I want to downscale UHD-1 videos to 1080p. According to this and on other websites the Lanczos method should give best result for downscaling.

However my goal is to downscale a video in a manner like it was captured natively in 1080p. Since there is more information available for UHD I fear that the downscaled video could result in better quality than it could have when captured natively by using a camera with 1080p resolution.

  • The "native" look of a high-quality 1080p image is one that preserves as much information as possible within the Nyquist limit. Downscaling UHD-1 videos to 1080p can only approach this ideal, just as any native 1080p imaging chip can only approach this ideal. – Michael Tiemann Oct 28 '16 at 12:12
  • so in short you mean that the quality of a downscaled video will always be worse than "native" 1080p content? – holzkohlengrill Oct 28 '16 at 12:19
  • 1
    No. The quality of a downscaled video will always be an approximation of an ideal, just as a native video will also be an approximation of an ideal. One approximation may be better or worse. Both will be massively, unmistakably altered (as far as the math is concerned) by compressing the video to 4:2:2 or 4:2:0, the typical chroma encoding schemes used when playing video on consumer devices. And both will be even more massively, unmistakably altered by video compression schemes such as H.264 or H.265. – Michael Tiemann Oct 28 '16 at 14:13
  • It also depends on the quality of your UHD camera. Even though they may be higher pixel resolution the amount of compression that they often apply is usually greater than a HD camera, meaning the image has more artefacts and doesn't necessarily resolve that much more detail. – stib Oct 29 '16 at 13:12

Your understanding of sampling seems to be a bit off. When downsampling from a UHD stream you don't have more information than a native 1080p sensor capturing it, you have less. When the 1080p sensor is capturing video, it has near infinite points and photons it can capture as it is looking at a real life source. Each photosite captures exactly what photons hit it (within the margin of error of the photosite anyway).

With downscaling, you have a much, much smaller amount of input. All you have is the samples available and you don't know exactly where within the pixel they came from. If the sample is an exact multiple, then it is still possible to make an exact 1:1 mapping by combining a group of pixels in to the only pixels they impact, however if you do not have an evenly divisible number of pixels, then you can't determine the exact impact of your new set of pixels from your old and you have to apply some type of imperfect translation between them.

This imperfect translation will produce less positionally accurate information, but also can have other advantages. One advantage of downsampling is the reduction of noise. Since noise is random, averaging multiple samples together will generally reduce noise when all other factors are equal. It's ideal for this to be done at an evenly divisible interval to avoid needing to translate data between pixels, but the advantage is still seen with any downscale.

Getting back to your question however, unless you are shooting an even multiple of 1080p or choose to crop your sensor input to an even multiple of 1080p, you will have some type of translation of pixels. If you want the most "natural" look, I would encourage using a resolution and aspect that allows you to directly map pixels.

  • Thanks for this great answer which makes perfectly sense with the comments from Michael. – holzkohlengrill Oct 31 '16 at 7:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.