The limits of 4K upscaling: why AI cannot fully recover lost video detail
New analysis suggests that while artificial intelligence is sharpening the process of video upscaling, it remains a predictive exercise rather than a true restoration of original data.
As 4K becomes the dominant standard for video creation, a significant portion of existing content remains in 1080p or standard definition. Engadget has published a technical explainer detailing the mechanics of upscaling these lower resolutions to 4K, highlighting that the process is fundamentally about estimating missing pixel data rather than recovering original clarity.
The distinction between resolution standards is precise. Standard ultra-high definition, commonly referred to as 4K, operates at 3,840 x 2,160 pixels, while cinematic "true 4K" or DCI 4K sits at 4,096 x 2,160 pixels. Moving from 1080p to standard 4K requires a four-fold increase in total pixel count, meaning three additional pixels must be generated for every original pixel recorded.
Traditional upscaling software relies on interpolation, a mathematical process that estimates the colour and brightness of surrounding pixels to fill in the gaps. While this method is predictable, it often results in softer images that lack genuine texture. Sharpening and noise reduction can improve the appearance, but excessive application can make the footage look unnatural.
AI-based upscalers offer a different approach by training on pairs of low and high-resolution images to predict how sharper features should appear. This method can produce more convincing results than traditional interpolation, though it is not without its own set of challenges. AI models can introduce unusual visual artifacts that traditional methods typically do not, and the quality of current results may appear dated as models continue to improve.
Source quality remains the critical determinant of success. The $12 million HD restoration of Star Trek: The Next Generation by CBS serves as a benchmark for high-end conversion. By using original 35mm film negatives, which contained far more detail than the original broadcast, the studio was able to remaster the series to 1080p. In contrast, upscaling from a low-bitrate digital file cannot remove inherent compression artifacts such as blocking or banding.
For home users and professionals alike, the leap from standard definition to 4K is particularly difficult. Older video formats often used interlacing, which requires de-interlacing before upscaling. Engadget notes that maintaining watchable quality when upscaling 480i NTSC video to 4K is questionable, suggesting that while AI is making the process more accessible, a like-for-like match to the original footage is rarely guaranteed.
