1. Workflow Overview

This workflow focuses on image generation comparison & optimization, featuring:

1. Dual-path Generation: Parallel outputs from Flux DiT and HiDream Sampler for side-by-side comparison.

2. Smart Prompting: Auto-generates prompts from input images via BizyAirJoyCaption2.

3. VRAM Optimization: Integrated PurgeVRAM nodes for low-end GPUs.

4. Output Fusion: Merges results into a comparison grid.

2. Core Models

3. Key Nodes

4. Workflow Groups

• Group 1: Input & Prompting

  • Input: Uploaded image (e.g., c298b158...jpg)

  • Output: Generated prompt (e.g., "Pirate ship in a cosmic maelstrom")

• Group 2: Dual-path Generation

  • Flux Path: Uses Flux DiT + T5 encoder

  • HiDream Path: Fast generation via HiDreamSampler

  • Key param: seed=114129529 (fixed for consistency)

• Group 3: Output Fusion

  • ImageConcanate merges images → Saves to ComfyUI/output

5. Inputs & Outputs

• Input Parameters:

  • Required: Image file, Resolution (default: 1024x1024)

  • Optional: HiDream steps, Flux quant strength (0.105)

• Output:

  • Comparison grid image (saved as PNG)

6. Notes

1. Compatibility:

   • Flux DiT requires nunchaku-fp16 config files.

   • HiDream does not support dynamic resolution.

2. VRAM Management:

   • Enable PurgeVRAM if GPU <8GB to avoid crashes.

3. Optimization:

   • Flux's svdq-int4 trades quality for speed—adjust as needed.