Hello developers,
I am working on a cinematic video-to-video project using Google Flow, and I am experiencing a persistent character consistency failure. Despite running 20 to 30 iterations across multiple clean and newly created projects, the model consistently fails to retain the character’s identity and facial features.
Here are the two workflows I tested extensively:
-
Character Consistency Feature: I created a dedicated character profile (named “Kadir”). I uploaded a high-resolution, neutral portrait for the face model and completely filled out the “Describe body and outfit” section. I used the Nano Banana Pro model for maximum adherence.
-
Image-to-Video Reference Method: I disabled the character agent and directly attached high-quality full-body and profile reference images to the video-to-video prompt, explicitly referencing “the person in the reference images.”
The Issue: The source video is a static clean plate (a coastal walkway with a bench). In every single render, instead of drawing the character from the references, the model hallucinates a completely different, generic default avatar. It sometimes catches the low-frequency data (like clothing color), but the high-frequency facial geometry is completely replaced by a random person.
Since this workflow used to yield stable results in previous builds, is this a known regression or an IP-Adapter weight issue in the recent updates? Am I missing a hidden prompt weight parameter, or is the model currently suffering from severe identity drift on blank canvases?
Any guidance or technical workaround from the engineering team would be highly appreciated.