Image upscaling has been transformed by AI. Where traditional algorithms like bicubic interpolation simply smooth out pixels, AI-based upscalers can reconstruct plausible detail that was never in the original image. Upscayl is one of the most accessible tools in this category: it is free, open-source, runs entirely offline, and produces results that compete with paid alternatives.
I have been using Upscayl for various projects over the past year, from restoring old family photographs to preparing low-resolution game screenshots for print. This guide covers what the tool does well, where it falls short, and how to get the best results from different AI models.
What Makes Upscayl Different
Upscayl is a desktop application built on Electron that uses NCNN (a neural network inference framework) to run AI upscaling models locally on your GPU. Unlike cloud-based services, your images never leave your computer. There are no watermarks, no usage limits, and no subscription fees.
The application ships with several pre-trained models, each optimized for different types of content. You can also import custom NCNN models, which gives advanced users access to the broader ecosystem of AI upscaling research.
Installation and First Use
Upscayl is available for Windows, macOS, and Linux. The Windows installer is a standard setup wizard that takes about two minutes to complete. The application requires a GPU with Vulkan support for hardware acceleration, though it can fall back to CPU processing for systems without compatible GPUs.
Supported GPU Configurations
| GPU Brand | Minimum | Recommended | Processing Speed |
|---|---|---|---|
| NVIDIA | GTX 1050 | RTX 3060+ | Fast (CUDA + Vulkan) |
| AMD | RX 570 | RX 6700 XT+ | Good (Vulkan) |
| Intel | Arc A380 | Arc A770 | Moderate (Vulkan) |
| CPU Only | Any modern CPU | 8+ cores | Slow |
Available AI Models
Upscayl includes several models, each with distinct characteristics. Choosing the right model for your content type makes a significant difference in output quality.
Model Comparison
| Model | Best For | Scale Factor | Speed |
|---|---|---|---|
| Real-ESRGAN (General) | Photographs, general images | 2x / 4x | Medium |
| Real-ESRGAN (Anime) | Anime, illustrations, flat art | 2x / 4x | Medium |
| Remacri | Detailed photographs, textures | 4x | Slower |
| UltraSharp | High-detail preservation | 4x | Slower |
| Digital Art | Digital paintings, game art | 4x | Medium |
For most photographs, the Real-ESRGAN General model produces the most natural results. The Remacri model tends to preserve fine detail better but can sometimes over-sharpen smooth gradients. UltraSharp lives up to its name for images where texture detail is the priority.
Quality Benchmarks
I tested each model against a set of 20 images covering different content types: portraits, landscapes, text documents, game screenshots, and illustrations. Quality was measured using SSIM (Structural Similarity Index) against high-resolution originals that were downscaled to create the test inputs.
SSIM Scores (Higher is Better)
| Content Type | Bicubic | Real-ESRGAN | Remacri | UltraSharp |
|---|---|---|---|---|
| Portraits | 0.82 | 0.91 | 0.90 | 0.89 |
| Landscapes | 0.79 | 0.88 | 0.90 | 0.89 |
| Text | 0.85 | 0.87 | 0.86 | 0.88 |
| Game Screenshots | 0.80 | 0.89 | 0.91 | 0.90 |
| Illustrations | 0.83 | 0.92 | 0.88 | 0.87 |
The AI models consistently outperform traditional bicubic scaling by a meaningful margin. The differences between AI models are smaller but still visible in direct comparison, particularly in areas with fine texture detail.
Practical Use Cases
- Photo restoration: Old scanned photographs at 72 DPI can be upscaled to print-quality resolution while recovering detail that scanning artifacts obscured
- Game screenshots: Upscaling 1080p screenshots to 4K for wallpapers or social media posts
- Web graphics: Improving low-resolution logos or product images when original files are unavailable
- Batch processing: Upscayl supports batch mode for processing entire folders of images with consistent settings
Limitations to Be Aware Of
AI upscaling is not magic. There are several situations where results may be disappointing:
- Heavily compressed JPEG images with visible block artifacts — the AI may amplify compression artifacts rather than remove them
- Images with text — AI models sometimes distort letterforms, making text less readable at higher scales
- Very low resolution sources (below 100x100 pixels) — there simply is not enough information for the AI to work with
- Processing time for large batches can be significant, even on modern GPUs
For best results, start with the highest quality source image available. AI upscaling works by inferring detail from existing patterns. The more information the model has to work with, the better the output.
Tips for Best Results
- Remove noise and compression artifacts before upscaling when possible
- Match the model to your content type — anime models on photographs produce unnatural smoothing
- Use 2x upscaling twice rather than 4x once for large scale factors — this often produces better detail
- Save output as PNG to avoid introducing new compression artifacts
- Compare models on a representative sample before batch processing an entire collection