need the coat buttoned or open, say only that: "coat open".
7. Treat each outfit as a version. A character in a winter coat and the same character in a summer dress are two reference sets with the same face. Name them clearly, such as Nina coat and Nina summer, and never use both in one generation. Mixing them gives you an outfit somewhere between the two.
8. Check wide shots first. When you review a batch, look at the widest shots before the close-ups; that is where clothing drift shows. If trousers or shoes keep changing, your full-body references are missing or unclear.
Common mistakes and checklist
The common mistakes are face references in different tops, no full-body back view, a busy print on the main garment, clothing described in the prompt, and two outfits mixed in one reference set.
Checklist: same outfit in every image; full-body front and back; one sharp detail close-up or a flat-lay; solid colors or large simple shapes; no clothing words in the prompt; one reference set per outfit.
In RefVid, save each outfit as its own subject, and each video uses only the photos of the subjects you add to it, so outfits are not mixed by accident. A video can hold up to 3 subjects, so do not add two outfit versions of the same character to one video.
Frequently asked questions
- Why does my character's shirt change color between AI video clips?
- Usually because the reference images show the shirt under different light, or because the prompt names a color that differs slightly from the images. Use evenly lit reference photos and leave clothing out of the prompt.
- Can AI video keep a patterned outfit consistent?
- Large, simple patterns hold reasonably well in AI video; small or detailed prints are usually re-drawn. If the pattern must be exact, use solid colors in video and show the print in still images.
- How do I show the same AI character in different outfits?
- Create a separate reference set for each outfit, reusing the same face. Generate each outfit from its own set and never mix two outfit sets in one request.