1. Workflow Overview

This workflow specializes in local inpainting and style transfer, leveraging ICEdit (In-Context Editing) and Flux models to:

• Edit Specific Areas: Modify masked regions (e.g., changing couch color to sky blue).

• Apply Artistic Styles: Support Van Gogh-style transformations.

• Auto-Upscaling: Default 512px output with ESRGAN enlargement.

Core Models:

• Flux Fill dev: Official redraw model (FP8 optimized).

• ICEdit-LoRA: Fine-tuned for local edits.

• T5-XXL & CLIP-L: Dual-text encoders for multilingual prompts.

2. Key Nodes

Built-in Nodes:

• DiptychCreate: Splits input image into original + mask pairs.

• ICEFConditioning: Fuses text instructions with image context.

• FluxGuidance: Controls semantic guidance intensity (default=50).

Custom Nodes:

1. LibLibTranslate (Install via ComfyUI Manager)

   • Translates Chinese prompts to English (e.g., "梵高画风" → "Van Gogh style").

2. easy cleanGpuUsed (From ComfyUI-Easy-Use)

   • Prevents OOM errors by clearing VRAM.

Model Dependencies:

• ESRGAN_4x: Upscale model loaded via UpscaleModelLoader.

• ae.sft: Custom VAE (place in models/vae).

3. Workflow Structure

4. Inputs & Outputs

Required:

• Image: Upload via DiptychCreate (e.g., 原图2.png).

• Prompt: Enter Chinese in LibLibTranslate (e.g., "Van Gogh style").

Optional:

• Seed: Randomized by default (editable in KSamplerAdvanced).

• Guidance Scale: Adjust in FluxGuidance (affects style strength).

Output:

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

• Real-time preview via PreviewImage.

5. Notes

1. VRAM: Recommend 12GB+ GPU. Disable ImageUpscaleWithModel for low-end GPUs.

2. Troubleshooting:

   • Missing VAE: Ensure ae.sft is in models/vae.

   • Mask issues: Verify DiptychCreate mask channels.

3. Optimization:

   • Reduce KSamplerAdvanced steps (default 28 → 20).

   • Use easy cleanGpuUsed to free VRAM periodically.