Real-Time Stable Diffusion-Based Image Super-Resolution. AdcSR super-resolves a 128×128 image to 512×512 in just 0.03s 🚀 on an A100 GPU

With Cog (recommended)

cog predict -i image=@input.jpg

From Python

from predict import Predictor
predictor = Predictor()
predictor.setup()
result = predictor.predict(image="input.jpg")

⚙️ Main Parameters

• image: Input image (required)
• scale_factor: Fixed at 4x (also supports 2x)
• No text prompt required

📝 Usage Tips

1. Best Results: Use with natural photographs and realistic images
2. Input Quality: Higher quality inputs produce better outputs
3. File Size: Consider input image size for optimal processing speed
4. Batch Processing: Process multiple images for maximum efficiency

📝 Important Notes

• Model weights (net_params_200.pkl) are downloaded automatically from HuggingFace
• AdcSR is optimized for 4x, but also supports 2x
• If AdcSR fails, bicubic is used as fallback

📄 License

MIT License – see LICENSE

Increase resolution and details up to (4k, 8k), reach out to me on twitter.

Follow me on Twitter/X | Website

🙏 Credits

Original Research: AdcSR - Adversarial Diffusion Compression for Real-World Image Super-Resolution

Model Authors: - Research team behind the AdcSR methodology - Original implementation and training - HuggingFace Model

Base Architecture: - Stable Diffusion (Stability AI) - Diffusion model components from the diffusers library

This model is intended for research and creative applications. Please ensure appropriate licensing for commercial use.