Clothing tells viewers where they are in a scene. A blazer, a uniform or a bright jacket might be part of the joke or the story. So it can be disappointing when a face replacement also changes the shirt, collar or sleeves.
This happens most often when the model treats the prompt as a complete scene recreation instead of a localized identity edit.
Tell the model which visual layer owns the outfit
The source video should usually determine the actor's body movement and clothing in a face-replacement task. The uploaded reference image should determine recognizable facial identity. Saying this clearly avoids the mixed signals of “make the person look exactly like the photo” when the photo contains a different outfit.
Try: “Transfer the uploaded person's facial identity to the designated source character. Keep the source actor's clothing, body, pose, and shot timing unchanged.”
Why collars and necklines are tricky
When the model widens or changes the head, the neck connection may also shift. Collars sit right next to that boundary, so they can become distorted or change color as the face moves.
Review the collar during a head turn and a speaking moment. If the clothing stays consistent but the jaw looks pasted on, you are dealing with a blending problem rather than wardrobe drift.
Choose the right scope of edit
A full scene-recreation model may have difficulty preserving every wardrobe detail. A purpose-built video-to-video editing model is usually a better candidate when you need close adherence to existing footage.
If the model still redesigns outfits, simplify your prompt, avoid creative wardrobe adjectives, and test a shorter segment before running a full export.
Clothing preservation is about giving the source clip authority over the parts of the image you do not intend to change.
