1. Workflow Overview

This workflow leverages FLUX ControlNet V3.0 for multi-condition controlled generation, combining HED soft-edge, Depth, and Canny edge preprocessing to precisely guide image synthesis. Outputs adhere to input structure and text prompts.

Key Features:

• Multi-ControlNet Integration (HED + Depth + Canny).

• Depth Anything V2 for depth map generation.

• WD14 Tagger for auto-prompt generation.

• Flux Sampler with FP8 optimization for efficiency.

Output:

• High-res images (default 1024x1024), stylized by prompts and ControlNet conditions.

2. Core Models

3. Key Nodes & Installation

4. Workflow Groups

1. HED Group (Blue)

   • Input: Source image (resized via ImageResize+).

   • Preprocess: HEDPreprocessor extracts soft edges.

   • Control Weight: 0.8 (set in ApplyFluxControlNet).

2. Depth Group (Orange)

   • Input: Same image.

   • Preprocess: DepthAnything_V2 generates depth map.

   • Control Weight: 0.7.

3. Canny Group (Purple)

   • Input: Same image.

   • Preprocess: CannyEdgePreprocessor (thresholds 100/200).

   • Control Weight: 0.6.

4. Generation Group

   • Prompts: Processed by CLIPTextEncode (e.g., "Makoto Shinkai style").

   • Sampling: XlabsSampler merges multi-ControlNet conditions.

5. Inputs & Outputs

Input Parameters:

• Image: Loaded via LoadImage (e.g., sample Redbook image).

• Prompts: Manual input or auto-generated by WD14Tagger.

• Resolution: Default 1024x1024 (set in EmptyLatentImage).

Output:

• Final images saved in /ComfyUI/output/ with metadata.

6. Notes

1. Hardware:

   • RTX 40 series recommended (FP8 support), VRAM ≥12GB.

   • Depth Anything V2 is VRAM-intensive; may crash at high resolutions.

2. Model Setup:

   • Download ControlNet V3 models to /models/controlnet/.

   • Missing models trigger download prompts.

3. Tips:

   • Total ControlNet weights should ideally ≤2.0 (e.g., HED 0.8 + Depth 0.7 + Canny 0.6).

   • FP8 mode may introduce noise; adjust denoise in ModelSamplingFlux.