Closed ziyaxuanyi closed 2 months ago
you can send the image (or actually any image) as negative conditioning. This can be very effective and the effect can be calibrated also adding noise
I also talk about this in this video https://www.youtube.com/watch?v=gmwZGC8UVHE
Note that one of the methods mentioned in the paper: subtracted clip's embeddings as distinguished representation, but it does not seem to be implemented in the relevant implementation of this code (the workflow used is: ComfyUI_IPAdapter_plusexamples/ipadapter_style_composition.json)
Has this code implemented this? If so, could you provide a sample workflow?