As a follow up to my last post regarding Consistent portraits using IP-Adapters for SDXL, this is a short comparison of the two face IP-Adapters for SDXL by h94 / xiaohu: namely, ip-adapter-plus-face_sdxl_vit-h.bin and ip-adapter-faceid_sdxl.bin.
Here is the workflow - as you can see I am using SG161222/RealVisXL_V3.0_Turbo to generate a portrait and then using that portrait in a subsequently generated image.
Here is the output at different IP-Adapter weights:
You can clearly see the progression of the original SDXL output towards an face influenced by IP-Adapter. Best to experiment further, but a few observations:
- Plus-Face weights seem to go from 0.0 to 1.0 but work better at a lower weight (the workflow is at
0.3) - too high and the image is not quite what the prompt describes anymore, but - FaceID weights seem to go from 0.0 to 2.0 quite comfortably (the workflow is at
1.3but perhaps under1is better) - Use a source image that matches your prompt i.e. a straight on portrait generates better straight on portraits, and a side view portrait generates a better side view portrait
- Higher weights will ruin the angle of the face as prompted and you get the angle of the source
- I do not really know... but I think I still prefer Plus-Face.
BTW for the curious: I used paint.net to create the animated .webp using the WEBP Animations and Images FileType Plugin (v1.4 2022-01-24) and Text+ Plugin (2014-04-27). I then converted the .webp to .webm.