README.md

metadata

title: UGC Photo Generation
emoji: 📸
colorFrom: purple
colorTo: pink
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: apache-2.0

🎨 Unified UGC Platform - Photogen1

A unified ComfyUI-powered platform for ultra-realistic UGC photo generation with face enhancement capabilities.

🚀 Features

• Authentic UGC Style: Generate photos that look like real smartphone photography
• Face Enhancement: Improve facial details using GFPGAN
• Multiple Workflows: Portrait, Full Body, Street Photography, and FLUX styles
• Authenticity Stack: Carefully curated models and LoRAs for maximum realism
• Real-time Generation: GPU-accelerated ComfyUI backend

🎯 The Authenticity Stack

This platform uses a carefully selected stack of models designed to produce images that look like genuine user-generated content:

Base Models

• RealVisXL V4.0: Foundation checkpoint for exceptional photorealism
• SDXL VAE: Standard VAE for accurate color and light rendering

LoRA Stack (Applied in sequence)

1. epiCRealism - Natural photographic (0.6): Creates the "phone camera" aesthetic
2. Detail Tweaker XL v2 (0.4): Adds realistic skin texture and fabric details
3. Film Grain Helper SDXL (0.3): Introduces authentic camera sensor noise

Face Enhancement

• GFPGANv1.4: For face restoration and enhancement

📁 Project Structure

photogen1/
├── app.py                          # Main application
├── requirements.txt                # Python dependencies
├── packages.txt                    # System packages for HF Spaces
├── setup.sh                        # Setup script (clones repos & downloads models)
├── README.md                       # This file
├── workflows/
│   ├── organic_portrait_ugc.json  # Main UGC workflow with Authenticity Stack
│   ├── organic_portrait_workflow.json
│   ├── face_enhancement_workflow.json
│   ├── full_body_workflow.json
│   ├── street_photo_workflow.json
│   └── flux_workflow.json
├── models/
│   ├── checkpoints/               # Base models
│   ├── loras/                     # LoRA models  
│   └── vae/                       # VAE models
└── outputs/                       # Generated images

🛠️ Setup

The application automatically:

1. Downloads all required models on first run
2. Clones ComfyUI and necessary custom nodes
3. Sets up the complete environment

No manual setup required!

🎨 Optimal Settings for UGC

Pre-configured for authentic results:

• Sampler: DPM++ 2M SDE Karras (adds organic randomness)
• CFG Scale: 5.0 (lower values = more natural, less "AI" look)
• Steps: 25-30 (sufficient for quality without overprocessing)

💡 Usage Tips

For Best UGC Results

• Use prompts that describe real scenarios: "selfie in bedroom, iPhone photo, natural light"
• Include imperfections: "slightly messy background", "casual outfit"
• Avoid professional terms: Skip words like "professional", "studio", "perfect"

Example Prompts

Good: "woman taking mirror selfie in bathroom, smartphone photo, morning light, no makeup, oversized t-shirt"

Avoid: "professional portrait of beautiful model, studio lighting, high quality"

🚀 Performance

• T4 GPU: ~30-45 seconds per image
• A10G GPU: ~15-20 seconds per image
• Face enhancement: +10-15 seconds

🐛 Troubleshooting

• Model download fails: Check internet connection, HF Space will retry
• OOM errors: Reduce image dimensions or batch size
• Slow generation: Normal on free tier, upgrade to faster GPU

📄 License

Apache 2.0