Adobe's VideoGigaGAN upscales blurry video 8x without the flicker

Upscaling a pixelated video used to mean choosing between two bad options: smear everything into a blurry smudge, or let an AI guess the details and watch the whole frame flicker like a strobe light.

Adobe researchers think they've cracked the code. In a new paper released on April 18 and reported on April 22, the company introduced VideoGigaGAN, a generative AI model that can boost blurry footage up to eight times its original resolution without the visual chaos.

Sharp details without the AI jitter

Generative Adversarial Networks (GANs) have long been the gold standard for blowing up static images, largely because they can hallucinate realistic micro-details out of thin air.

Bring that tech to moving pictures, though, and it usually falls apart. When frame A guesses one set of wrinkles and frame B guesses another, the result is an unwatchable, jittery mess. Traditional video super-resolution (VSR) tools avoid this by playing it safe, but their outputs tend to look soft and muddy.

VideoGigaGAN essentially bridges that gap. It borrows the intense detail generation of image GANs while enforcing temporal consistency across frames.

In Adobe's demos, complex elements like human skin textures and tiny facial creases look remarkably crisp. The catch? Most of those fine details are entirely artificial, generated by the model. Yet they look so natural that it pushes the old "what is a photo" debate right into video territory.

Still locked in the research lab

Before you look for an update button, keep in mind this is strictly a research preview. Adobe hasn't made any promises about bringing VideoGigaGAN to Premiere Pro or the wider Creative Cloud suite anytime soon.

It is also not Adobe's only bet in the space. Back at its MAX event in October 2023, the company teased Project Res-Up, a diffusion-based experiment aimed at cleaning up low-res GIFs and video clips. Meanwhile, heavyweights like Microsoft and Nvidia have already shipped their own VSR solutions.

Why it matters

For editors, archivists, and documentary filmmakers sitting on decades of standard-definition tape, an 8x upscale that actually holds together frame-by-frame is the holy grail. It would turn unwatchable archival footage into broadcast-ready assets without manual touch-ups.

Technically, it's also a refreshing validation for GANs. While the entire industry has pivoted toward compute-heavy diffusion models, Adobe's researchers proved that GANs still have plenty of fight left if you can tame their temporal instability.

That said, we think this capability opens up a tricky ethical can of worms. When an algorithm literally invents realistic pores, blemishes, and facial creases that never existed in the original recording, can we still call it restoration? As tools like this mature, creative suites will eventually need clear provenance standards so viewers know where reality ends and AI hallucinations begin.