Distribution Matching Enables Better Use of Pretrained Diffusion Priors for Super-Resolution

2000, Jan 01    

🔗 Links

• 📄 Paper — (ToDo)
• 💻 Code — (ToDo)
• 🧪 Colab Notebook — (ToDo)
• 📦 Model Weights — (ToDo)

---

Distribution Matching Super-Resolution (DM-SR)

DM-SR aims at developing a Practical Algorithm for Real-world Super-Resolution. DM-SR leverages the generative capabilities of a pretrained diffusion model, while requiring only a single step to perform super-resolution.

---

Demo

The following images show a comparison between our DM-SR results and bicubic upsampling. You can drag the slider to visually compare the two methods side by side.

---

Visual Comparisons

The left half shows the bicubic upsampled image, while the right half displays the super-resolved result produced by our DM-SR method. Please zoom in for better view.