1. Workflow Overview

This workflow is designed for multi-image fusion and semantic redrawing, leveraging FluxKontextMaxImageNode to merge input images based on text prompts. Key applications include:

• Person-scene composition (e.g., blending portraits with sculptures/architecture)

• Creative collages (e.g., KTV scenes, wedding photo background replacement)

• Image upscaling & detail enhancement

2. Core Models

• Flux Dev Model: Base generation model with integrated 4-step Lora for fast inference.

• 4x_NMKD-Siax_200k: Upscaling model for resolution enhancement.

• Bilingual CLIP Encoder: Supports EN/CN prompts (loaded via DualCLIPLoader).

3. Key Nodes

4. Workflow Structure

• Group 1: Dual-Person Pose

  • Input: Girl portrait + David sculpture

  • Process: ImageConcanate merge → FluxKontext generates "embracing with church background"

• Group 2: Person-Scene Merge

  • Input: Bride + Castle image

  • Process: Semantic prompt-guided fusion

• Group 3: Upscale & Repair

  • Pipeline: ImageUpscaleWithModel → TTP_Image_Tile tiling → Reassembly

5. Inputs & Outputs

• Input Parameters:

  • Images: Multiple formats (.png, .jpeg)

  • Prompts: Bilingual (auto-translated)

  • Seed: Fixed or randomized

• Output: High-res PNG (default 3:4 or 16:9 ratio)

6. Notes

1. VRAM: ≥12GB recommended; use --medvram for large outputs.

2. Plugin Dependencies:

   • FluxKontext: GitHub Link

   • Tiled Diffusion: Install via ComfyUI Manager

3. Troubleshooting:

   • Chinese prompts must be translated via LibLibTranslate

   • Resize mismatched images with ImageResizeKJ node