Basically I need to take a h.264 8 bit 1080p render and upscale it to 4k without re-encode. Is it possible? And if it is, does the container matter? Cause ideally it needs to be mp4 container, otherwise avi, otherwise mkv.

Reason: I made a video project in a 1080p space. But now I need the final output as a 4k. So the problem is if I change the resolution of the space it offsets and shrinks all the effects in the video editor. So I could render everything as a high format like png 16 bit (or something better...?) but that's alot of rendering and data for a 30 minute project. The other problem is the footage in the video editor is h.264 already, so I need to vehemently avoid double renders.

  • 2
    Not possible. If you're changing the resolution, you have to encode. – Gyan Nov 24 '18 at 5:02
  • 1
    Just play it back on a bigger television. Done. – stib Nov 27 '18 at 3:50
  • 1
    Who are all these people who are going to die if their video gets re-encoded? It's not like losing a generation on a VHS machine, you just won't notice. – stib Nov 27 '18 at 3:52

To get from 1080p to 4K you're going to need to re-encode, but container won't matter (much), so 16bit png is overkill just for the sake of 4K. H.264 works fine.

But before you spend the time to upsample, you should double-check to make sure that you actually need to. Most modern 4K televisions upsample 1080p on the fly. Most mobile devices aren't high enough resolution to display 4K. If you can leave your project in 1080p and let the display device do the work, then you're already done!

But if you're sure that's what you need, remember that normal, traditional methods of upsampling such as bicubic or bilinear interpolation don't really increase the actual resolution of the image. Sure, the output pixel dimensions will change from 1920x1080 to 3840x2160, but the new image size just comes from (to over-simplify) doubling each pixel's width and height. In other words, the basic scaling algorithms don't add any information to the image, they just make the 1080p "bigger."

There are newer, fancier upsampling algorithms which use sophisticated methods to scale images. Photoshop calles theirs "Preserve details 2.0." DaVinci Resolve calls it "Super Scale." After effects calls it a "Detail Preserving Upscale" effect, but they're all very similar methods of upscaling an image while preserving sharpness. Bear in mind that none of these methods actually add information to the image, either. They're just a tad better than ordinary, re-scaled 1080p. They're also fairly computationally expensive.

In a perfect world, the best thing to do is to shoot in a resolution that's at least as high as your deliverable, but preferably higher, and then downsample. When that doesn't happen, you'll have to evaluate the importance of having a higher resolution and weigh that against the extra time and effort it takes to generate such an image.

4K is still pretty difficult to work with under many circumstances, in terms of bandwidth requirements, latency, storage, etc. There's a lot of overhead that goes into those extra pixels, and it's not exactly a 1:1 relationship. Quadrupling the number of a pixels means you can expect MORE than four times the effort to complete a project, especially if you're working with long timelines, multiple streams, multicam, effects, etc.

| improve this answer | |
  • 1
    A.I. is going to change this techxplore.com/news/2017-10-small-pixel-perfect-large.html – stib Nov 27 '18 at 1:55
  • Everything old is new again. ams.org/journals/notices/199606/barnsley.pdf . Notice barnsley's AOL email address. I know A.I. will improve these algorithms, but at some point you can't create something from nothing. – Jason Conrad Nov 27 '18 at 6:09
  • That's the difference between AI approaches and signal processing approaches. AI will indeed create extra content. The weirdly distorted puppies in the images on this page are examples of created content, rather than interpolated content. fastcompany.com/90244767/… – stib Nov 28 '18 at 2:52
  • Right, but once you start enlarging pixels, it doesn't matter if you fill in the missing bits with similarly colored pixels, fractals, wavelets, or Google's creepy fever dream of eyeballs and spiders, no real visual information from the original scene is recovered. You're right that they'll eventually be able to to present a convincing illusion of continuity, but as it stands, they use 128-512 TPU cores to create a single, completely generative image. Upscaling a user-specified original is a different problem, and one that's multiplied by a factor of 24 at least for video. – Jason Conrad Nov 28 '18 at 13:07
  • True, you'll never restore information that isn't there ("computer - zoom in and enhance"), but an illusion is going to be good enough in most non-forensic cases. And yeah, the processing power involved is stupendous. To train Google's BigGan model took enough electricity to power an average American house for 6 months. – stib Nov 29 '18 at 4:14

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.