What is Flux Redux and how does it work?
Flux Redux functions as an adapter designed to work in a similar fashion to IP Adapters, providing style, composition, and other defining qualities and characteristics from a reference image into the final generated image. IP Adapters offer excellent style replication from another image, and the speaker believes Flux Redux offers similar functionality.
It lets you effectively clone or transfer the essence of a reference image, injecting its artistic qualities into your new creation. This level of control is essential for generating images that match a pre-defined aesthetic, replicate a known style, or simply add a personal touch.
Key features of Flux Redux:
- Style Transfer: Copies the look of reference images
- Composition Replication: Transfers layouts of different pictures to a base image to be modified or Altered.
- Detail Replication: It can account for small details that matter to the user. Black Forest Labs offers some of its own examples, like changing a man's portrait into an anime version.
Flux Redux Workflow for ComfyUI: Step-by-step guide
To use Flux Redux in ComfyUI, download these:
- Flux Redux safe tensors
- Sigclip Vision file.
The presenter includes these links in the video description.
Download and place the flux-redux-dev safe tensors into the style models folder.
Next, place sigclip vision files into the Clip Vision folder in ComfyUI models.
You can alternatively install it using the comfy UI manager. In Comfy UI's manager, select custom nodes, and search for 'Apply style model adjust'.
When it comes to setting up the workflow, the presenter describes the process of using different nodes such as Clip Vision Loader and the Flex Guidance tool. He also describes how the strength setting on the model has an inverse relationship with adherence to the prompt; lower settings tend to adhere more closely to the prompt instructions, while higher settings emphasize characteristics from the initial image.
The workflow includes:
- CLIP Vision Loader: Loads the specified vision model
- CLIP Vision Encode: Encodes the image data
- Load Style Model: Loads specified style data.
- Apply Style Model (Adjusted): Adjusts the impact of the style on the diffusion. The user states that values closer to zero follow the prompt more carefully. 0.3 to 0.6 balance between both sides of the equation, and values from 0.6 to 1 follow the init image much closer.
- Flux Guidance Use flux guidance to increase or decrease prompts for various outcomes.
Flux Redux in ComfyUI allows for a seamless and intuitive workflow. Remember to restart ComfyUI in order for these settings to function correctly.