1. Workflow Overview

This workflow is designed for multi-element fusion, combining reference images, style models, and conditional controls. Key features:

• Multi-Image Fusion: 3 image+mask inputs via ReduxAdvanced nodes.

• Advanced Control: FluxGuidance adjusts generation strength (default=3.5).

• Custom Models: Independent UNet, CLIP, and VAE loading.

Core Models:

• UNet: flux1-dev.sft (FP8 optimized)

• CLIP: Dual encoders (ViT-L-14-TEXT-detail + t5xxl_fp8)

• Style Model: flux1-redux-dev.safetensors

• VAE: ae.sft

2. Node Breakdown

Dependencies:

• Models:

  • Style Model: flux1-redux-dev.safetensors

  • CLIP Vision: google--siglip-so400m-patch14-384/model.safetensors

• Plugins:

  • Redux Plugin: For image fusion (GitHub).

  • Flux Plugin: Advanced conditioning.

3. Workflow Groups

1. Input Group:

   • Images: 3 sets of images+masks (via LoadImage).

   • Text: Dual CLIP encoders for prompts.

2. Fusion Group:

   • ReduxAdvanced Nodes:

     • Input: Image+Mask+Style Model+CLIP Vision.

     • Output: Fused conditioning.

3. Generation Group:

   • Sampling: Euler, 30 steps, CFG=1.

   • Output: 1024x1024 image (decoded by VAEDecode).

4. Inputs & Outputs

• Inputs:

  • Images: ≥1 reference image (3 recommended with masks).

  • Prompts: Positive/negative text via CLIPTextEncode.

  • Strength: FluxGuidance parameter (default=3.5).

• Output: HD image (1024x1024) saved to ComfyUI/output.

5. Notes

1. VRAM: ≥12GB GPU required for multi-image fusion.

2. Model Compatibility: Ensure flux1-redux-dev.safetensors matches UNet version.

3. Troubleshooting:

   • Missing Node: Install Redux/Flux plugins.

   • Image Size Mismatch: Uniform resolution required.

4. Optimization: Use FP8 models (e.g., t5xxl_fp8) to reduce VRAM usage.