Demo video @ Youtube.
First, read the IP Adapter Plus doc, as well as basic comfyui doc.
Then use comfyui manager, to install all the missing models and nodes, i.e. the Clip VIT H from ipadapter, the sdxl vit h ipadapter model, the big sdxl models, efficient nodes...
- IPAdapter weight: Makes the face / style copy from the top image stronger
- Unclip conditioning strength: The 2nd image is encoded into a CLIP prompt, but you can use additional text to modify the images, i.e. make them smile.
- The KSampler Adv. (Efficient) has:
- a "start at step" parameter, the later you start the closer the image is to the latent background image.
- an amount of steps depends on your model. SDXL turbo works fine with 1, the others require 10-20
- the cfg-scale, i.e. how "strong" your model should be. Too high and you get "deep fried" / "burnt" images with too high contrast. turbo works fine with 1, the others work best with 2-3.
- On top there are disabled XY plot nodes, so you can run multiple parameters in one batch
- For faster runs, reduce amount of steps and the resolution of the latent (therefore output) image. SDXL only performs well on very limited resolutions.
To run the whole grid, press the "queue full grid" button. The node emits outputs from 1-N, but the index of load image batch expects 0-N, so you have to insert a dummy 0th image.
Try to pre-crop your images to a square and 1024x1024 with the face close to center for best results.
For my 13700k / 2070 super, creating 1024^2 images takes roughly 7 seconds for sdxl turbo (1 step), 13 seconds for sdxl fae (10 steps) and 21 seconds for sdxl base (20 steps).
All of the following images are licensed CC0, do whatever you want with them with or without mentioning this repo.