1. Workflow Overview

• Purpose: Fuses user-uploaded images (e.g., product photos) with abstract fluid art styles to generate high-quality backgrounds for PPTs/posters/webpages, with style control and upscaling.

• Core Models:

  • Stable Diffusion XL: Base image generation.

  • 4x-UltraSharp: Image super-resolution.

  • Meta-Llama-3.1-8B: Auto-generates image captions.

  • Florence-2: Multimodal image analysis.

2. Key Nodes & Installation

1. StyleModelApply

   • Function: Transfers style from reference image (strength adjustable, default=0.3).

   • Install: Built-in; requires flux1-redux-dev style model.

2. UltimateSDUpscale

   • Function: Tile-based upscaling to 1536px (avoids OOM).

   • Install: Install Impact Pack via ComfyUI Manager.

3. Joy_caption_two

   • Function: Generates descriptive text via Llama-3.

   • Install: Manual install of unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit.

Dependencies:

• LoRA Models: e.g., Dynamic Abstract Background v1.0 (download from LibLibAI).

3. Workflow Structure

Group 1: Image Input & Preprocessing

• Input: User uploads via LoadImage.

• Process: Resize to 1024x1024 (ImageResizeKJ), extract depth (DepthAnythingV2Preprocessor).

Group 2: Style Fusion & Generation

• Core Nodes:

  • StyleModelApply: Style blending (controlled by Float node).

  • KSampler: Image generation (Euler, 25 steps, CFG=10).

• Output: Latent representation.

Group 3: Super-Resolution

• Flow: VAE decode → Tile upscale (UltimateSDUpscale) → HD output.

Group 4: Text Generation

• Pipeline: Florence2Run analysis + Llama-3 captioning.

4. Inputs & Outputs

• Required Inputs:

  • Source image (e.g., product photo).

  • Style reference image (e.g., fluid art).

  • Style strength (default=0.15, recommended≤0.2).

• Outputs:

  • HD fused image (PNG).

  • Text description (e.g., "Abstract fluid style, gold-to-black gradient").

5. Notes

• VRAM: ≥12GB GPU; use --medvram for upscaling.

• Common Issues:

  • Over-stylization: Reduce StyleModelApply strength.

  • Resolution overflow: Ensure input image ≤2048px on the longer side.

• Optimization:

  • Use TAESD decoder for faster previews.

  • Disable unused groups (e.g., text generation) to save resources.