1. Workflow Overview

This workflow specializes in controlled image generation with:

1. Multi-Modal Inputs: Text prompts + reference images

2. Precision Control: Flux sampling parameters + LoRA adjustments

3. Bilingual Processing: Integrated Baidu translation for prompts

4. Comparative Outputs: Side-by-side result analysis

Key Applications:

• Advertising material generation

• Social media content creation

• Product visualization

2. Core Models

3. Critical Nodes

Core Processing Nodes

Special Dependencies

1. Flux Models:

   git clone https://github.com/FluxAI/ComfyUI-Flux

2. Translation Service:

   • Requires Baidu API key

   • Configure in config/baidu_translate.json

4. Workflow Architecture

Processing Stages

Data Flow

graph LR
A[Text Prompt] --> B[CLIP Encoding]
C[Reference Image] --> D[Latent Space]
B --> E[Flux Sampling]
D --> E
E --> F[VAE Decode]
F --> G[Result Comparison]

5. I/O Specifications

Input Requirements

1. Text Prompts:

   • Chinese/English bilingual support

   • Example: "Urban fashion photoshoot with graffiti backdrop"

2. Image Inputs:

   • Recommended 1024x1024 PNG

   • Alpha channel for masking

Outputs

• Primary Image: High-res generated result

• Parameter Analysis: Visualized sampling metrics

• Bilingual Captions: JSON metadata

6. Optimization Guide

1. Performance Tweaks:

   # In custom_nodes/flux_sampler.py:
   torch.set_float32_matmul_precision('high')

2. Quality Adjustment:

   • Flux guidance_scale: 3.5-5.0

   • LoRA strength: 0.7-1.0

3. Troubleshooting:

   • Blurry outputs: Increase steps (20→30)

   • Over-saturation: Reduce cfg_scale (3.5→2.5)

7. Deployment

Step 1: Dependency Installation

pip install baidu-aip torchvision>=0.15

Step 2: Model Placement

• Flux models: models/flux

• LoRAs: models/loras

Verification Command

# Check translation service
from aip import AipNlp
print(AipNlp('APP_ID','API_KEY','SECRET_KEY').detectLang("test"))

Real-World Use Case

Scenario: Sneaker ad generation

1. Input:

   • Prompt: "限量版运动鞋，未来主义设计，霓虹灯光效"

   • Blank canvas 1024x1024

2. Process:

   • Applies hyper-realistic texture LoRA

   • Generates 3 style variants

   • Outputs English/Chinese captions

3. Output:

   {
     "image": "sneaker_ad_final.png",
     "caption_en": "Limited edition sneakers with cyberpunk neon lighting",
     "parameters": {"steps":25, "cfg":3.8}
   }

Processing Time: ~38s (RTX 4090)

Note: Ideal for marketing teams needing rapid visual prototyping with precise control.