This adapter for Stable Diffusion 1.5 and SDXL is designed to inject the general composition of an image into the model while mostly ignoring the style and content. Meaning a portrait of a person waving their left hand will result in an image of a completely different person waving with their left hand.

I do a lot of experiments and other things. To keep up to date, follow me on Twitter.

I want to give a special thanks to POM with BANODOCO. This was their idea, I just trained it. Full credit goes to them.

Use just like other IP+ adapters from h94/IP-Adapter. For both SD1.5 and SDXL variants, use the CLIP vision encoder (CLIP-H)

You may need to lower the CFG to around 3 for best results, especially on the SDXL variant.

Control nets are more rigid. A control net will spatially align an image to nearly perfectly match the control image. The composition adapter allows the control to be more flexible.