When a generated image does not look like your act, the instinct is to rewrite the prompt. That is usually the wrong lever. What constrains a face is the reference photos on file for that person, so start there.
Check every member has references
Open Members and go through each person. Each member profile has its own place to upload likeness photos, separate from Professional Photos.
The common pattern is one person covered well and everyone else not at all. That produces exactly what you would expect: one face that looks right and the rest drifting. Three or four solid references each beats ten for the singer and none for anyone else.
Check what kind of photos they are
Likeness photos and Professional Photos do different jobs, and putting the wrong thing in the likeness slot is the most common cause of a poor result.
Press shots are the usual culprit. Heavy retouching, dramatic lighting and stylized color all make a face harder to learn. What helps instead:
- Several angles rather than the same one repeated.
- Even, neutral lighting, without hard shadows across the face.
- The face unobstructed. No sunglasses, no hands, no microphone in front of the mouth.
- Recent photos, so the likeness matches how the person looks now.
- One person per photo. A reference with two faces does not clearly say which one belongs to this member.
Then look at the prompt
Once references are in place, prompt wording is worth adjusting, but it is the second lever, not the first. A prompt cannot describe a face into existence that the reference set does not support.
If a particular image still misses after you have added references, add more references rather than adding more adjectives.
If the whole image is wrong, not just the faces
A picture that misses the mood rather than the people is usually a profile problem instead. Biography, Genre and Location on Artist Profile are what the agents read for context, so a thin profile produces generic imagery.
Still stuck
Open Help & Support from your account menu. Support can see your workspace, so it can check whether the references actually uploaded rather than guessing from a description.
