GAN Upscale

Breathing New Life into Old Footage

For decades, increasing the resolution of video (upscaling) was a simple mathematical process. Traditional methods like bicubic or nearest-neighbor interpolation essentially make educated guesses to fill in the missing pixels when stretching a low-resolution image to a larger size. The result is often acceptable but typically appears soft, blurry, and lacking in fine detail. A GAN upscale represents a paradigm shift from interpolation to generation. Instead of just stretching pixels, it uses a sophisticated AI model to intelligently ‘hallucinate’ or generate new, believable detail that was not present in the source footage.

This technology has become a game-changer for film restoration and preservation, allowing older content shot on standard-definition video or lower-resolution film stocks to be presented on modern 4K and 8K displays in a way that was previously impossible. It offers a path to revitalizing vast back-catalogs of television shows, documentaries, and films for a new generation of viewers.

How Generative Adversarial Networks Work

The technology’s name comes from its core architecture: two dueling neural networks. The process can be understood as a competition between an artist and a critic:

1. The Generator: This network acts as the artist. It takes the original low-resolution image and attempts to create a high-resolution version. Its initial attempts are poor, like a forger learning their craft.
2. The Discriminator: This network acts as the art critic. It has been trained on a massive dataset of thousands of real, high-resolution images and has learned to distinguish between authentic detail and fake, generated detail.

During the training process, the Generator creates an upscaled image and shows it to the Discriminator. The Discriminator judges whether the image is ‘real’ or ‘fake’ and provides feedback. They compete against each other in a feedback loop millions of times. The Generator constantly refines its technique to better fool the Discriminator, and in doing so, it learns to generate incredibly realistic textures, edges, and details—such as the fabric of a coat, the stubble on a face, or the grain of wood—that make the upscaled image appear authentically high-resolution.

The Authenticity Debate in Restoration

While GAN upscaling can produce stunning results, it is also a source of intense debate among film purists and archivists. The central question is one of authenticity. Is the detail created by the AI a faithful restoration of what was originally there, or is it an artificial fabrication—a ‘best guess’ based on the model’s training data? For example, if the AI was trained on many images of human faces, it might add skin pores to a blurry face in a way that looks realistic but may not precisely match the actor’s actual skin texture.

This can sometimes lead to an undesirable ‘plastic’ or waxy look, especially if the process is pushed too far. It raises a philosophical question about the goal of film preservation. Is the objective to preserve the source material exactly as it was captured, limitations and all? Or is it to adapt the material to meet the expectations of a modern audience on modern hardware? Director Peter Jackson’s team used highly advanced AI techniques to restore and colorize World War I footage for They Shall Not Grow Old, a process praised for its immersive power but also questioned for its interpretive, rather than purely archival, approach. GAN upscaling exists on this same complex spectrum between faithful preservation and creative reinterpretation.