1. Workflow Overview

This is a FLUX-based multimodal workflow supporting both text-to-image and image captioning, featuring:

• Llama-3 8B for automatic image description

• FLUX MIX V2 as base model (FP8 quantized)

• Multi-LoRA stacking (Shining Nikki series)

• Built-in Baidu translation node

Core Models:

• FLUX MIX V2: Base generation model

• Meta-Llama-3-8B: Image captioning model

• Shining Nikki LoRAs: Costume style control

2. Critical Nodes

Dependencies:

• Model Files:

  • Place 小白_FLUX_MIX_V2.safetensors in models/unet

  • Download t5xxl_fp8_e4m3fn CLIP model

• Plugins: Essential to install Impact Pack

3. Workflow Structure

Key Groups:

1. Prompt Input (Top-Left):

   • Supports CN/EN input + auto translation

   • Dynamic LoRA trigger word insertion

2. FLUX Generation (Center):

   • 896x1152 resolution + Euler sampler

   • Three-stage LoRA stacking (weight 0.2~0.8)

3. Output (Right):

   • Batch generation (max 3 for non-VIP)

   • Auto-upload to Liblib cloud

4. Inputs & Outputs

Required Inputs:

• Positive prompt: Chinese directly accepted

• Optional image: For captioning (toggle② must be off)

• LoRA selection: Dropdown menu

Final Output:

• Format: JPG/PNG with metadata

• Path: Auto-saved to Liblib cloud

5. Notes

• VRAM: ≥16GB required (FP8 + multi-LoRA)

• Common Errors:

  • LoRA not found: Avoid spaces in filename

  • Translation failed: Configure Baidu API key

• Optimization:

  • Disable unused LoRA groups

  • Reduce resolution to 768x1024 for batch generation