1. Workflow Overview

• Purpose: Automatically reverse-engineer prompts from input images and redraw them with stylized LoRAs (e.g., anime style).

• Key Features:

  • Captioning: Uses Florence-2-base-PromptGen-v2.0 to generate tags.

  • Redraw: Applies Flux洗图神器.safetensors LoRA and svdq-int4-flux.1-dev base model for image-to-image generation.

  • Custom Control: Supports manual trigger words (e.g., kk3d ip) fused with auto-generated tags.

2. Key Nodes

3. Workflow Groups

1. Captioning Group

   • Input: User-uploaded image (e.g., ComfyUI_temp_btare_00003_.png).

   • Output: Tags like 1girl, solo, long hair....

2. Prompt Fusion Group

   • Merges auto-tags with manual triggers via LayerUtility: TextJoin.

3. Generation Group

   • Models: svdq-int4-flux.1-dev + Flux洗图神器 LoRA (weight=0.8).

   • Output: Resolution matches input (default 512x512).

4. Inputs & Outputs

• Inputs:

  • Image (required).

  • Optional trigger words (e.g., kk3d ip).

  • Denoising strength: Default 0.7 (via BasicScheduler).

• Outputs:

  • Generated image (saved to ComfyUI/output).

  • Caption text (displayed in ShowText node).

5. Tips & Warnings

• ⚠️ Compatibility:

  • Requires ComfyUI-nunchaku and ComfyUI-Florence2 extensions.

  • Florence-2 model must be manually downloaded.

• ✅ Optimization:

  • Use svdq-int4 for low-VRAM GPUs (8GB).

  • Adjust FluxGuidance strength (default: 3.5) for style intensity.