1. Workflow Overview

This workflow, named "Flux+ICEdit Single-Command Image Editing with Consistency", is designed for text-guided image editing. It allows modifying images via natural language (e.g., changing outfits, colors, or accessories) while preserving structural consistency. Key features:

• Image Editing: Uses ICEdit-MoE-LoRA for localized/global edits.

• Text Guidance: Accepts Chinese commands (auto-translated to English).

• Upscaling: Enhances resolution with 4x-UltraSharp.

• Style Control: Applies StyleModel (e.g., flex1_redux_siglip2_512).

Core Models:

• Stable Diffusion: Base generative model (loaded via UNETLoader).

• ICEdit-MoE-LoRA: Fine-tuned LoRA for editing.

• CLIP Vision (sigclip_vision_patch14_384): Encodes image semantics.

• 4x-UltraSharp: Upscaling model.

2. Key Nodes

1. ICEdit-MoE-LoRA (Node 345)

   • Function: Loads the editing-specific LoRA.

   • Installation: Manually download .safetensors to models/loras.

2. ACEPlusLoraProcessor (Node 472)

   • Function: Processes input image/mask for edits.

   • Dependency: Requires ComfyUI-ICEdit (GitHub).

3. DeepTranslatorTextNode (Node 424)

   • Function: Translates Chinese commands to English (Google API).

   • Install: Via ComfyUI Manager (search DeepTranslator).

4. StyleModelApply (Node 443)

   • Function: Applies style models (e.g., flex1_redux_siglip2_512).

5. ImageUpscaleWithModel (Node 439)

   • Function: Upscales images with 4x-UltraSharp.

Special Dependencies:

• Model Files:

  • ICEdit-MoE-LoRA: Download from CivitAI.

  • sigclip_vision_patch14_384: Available on Hugging Face.

3. Workflow Structure (Groups)

1. Group 1: Upload Image (Node 240)

   • Input: Original image (e.g., 768-1152.png).

   • Output: Image data for editing.

2. Group 2: Load Models (Nodes 329, 345, 32, 227)

   • Loads Stable Diffusion, ICEdit-LoRA, VAE, CLIP, etc.

3. Group 3: Text Command (Nodes 461, 424, 114)

   • Input: Natural language command (e.g., "Change to yoga pants").

   • Output: Translated English prompts for CLIPTextEncode.

4. Group 4: Generate & Upscale (Nodes 454, 439, 437)

   • Edits image → Upscales → Downscales for optimization.

5. Group 5: Output (Nodes 493, 464)

   • Output: Comparison view (original vs. edited) and saved final image.

4. Inputs & Outputs

Inputs:

• Original image (e.g., 768x1152 resolution).

• Chinese text command (e.g., "Change shirt color").

• Seed value (fixed or randomized).

Outputs:

• Edited high-res image (saved as PNG).

• Slider comparison tool (original vs. edited).

5. Notes

1. Common Issues:

   • Missing ICEdit-LoRA → Check models/loras.

   • GPU OOM → Reduce resolution or use --medvram.

2. Optimization:

   • Enable --xformers for faster inference.

   • Increase KSampler steps (default: 20) for complex edits.

3. Compatibility:

   • Requires ComfyUI ≥ 0.3.30 and PyTorch 2.0+.

   • ACEPlusLoraProcessor needs manual dependency install.