Image Blending with Attention Masking - IPAdapter

 

How it works:

Attention masking basically focusses on parts of the Empty Latent Image. Those specific areas are the ones you’ll fill in with the desired image, while the entire jigsaw puzzle will be getting blended together with the background. This allows you to actually blend different images together into a coherent whole. (Even different types if you want, you can blend Realism and Cartoons for example if you so choose.)

The workflow consists of 3 separate parts:

1. Image Preparation

2. Image Creation

3. HiRes.Fix

 

The Workflow has been made super easy to expand upon as you can easily add more Image Loaders with minimal effort, this up to a total of 5.

Keep in mind: The more images you’ll load into the workflow, the harder it’ll become for the IPAdapter to create a coherent whole. 1-2 images + a Background is pretty standard and doable more than that will require a bit more finetuning in the settings.

 

What is needed:

• As always, I suggest you Download ComfyUI Manager by Lt.Dr.Data for easy instalment of any additional nodes you currently lack: https://github.com/ltdrdata/ComfyUI-Manager

• You'll have to Install and Download IPAdapter Files from the official Source: https://github.com/cubiq/ComfyUI_IPAdapter_plus

  • This because they aren't automatically downloaded into ComfyUI upon first use.

• The Loader for Checkpoints also has a single LoRA slot added for a Detailer LoRA of Choice these will be linked in the Suggested resources.

• Because there is an Upscale component to this workflow I suggest a couple of Upscale Models.

  • 4x Animesharp (For Any and all Anime Images)

  • 4x Ultrasharp (For Mixes that Blend Cartoon and Realism together)

  • 4x NMKD-Superscale-SP_178000_G (For Photographic Realism)

    • These Will be provided in a separate Download File for ease.

Where to Put the Files:

• All information for IPAdapter is explained on the official github webpage.

  • This Workflow doesn't use FaceID Nodes so if you don't want to ImagineFace isn't a requirement for this at all.

• The LoRAs go into: "ComfyUI\models\loras"

• The Upscale Models got into: "ComfyUI\models\upscale_models"

As always, more info is provided in the Note Nodes inside the Workflow. in this case it'll pertain to mostly differences between SD1.5 and SDXL as this Workflow is working on both variations of Stable Diffusion.

描述:

Workflow Updated to v1.1

• hiresfixScale Node malfunction due to node update. (Fixed)

训练词语: