What are Flux.1 Tools?

For Flux.1, derivative models equivalent to ControlNet and IP-Adapter have been released by Flux officially, separate from the base model.

  • FLUX.1 Fill ... Model for inpainting / outpainting
  • FLUX.1 Depth / FLUX.1 Canny ... Models that redraw while keeping the shape with structure-based guides (Depth / Canny)
  • FLUX.1 Redux ... An IP-Adapter-like model for Flux that mass-produces variations that look exactly like the reference image

FLUX.1 Fill

It can be used just like an inpainting model.

Model Download

📂ComfyUI/
└── 📂models/
    └── 📂diffusion_models/
        └── FLUX.1-Fill-dev_fp8.safetensors

Workflow

FLUX.1-Fill.json

  • 🟪 Load flux1-fill-dev.safetensors with the Load Diffusion Model node.

  • 🟩 Even specifically for Flux.1 workflows, add the InpaintModelConditioning node just like in Stable Diffusion 1.5 inpainting.

    • Input the image and mask.

FLUX.1 Depth / FLUX.1 Canny

It can be used with the same feeling as ControlNet Depth / Canny.

Model Download

📂ComfyUI/
└── 📂models/
    └── 📂diffusion_models/
        ├── flux1-depth-dev-fp8.safetensors
        └── flux1-canny-dev-fp8.safetensors

Workflow

  • 🟩 Input the control image to the InstructPixToPixConditioning node.

  • 🟦 Since it is Depth this time, create a depth map with Depth Anything V2.

    • Since the image size of this depth map becomes the output image size as is, resize it to an appropriate size.

For the Canny version, input the Canny edge image with the same configuration.

FLUX.1 Redux

FLUX.1 Redux is a model where you pass one or more reference images and generate variations strongly biased towards those images. It is close to IP-Adapter, but Redux is quite resistant to prompts, and the appearance of the reference image comes out almost as is.

Model Download

Redux is loaded as a "style model" separate from the Flux body. Furthermore, CLIP-ViT for encoding reference images is also required.

📂ComfyUI/
└── 📂models/
    ├── 📂style_models/
    │   └── flux1-redux-dev.safetensors
    └── 📂clip_vision/
        └── sigclip_vision_patch14_384.safetensors

Workflow

FLUX.1-Redux.json

  • 🟩 Add the Apply Style Model node and connect the Style model and CLIP Vision Encode.

    • Connect sigclip_vision_patch14_384.safetensors and the reference image to the CLIP Vision Encode node.

Mixing Multiple Images

If you line up the chunks of Apply Style Model horizontally, you can also reference and mix multiple images.

FLUX.1-Redux_multi.json

Problem with Redux

Again, Redux ignores most other parameters such as prompts and LoRA.

I think it is best to use it as a "tool to mass-produce variations that look exactly like the reference image" by cutting it off like that, but there are also custom nodes that allow some control with prompts. For your reference.