You have a 400×300 pixel image and you need it at 1600×1200. Traditional upscaling — bilinear or bicubic interpolation — just blurs the image, smearing existing pixels across more space. AI upscaling takes a fundamentally different approach: it generates new detail that wasn't in the original image, producing results that look dramatically sharper than anything interpolation can achieve. Here's how it works, what the limitations are, and when it makes sense to use it.
How traditional upscaling works (and fails)
When you increase the pixel dimensions of an image using traditional methods, the software has to invent new pixels. A 400px image scaled to 1600px needs 16 times more pixels than the original contains. The question is: what color should each new pixel be?
- Nearest neighbor — copies the closest existing pixel. Fast, but produces a blocky, pixelated result. The 400px image looks like a 400px image displayed at 4x zoom.
- Bilinear interpolation — averages the 4 nearest pixels. Smoother than nearest neighbor, but everything looks soft and out of focus. Fine edges become blurry gradients.
- Bicubic interpolation— considers a 4×4 grid of neighboring pixels and uses polynomial fitting. Slightly sharper than bilinear, but still fundamentally limited: it can only smooth between existing data points. It can't recreate detail that doesn't exist.
The core problem: interpolation can only redistribute existing information. It can't add new information. If the original image doesn't contain the texture of individual hair strands, no amount of interpolation will put them there. The result is always softer than a native high-resolution image.
How AI upscaling works
AI upscaling (also called super-resolution) uses neural networks trained on millions of image pairs: a high-resolution original and its downscaled version. The model learns patterns — what a sharp edge looks like when downscaled, what fine texture corresponds to a blurry area, how detail distributes across natural images. When given a new low-resolution image, the model generates plausible high-frequency detail based on these learned patterns.
The key word is plausible. The AI doesn't recover the original detail — that information is genuinely lost. It invents new detail that is statistically consistent with what the original might have looked like. The texture it adds to an upscaled face isn't the person's actual skin texture; it's a realistic-looking texture that fits the overall lighting, color, and structure of the face.
Real-ESRGAN
Real-ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) is the most widely deployed AI upscaling model. Developed by researchers at Tencent and published in 2021, it builds on the ESRGAN architecture with training specifically designed to handle real-world degradation — not just clean downscaling, but JPEG compression artifacts, noise, and blur.
The model uses a generator network (which produces the upscaled image) and a discriminator network (which judges whether the output looks realistic). During training, these two networks compete: the generator tries to produce images that fool the discriminator, and the discriminator tries to distinguish generated images from real ones. This adversarial training pushes the generator toward producing images with realistic textures and sharp details.
Real-ESRGAN supports 2x, 3x, and 4x upscaling. The results at 2x are typically excellent — almost indistinguishable from a native high-resolution photo under normal viewing conditions. At 4x, the quality is good but you can start to see hallucinated textures if you examine closely. 4x on a very small source image (under 200px) pushes the model hard and results become inconsistent.
What AI upscaling does well
Photographs of people and faces
AI upscalers are heavily trained on facial data. Skin texture, eyes, hair — these are reconstructed convincingly at 2-4x. A 200px headshot from a web page can be upscaled to a usable 800px portrait with realistic detail.
Natural textures
Grass, trees, fabric, water, stone — natural textures have statistical regularity that the model exploits. A blurry patch of grass in a landscape photo will get plausible grass-like detail added.
Architecture and hard edges
Buildings, furniture, vehicles — subjects with clean geometric edges upscale well. The model sharpens edges convincingly and fills in surface detail.
Old or low-quality photos
AI upscaling genuinely shines on old family photos, scanned prints, and screenshots from old websites. These images are often small and degraded by compression, scanning artifacts, or age. AI upscaling can produce a dramatically improved version — not a restoration of the original, but a plausible higher-quality version.
What AI upscaling does poorly
Text and fine lettering
Text in an image that's too small to read at the original resolution will not become readable after AI upscaling. The model doesn't understand language — it generates textures that look letter-like but don't form correct characters. If you need to upscale a document or screenshot where text legibility matters, the text must already be readable in the source. AI can sharpen it, but it can't reconstruct characters from a blur.
Heavily compressed images
JPEG block artifacts at very low quality (below 30%) are sometimes interpreted by the model as legitimate image features. The upscaled result can show exaggerated block patterns or hallucinated textures where the model tried to “enhance” a compression artifact. Real-ESRGAN handles moderate compression well, but extreme cases are problematic.
Extreme upscaling (8x+)
Running 4x upscaling twice to get 16x doesn't work well. Each pass generates new detail, and the second pass generates detail on top of already-generated detail. The result looks increasingly synthetic — oversharpened textures, artificial-looking skin, and pattern repetition. If you need more than 4x, accept that quality will be limited.
Pixel art
AI upscalers trained on photographs will destroy pixel art. They interpret the deliberate blocky pixels as low-resolution detail and try to smooth them into photorealistic textures. Use nearest-neighbor scaling for pixel art — it preserves the sharp pixel boundaries that are the whole point.
Realistic expectations
AI upscaling is impressive, but it has a ceiling. The technology cannot perform miracles. Some guidelines for setting expectations:
- 2x upscaling — results are typically excellent. A 600px image upscaled to 1200px will look natural and sharp.
- 3x upscaling — results are good. Noticeable generated detail if you compare closely with a native 3x image, but very usable for most purposes.
- 4x upscaling — results are acceptable. Fine for web use and moderate print sizes. Close inspection reveals AI-generated textures.
- Beyond 4x — diminishing returns. Run multiple passes if you must, but expect visibly synthetic results.
The quality also depends heavily on the source image. A well-lit, in-focus, moderately compressed photograph will upscale beautifully. A dark, noisy, heavily compressed thumbnail will produce a better result than interpolation but won't look like a professional photo.
When to use AI upscaling
- Printing photos that are slightly too small. Your 2400px photo needs to be 3600px for a poster at 300 DPI. AI upscaling at 1.5x (round up to 2x and crop) is a clean solution.
- Rescuing old or low-resolution images. Family photos from early digital cameras, scanned prints, web-sourced images.
- Product images from vendors. You receive a 500px product photo and need a 1500px version for your e-commerce site.
- Social media content. Repurposing older content that was created at lower resolutions.
When NOT to use AI upscaling
- When you can reshoot. A native high-resolution photo will always look better than an upscaled version. If you have the option to capture the image again at higher resolution, do that instead.
- Legal or medical images.AI upscaling generates detail that isn't real. For images where accuracy matters (evidence photos, medical scans, legal documents), fabricated detail is actively harmful.
- Images that will be closely inspected. If the viewer will be examining the image at 100% zoom (fine art prints, technical documentation), AI artifacts may be visible.
Try it yourself
MakeMyImgs' AI upscale tool uses Real-ESRGAN on the server side. Upload your image, select 2x, 3x, or 4x, and download the upscaled result. It handles photographs, illustrations, and mixed content. Processing takes a few seconds depending on the image size and chosen scale factor.
The bottom line
AI upscaling is the single biggest improvement in image resizing since digital photography began. It produces results that traditional interpolation simply cannot match. But it's generating plausible detail, not recovering lost information. Use it for making images look better at larger sizes, not for extracting hidden detail. Keep expectations realistic, stick to 2-4x scaling, and start with the best source image you have.