1. Workflow Overview

This workflow leverages ICEdit-LoRA to generate/edit designs via natural language commands (e.g., "Turn this man into Hulk"). It combines text-guided generation and local inpainting, ideal for ad concepts or art prototyping.

2. Core Models

• Stable Diffusion (UNETLoader): Base model flux1-fill-dev.safetensors (FP8 optimized).

• ICEdit-MoE-LoRA: Enables precise local edits (strength=1.0).

• Dual Text Encoders: clip_l.safetensors + t5xxl_fp8_e4m3fn for complex prompts.

• VAE: ae.safetensors for latent-to-image decoding.

3. Key Nodes

4. Workflow Structure

• Group 1: Text & Model Loading

  • DualCLIPLoader: Loads dual text encoders.

  • CLIPTextEncode: Processes voice commands.

• Group 2: Generation & Editing

  • KSampler: Generates images (20 steps, Euler).

  • InpaintModelConditioning: Applies local edits.

• Group 3: Output & Optimization

  • VAEDecode: Renders final image.

  • easy cleanGpuUsed: Auto-clears GPU memory.

5. Inputs & Outputs

• Input Parameters:

  • Required: Voice command text, base image (via LoadImage).

  • Optional: Mask area, CFG (default=1), seed (random).

• Output: Edited PNG (ComfyUI_XXX.png).

6. Notes

1. Plugins: Install ICEdit and comfyui-custom-scripts.

2. Model Paths: Place ICEdit-MoE-LoRA.safetensors and VAE in correct folders.

3. Performance: ≥10GB VRAM recommended; FP8 reduces usage.

4. Debug: Verify mask coverage if edits fail.