For Decades, Enlarging an Image Always Meant Losing Quality. Here Is How AI Changed That in 2014.

Thirty-Three Years of Getting Better at Guessing

The progression of image upscaling algorithms before neural networks tells a story of incremental improvement within a fixed mathematical framework. The simplest method, nearest-neighbor interpolation, assigned each new pixel the value of the closest existing pixel. The result was blocky and pixelated, like enlarging a low-resolution screenshot. It was fast and visibly wrong.

Bilinear interpolation replaced nearest-neighbor as the default in most applications. Instead of picking the single closest pixel, it averaged the values of the four surrounding pixels weighted by distance. The blocky artifacts disappeared. The image looked smooth. Smooth, unfortunately, meant blurry, and the blurriness became more obvious with each doubling of the image size.

Bicubic interpolation, proposed by Robert G. Keys in a 1981 paper, extended this idea to consider 16 surrounding pixels in a 4x4 grid rather than a 2x2 grid. It used cubic polynomial curves to calculate weights, producing transitions that were smoother and more realistic than bilinear in most cases. Bicubic became the standard method in professional tools and remained there for over three decades. Photoshop shipped it as the default. Camera RAW processing used it. When a printer needed to upscale an image for output, bicubic was the calculation.

The limitation was structural. All three methods, from nearest-neighbor to bicubic, were working locally. They looked at the immediate neighborhood of each pixel and produced a mathematically reasonable estimate of what the missing pixel should be. They had no information about what the object in the image was. A pixel in a bicubic upscale of a portrait has no idea it is part of an eyelash. It just knows the numerical values of its 16 nearest neighbors. This is why upscaled faces looked slightly artificial even with bicubic: the algorithm treated a hair the same way it treated a sky gradient, because it had no way to distinguish them.

The 2014 Paper That Changed the Framework

The SRCNN paper was published by Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang at the Chinese University of Hong Kong and presented at the European Conference on Computer Vision in 2014. The title was "Image Super-Resolution Using Deep Convolutional Networks." Its core proposition was that the relationship between low-resolution and high-resolution image patches could be learned from data rather than derived from mathematical first principles.

The architecture was three-layer by modern standards. The first layer extracted patches from the low-resolution input and mapped them into a high-dimensional feature representation. The second layer performed nonlinear mapping of those feature vectors into the domain of high-resolution patches. The third layer aggregated the predicted patches back into a coherent high-resolution image. The network was trained on pairs of high-resolution images and artificially degraded versions of the same images. After training on enough pairs, the network learned which patterns in low-resolution input corresponded to which textures, edges, and structures in high-resolution output.

The measured results were modest in PSNR terms, the standard mathematical metric for image quality. SRCNN improved over bicubic by 0.15 to 0.17 decibels on standard test datasets at 3x upscaling. But the visual quality difference was immediately apparent to human observers in a way that PSNR did not fully capture: edges were sharper, fine textures were partially recovered, and the characteristic blurred halo around high-contrast boundaries was substantially reduced. The network had not guessed from a local neighborhood. It had recognized that certain low-resolution patterns tend to correspond to specific high-resolution structures, because it had seen that relationship demonstrated tens of thousands of times during training.

From SRCNN to Modern Upscalers

SRCNN opened a line of research that moved quickly. In 2018, a team published ESRGAN, the Enhanced Super-Resolution Generative Adversarial Network, which introduced a fundamentally different training objective. Where SRCNN optimized for pixel-level accuracy, ESRGAN trained a generator network against a discriminator that learned to distinguish real high-resolution images from generated ones. This adversarial training pushed the generator toward outputs that looked perceptually realistic rather than mathematically optimal, which proved to produce results that human evaluators rated higher even when the pixel-level metrics were similar. ESRGAN won the PIRM Challenge on perceptual super-resolution that year.

Real-ESRGAN, published in 2021 by Xintao Wang and colleagues at Tencent ARC Lab, addressed a practical limitation of earlier networks: they were trained on clean, artificially degraded images and often failed on real-world photos that had been compressed multiple times, printed, rescanned, or damaged in other ways. Real-ESRGAN trained on images that had been subjected to simulated multi-stage degradation: randomized blur, noise, downsampling, and JPEG compression applied in various sequences. The result was a model that handled the complex, layered damage present in real photographs far better than models trained on clean degradation pipelines.

This progression, from SRCNN to ESRGAN to Real-ESRGAN and its contemporaries, is what powers most consumer upscaling tools today. The tools that describe themselves as AI-based image enlargers are generally drawing on architectures in this lineage.

Conclusion

The important limitation of all learning-based upscaling is that the network adds plausible detail, not original detail. When a 400x300 image is enlarged to 1600x1200, the additional pixels did not exist in the original file. The network predicted them based on patterns it learned during training. Those predictions are statistically likely to be correct in the sense that they look like what a human observer would expect to see, but they are not a recovery of original information.

For most practical uses, this distinction does not matter. Enlarging a product photo for a larger display, rescaling a headshot for a presentation, or improving a small image for printing all benefit from AI upscaling because the result is visually better than bicubic while the exact pixel values do not carry evidential weight.

The situation where it matters is forensic or technical contexts: any application where pixel data needs to be accurate rather than plausible. Upscaling a surveillance image for identification analysis, processing medical imaging, or enlarging satellite photography for measurement are all cases where learned prediction would compromise the integrity of the result. These applications require documented interpolation methods with known, bounded error characteristics rather than network predictions that cannot be formally bounded.

For practical photo and graphic work where you need a larger version of an image at better quality than a simple resize produces, an AI image enlarger applies the kind of network-based upscaling that has replaced bicubic interpolation as the standard since 2014. The bicubic algorithm still ships in every major tool, because it is fast and deterministic, but for quality-critical enlargement, the 33-year reign of guessing from local pixel neighborhoods is over.

Frequently Asked Questions

Try These Free Tools