Hello Google Gemini and Imagen Team,
I am writing to report a significant and reproducible limitation in Gemini’s image generation system related to identity preservation in multi-person image generation and upscaling.
Testing confirms that single-person image enhancement generally performs well. When only one subject is present, facial detail and overall identity are preserved at an acceptable level.
However, consistent failures occur when multiple people are involved.
First, when enhancing or upscaling group photos, individual identities are not reliably preserved. Facial structures subtly shift, causing people to no longer resemble the original subjects. This occurs even when prompts clearly instruct the model to preserve identity and avoid any form of stylization or beautification.
Second, when users attempt to create a group image by combining multiple individual photos into a single scene, identity degradation becomes severe. Each person’s facial geometry, proportions, and visual characteristics drift toward synthetic or averaged appearances, despite each subject being provided with a clear reference image.
This suggests that Gemini’s current image conditioning pipeline struggles to handle multiple identity constraints simultaneously, with identity signals interfering rather than remaining independently anchored.
These limitations directly affect real-world use cases such as reconstructing missed group selfies, family photo restoration, professional team portraits, and shared memory preservation.
The issue is not realism or resolution quality, but multi-identity coherence.
Addressing this problem through independent identity conditioning, stronger identity-lock mechanisms, and dedicated evaluation benchmarks for group-photo preservation would significantly improve the reliability and trustworthiness of Gemini’s image generation system.
Thank you for your continued efforts, and I hope this feedback contributes to future improvements.