Issue with Face Clarity in Scene Generation with InstantID in ComfyUI

[wangyanmeego]

Hello,

I've been using the InstantID feature within the ComfyUI, as implemented in the repository by cubiq, to generate contextual scenes using a reference facial image of a person, combined with prompt words. The feature works impressively well when the generated person's face occupies a large portion of the image, seamlessly integrating into the scene.

However, I encounter issues when the face of the generated person occupies a smaller area of the image. In such cases, the facial features become unclear and can even appear distorted.

Attached are some reference images from my tests. Could you please advise if there are any methods or settings available within the repository that might help address and fix this issue with facial clarity?

The face reference image:

The full body result images:

This is my comfyui workflow: workflow (6).json

Thank you for your assistance.

[meatflavourdev]

Care to provide an example of the failure state to which you are referring so we can see the problem itself? (scene generation with small resolution reference face)

[wangyanmeego]

I have put some failed examples in Google Drive. When the face occupies a small proportion of the scene, the final face will be distorted and deformed.