SPACE_README.md

metadata

title: ResShift Super-Resolution
emoji: 🖼️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.0.0
app_file: app.py
pinned: false
license: mit

ResShift Super-Resolution

Super-resolution using ResShift diffusion model. Upload a low-resolution image to get an enhanced, super-resolved version.

Features

• 4x super-resolution using diffusion model
• Works in latent space for efficient processing
• Full diffusion sampling loop (15 steps)
• Real-time inference with Gradio interface

Usage

1. Upload a low-resolution image
2. Click "Super-Resolve" or wait for automatic processing
3. Download the super-resolved output

Model

The model is trained on DIV2K dataset and uses VQGAN for latent space encoding/decoding.

Technical Details

• Architecture: U-Net with Swin Transformer blocks
• Latent Space: 64x64 (encoded from 256x256 pixel space)
• Diffusion Steps: 15 timesteps
• Scale Factor: 4x

Citation

If you use this model, please cite the ResShift paper.