Redux is an excellent tool for maintaining consistency after Flux releases the family bucket.
No more worrying about overly differentiated scenes and characters when working on multi-screen creations.
There are enthusiasts who combinePuLID(Plugin developed by Byte) Face Swap feature, using Redux to achieve near-perfect Tutu-born one-click face swap.
This paper describes the workflow, with nodes and models for download.
PuLID uses version 0.91 of the model, due to the conflict between the PuLID plugin and the latest version (as of December 4, 2024) of ComfyUI, a modified version of the PuLID plugin is attached to this article, which needs to be manually installed instead.
I. Redux model
download:
https://pan.quark.cn/s/3af84d557f9f
Model Installation:
clip_vision model sigclip_vision_patch14_384.safetensors Install to
ComfyUI\models\clip_vision
Style Model flux1-redux-dev.safetensors Install to
ComfyUI\models\style_models
II. PuLID model and PuLID correction node installation package
download:
https://pan.quark.cn/s/3413c1eb6c1c
Node Installation:
ComfyUI upgrade, PuLID plugin has not been updated, the execution will report an error, you need to use this article's ComfyUI-PuLID-Flux-master.zip unpacked to replace (folder delete -master).
Unzip and replace the folder:
ComfyUI\custom_nodes\ComfyUI-PuLID-Flux
Model Installation:
pulid_flux_v0.9.1.safetensors Install to
ComfyUI\models\pulid
III. Introduction to workflow nodes
This article uses the gguf model, which is installed in the directory
ComfyUI\models\unet
Depending on the size of the video memory, select the guff model used locally.
For 11G or less video memory (dev and schnell versions):
https://pan.quark.cn/s/ba0ef44a35f5
11G+ video memory used (dev and schnell versions):
https://pan.quark.cn/s/8ded08171604
1, Clip loader and style model loader, in accordance with the illustration to select the model, if you can not choose to check the first step of this article, the model installation path.
2, GGUF Model Loader and Clip Loader, choose the appropriate quantized version according to the size of the video memory.
3. Load the picture
This workflow requires two images.
The original image is on the left, and the image that needs a face swap is on the right.
IV. Running the map
Original image
Generate a diagram:
Caveats:
After replacing the PuLID node, you need to reboot the initiator.
Multiple runs will inexplicably report an error, restart the launcher and you'll be fine, probably because this magically modified version of the PuLID plugin is not perfect.