Different results with /ai/annotate-image (controlnet mask generation)
Opened this issue · 1 comments
Aedial commented
There is a slight difference in the /ai/annotate-image compared to the frontend, due to some unknown conversion frontend-side.
Base image: image
Base image from frontend: image_input_frontent
Putting the image_input_frontent through the endpoint yields the same result as the frontend.
Aedial commented
From what I found, the difference is only -1/+1 in one of the RGB channels for ~half of the pixels of the top 25% of the image (so ~1/8 of the image). This is done by the frontend before sending the image to the backend to get the mask, but the reasons for such a subtle change are unknown.