1. Workflow Overview

This workflow is a multi-ControlNet template designed for effortless control, featuring:

• Dynamic ControlNet Switching: Toggle between HED/Depth/Canny via ImpactSwitch.

• Auto-Prompting: Uses JoyCaption (Llama-3 based) to generate prompts from images.

• FP8 Optimization: Low-VRAM models like flux_dev.safetensors for consumer GPUs.

Key Models:

• Flux-dev-FP8: Base generation model (FP8 quantized).

• ControlNet-v3 Series: Includes HED/depth/canny control models.

• Depth-Anything-V2: Depth map generator (requires separate install).

2. Key Nodes

ControlNet Modules:

1. LoadFluxControlNet

   • Loads HED/depth/canny models from models/controlnet/.

2. ApplyFluxControlNet

   • Applies ControlNet with adjustable weights (HED:0.4, Depth:0.8, Canny:0.3).

Preprocessors:

• HEDPreprocessor: Soft-edge maps (ideal for shape retention).

• DepthAnything_V2: Depth maps (install ComfyUI-Depth-Anything).

• CannyEdgePreprocessor: Edge sketches (for detail control).

Custom Nodes:

1. JoyCaption (Install via ComfyUI Manager)

   • Generates prompts using Llama-3.

2. ImpactSwitch

   • Switches between ControlNet inputs (enable Multi-Input).

3. Workflow Structure

4. Inputs & Outputs

Inputs:

• Required: Upload image via LoadImage (e.g., .webp in example).

• Optional:

  • Resolution (default 1024x1024, edit in EmptyLatentImage).

  • ControlNet weights (adjust in ApplyFluxControlNet).

Outputs:

• Images saved to /ComfyUI/output/.

• Real-time previews of HED/depth/canny intermediates.

5. Notes

1. Model Dependencies:

   • Download depth_anything_v2_vitl_fp32.safetensors to models/depth_anything.

   • Flux models require official download (e.g., flux-dev-fp8).

2. VRAM:

   • 8GB+ GPU recommended. Disable unused ControlNet branches for low-end GPUs.

3. Troubleshooting:

   • Poor auto-prompts: Switch JoyCaption to Descriptive/Long mode.

   • Control failure: Verify ImpactSwitch selection (HED=1, Depth=2, Canny=3).