Adjustable Real-World Super-Resolution
We’ve all got them: a blurry image saved off the web, a grainy screenshot, an old phone photo that’s lost its crispness. PiSA-SR makes them bigger and clean — stripping out blur, noise, and compression artifacts in one pass. Its party trick is a dial: one module removes damage, another adds fine detail, and you decide how much of each at run time. Want it natural and smooth? Turn it down. Want maximum sharpness? Turn it up. No retraining required.
🏆 Leads OSEDiff, StableSR, and DiffBIR on every perceptual metric — CLIPIQA up to 0.74–0.75 and LPIPS as low as 0.235 on RealSRSet + DIV2K
What it does
PiSA-SR turns blurry, low-resolution images into sharp, high-quality 2× / 4× upscaled versions. Under the hood it uses two specialized modules: one works at the pixel level to remove noise and artifacts, the other works at the semantic level to add perceptual detail. The key is that you control the balance between the two at inference time — lean toward faithful cleanup or toward richer detail — all from a single trained model with no retraining.
Problem it solves
- Downloaded low-res image from the web – Make it usable size with quality restored
- Screenshot or thumbnail from a video – Upscale and clarify for better viewing
- Old camera or phone photo – Breathe new life into aged/degraded photos
- Intentionally compressed file to save space – Recover quality from compressed original
- Video frame extracted at low quality – Restore sharpness for better detail
- Real-world degraded photos – Handles blur, noise, compression, downsampling all at once
- Flexible quality preference – Customize output between detail-focused and natural-looking
Input/Output
- Input:
- A low-resolution RGB image (real-world blur, noise, compression, downsampling)
- Output:
- A high-resolution RGB image (2× / 4× dimensions) with restored sharpness, removed artifacts, and enhanced detail
Performance
How PiSA-SR compares
Benchmarked against leading diffusion-based super-resolution models on RealSRSet + DIV2K validation (real-world SR):
Metric | PiSA-SR | OSEDiff | StableSR | DiffBIR |
LPIPS ↓ | 0.235–0.240 | 0.270 | 0.285 | 0.295 |
DISTS ↓ | 0.285–0.290 | 0.325 | 0.345 | 0.355 |
CLIPIQA ↑ | 0.74–0.75 | 0.70 | 0.68 | 0.67 |
MUSIQ ↑ | 0.58–0.59 | 0.55 | 0.52 | 0.50 |
PiSA-SR also has three siblings inside CNAPS Studio’s Image Upscaling lineup — Swin2SR, SwinIR, and LatticeNet. Compare them side-by-side.
Model Source
- GitHub Repository: https://github.com/csslc/PiSA-SR (official, CVPR 2025)
- Paper: Liang et al., "Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach" arXiv:2412.03017, CVPR 2025 https://arxiv.org/abs/2412.03017
Compliance & Provenance
Provider | Open-source |
Provider type | Specialized |
License | |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.