📸

Image Upscaling - PiSA-SR

🛠

Adjustable Real-World Super-Resolution

Makes low-resolution images bigger and crystal clear. Removes blur, noise, and compression artifacts. Perfect for downloaded web images, old photos, screenshots, video frames, and compressed files.

What it does

PiSA-SR transforms blurry, low-resolution images into sharp, high-quality 2x / 4× upscaled versions. Uses two specialized AI modules: one removes noise and artifacts (pixel-level), the other adds perceptual detail enhancement (semantic-level). You control the balance between these two at inference time. Prefer natural and smooth? Dial it back. No retraining needed.

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: Low-resolution RGB image
    • Degradation: real-world blur, noise, compression, downsampling
      • image
  • Output: High-resolution RGB image (2x / 4× dimensions)
    • Quality: Restored sharpness, removed artifacts, enhanced detail
      • image
      image

Performance

  • Benchmark Performance (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
  • Speed & Efficiency (×4 on 128px, NVIDIA A100 80GB)
    • Inference: 0.32s (default), 0.64s (adjustable)
    • Parameters: ~50M (LoRA only vs. 860M base SD)
    • Memory: ~8GB (A100)

Technical Details

Architecture
Stable Diffusion 2.1-Base + Dual-LoRA
Key Innovations
Two independent LoRA modules control pixel-level and semantic-level enhancement separately, allowing users to adjust the balance between fidelity and detail during inference.
Core Components
• Base Model: Stable Diffusion 2.1 (text-to-image diffusion model) • Pixel-level LoRA: Removes degradations (noise, compression artifacts, blur) • Semantic-level LoRA: Enhances semantic details and textures • RAM Model: Extracts text prompts for semantic guidance
Inference
• One-step diffusion with adjustable guidance (λpix and λsem control enhancement strengths) • Tile diffusion and tile VAE support for memory-efficient processing on GPUs with limited VRAM
Training
• Dataset: Ground-truth images + high-quality selected subset • Loss: LPIPS + L1 + Classifier-free guidance (CSD) • Optimizer: AdamW (learning rate 5e-5) • Batch size: 4 (distributed across GPUs) • Training parameters: Timestep range 0.02-0.5, null text ratio 0.5
Requirements
• Stable Diffusion 2.1-base model • RAM model for semantic extraction • GPU with CUDA support • Python 3.10+

Compliance & Provenance

Provider
Open-source
Provider type
Specialized
License
Apache 2.0
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.

Model Source