The first instinct after a failed face replacement is to add more “do not” instructions. Do not distort the face. Do not move the camera. Do not change the hair. Do not change the background. Eventually the prompt becomes longer than the original scene description.
Negative constraints work best when they clarify the boundary of the edit rather than attempt to describe every possible failure.
Start with the main rule
“Replace only the intended character” is the core objective. Add the source-character description immediately after it. If there are two identity references, map each to one named role.
Then mention the two or three things that actually went wrong in your last render. For example: “Do not replace the person in the red jacket. Keep both uploaded identities separate. Preserve original speaking timing.”
That is far more useful than twenty generic quality warnings.
Distinguish a non-target from an unwanted artifact
A non-target is someone who should not be changed. An artifact is a defect such as distorted ears or flickering glasses. Give them separate instructions. This makes the prompt easier to understand and easier to debug.
If a model supports a dedicated negative prompt field, follow that field's documented behavior. Do not assume every API accepts negative prompts or interprets them the same way.
Be careful with impossible demands
Asking for a perfect identity change while preserving every pixel of the original face is contradictory. Ask to preserve the motion and scene, not the original person's identity. Likewise, “keep all hair unchanged” may conflict with a replacement reference whose hairstyle is completely different.
Choose what belongs to the source performance and what belongs to the uploaded identity.
Test one correction at a time
After a failed render, write down the problem and make the smallest prompt change that addresses it. If the red-jacket bystander keeps changing, tighten that exclusion before adjusting color, sharpness or camera motion.
Clear boundaries beat a wall of prohibitions. They also make successful prompts much easier to reuse.
