Consistent color shift/oversaturation in image-to-image with NanoBanana (NB Pro & NB2)

Gemini API
1

Hi everyone,

I’m experiencing a consistent color shift issue when using NanoBanana models (NB Pro and NB2) for virtual try-on (image-to-image), and wanted to check if others have seen this or if there’s a known workaround.

Setup

  • Task: Image-to-image (virtual try-on)

  • Input: Base model image + garment image

  • Models:

    • NanoBanana Pro

    • NanoBanana NB2

  • Prompt (strict constraints):

    • Only change the clothing

    • Preserve face, skin tone, lighting, and background exactly

Expected Behavior

  • Only the garment should change

  • All other elements (skin tone, lighting, background) should remain identical to the base image

Actual Behavior

  • Overall saturation increases

  • Skin tone becomes reddish / blotchy

  • Slight background color shift

  • Looks like unintended global color grading or relighting

This happens consistently across multiple inputs and both NB Pro and NB2.

Observations

  • The issue persists even with very strict prompts (e.g., explicitly telling the model not to change lighting, skin, or background)

  • It feels like some internal harmonization / relighting step is being applied globally

  • The effect is especially noticeable on skin tones

Question

  • Is this expected behavior for NanoBanana image-to-image?

  • Is there any way to:

    • Disable or reduce global color adjustments (e.g., relighting / harmonization)?

    • Better preserve original color fidelity, especially for skin tones?

Additional Info

  • The issue occurs regardless of prompt strictness

  • The issue occurs across most inputs, but becomes significantly worse on certain images

Any insights or suggestions would be greatly appreciated. Thanks in advance!