The current pipeline appears to place significant weight on aesthetic priors relative to identity embeddings. This can lead to identity drift even under minor pose changes, requiring users to perform multiple generations to maintain character consistency. In practice, this increases credit consumption for tasks that should otherwise be straightforward.
#Identity Drift Under Pose Variation
olive pilot
olive pilot
Suggested Improvement:
Stronger identity embedding persistence or adjustable aesthetic bias could help keep the same character consistent during small pose changes, reducing repeated generations and unnecessary credit usage.
olive pilot
Additional Observation:
Based on my personal experience using Nano Banana Pro, it appears that the model now places even stronger weight on aesthetic priors compared to earlier versions. While this enhances overall visual appeal, it can further override original facial features and reduce character consistency, especially during minor pose or angle adjustments.
olive pilot
It feels like Gemini is spreading resources across videos, music, and immersive content, but none are fully polished. Focusing priorities or defining clear KPIs might help improve quality.
valid kernel
Yes
supple seal
@olive pilot hey, thanks for the feedback/suggestion, it's really helpful!