Generating consistent character images within my custom GPT has become extremely difficult. Previously, it was possible for me to reliably create the same character by referencing my avatar, along with strict instructions and a detailed knowledge file for enforcement. This combination worked well for building unique brand character that would stay visually consistent across generations. I have been using it for months.
Now, that consistency is gone—even when supplying a reference image, the results regularly drift from the intended look (face, hair, outfit, and style). Getting the same character twice is almost impossible, no matter how clear the prompt is, if I supply the reference (Avatar) within the chat.
On top of that, more prompts are being flagged or blocked, even though they’re completely safe and unchanged from what worked before. This regression has made creative workflows inside my custom GPT frustrating and unreliable.
An option to lock character features more strictly to a provided reference (avatar + knowledge file + instructions) was working before and would be a huge improvement if restored. Right now, the loss of consistency and the flagging of prompts that are as innocent as they come is a major setback for me.
I have spent days trying to find work arounds with minimal success and even if I achieve something it is not repeatable.
If anyone has a magic trick up their sleeve I am all ears.