How much can you uprez or interpolate an image before it falls apart? It’s a common question by anyone who has to worry about the following workflows:

  • SD to HD
  • 720p to 1080p
  • 1080p to 2K
  • 2.5 or 3K to 4K, etc.

Can there be an absolute answer to this question (by ‘absolute’, I mean an answer that is true regardless of the scenario)?

No. Why not?

Well, people uprez images all the time. If you watch DVDs or SD TV on an HDTV or high-resolution computer display, you are engaging in image interpolation. Millions of viewers watch television in this fashion all over the world, every passing hour. No one complains.

This is why, in absolute terms, it doesn’t matter if you uprez. People only complain under the following two conditions:

  • When heavy compression artefacting is evident, or if features like text are illegible.
  • When presented with the ‘original’ for comparison. Then, the differences are obvious.

High resolution, then, becomes an acquired taste. We are happy with our car, until we see a newer shinier car. Today, 4K is that car. I’ve already written about what 4K TV is, and how you go about deciding whether you need one or not. This article will purely focus on the uprezzing or interpolation part. It’s for the image makers, the ones whose responsibility it is to present the best image. If videographers worldwide spend sleepless nights worrying about interpolation, it’s because they are concerned about their images.

Conditions for testing

Before we can test interpolation we must first stake the point at which the image will no longer be acceptable. What could these factors be?

  • At the ideal viewing distance there is a considerable deterioration in image quality – edges are jagged, gradients look posterized.
  • Resolution is considerably reduced. E.g., two lines of size 1px separated by 1px can no longer be distinguished apart.
  • Compression should not be a criterion, so we must compare uncompressed raster images.
  • The image must be formed at different sizes without rasterization entering the equation. The solution is to use vector images to form the raster images. Unfortunately, no camera produces images at different resolutions sufficient for testing.
  • There must be zero chroma subsampling.
  • There must be no involvement of factors that affect resolution – like lens, optical low-pass and other in-camera filters, sensor, formats, motion blur, optical and temporal artefacting, etc.
  • The image must have straight and diagonal lines, curves and shapes, text, solid color and gradients.

Since the ultimate criterion is subjective image quality, I leave it to you to decide at what point these factors kick in.

The test

Based on the above, I have chosen this methodology:

  • Use Adobe Illustrator to produce vector images. These images can be scaled to any proportion without rasterization effects coming into play. In other words, these are the sharpest images possible known to man.
  • Output these vectors to raster images, namely TIFF. No compression is applied. All images are 8-bit, to reflect video.
  • Maximum horizontal resolution (of the primary image) is 900 px x 450 px. Not everyone has 4K or 2K monitors, and many might view this page on mobile devices. The principles remain the same, no matter what. Whatever you are seeing is already in 100%, provided your browser is not zoomed in or out, or if you have not altered the resolution settings of your display or graphics card in any way.
  • Color space is sRGB throughout.
  • Uprez algorithm used is Photoshop bicubic smoother.

Images are output in the following resolutions:

  • 900 px – No uprez or 1x
  • 855 px – 1.05x
  • 809 px – 1.11x
  • 675px – 1.33x
  • 600px – 1.5x
  • 542px – 1.66x
  • 450px – 2x
  • 300px – 3x

Here are the original images, output from Illustrator, in TIFF:

Warning! These images are JPEGs at 650px because browsers don’t display TIFF files directly and this website has a width of 650px. Click on them to view the original TIFF files and compare those. Do not compare these images as shown!

Original Reference 900pxWeb

Reference 855px

Web

Reference 809px

Web

Reference 675px

Web

Reference 600px
Web

 

Reference 542px

Web

Reference 450pxWeb

Reference 300pxWeb

As you can see, all images look similar at their native resolutions. This is only possible with vector images. Cameras don’t produce images in such increments so we can test them so.

The results

Here are the results, of the above images resampled (uprezzed) to 900px:

Warning! These images are JPEGs at 650px because browsers don’t display TIFF files directly and this website has a width of 650px. Click on them to view the original TIFF files and compare those. Do not compare these images as shown!

300px uprezzed to 900px

Web

450px uprezzed to 900px

Web

542px uprezzed to 900px

Web

600px uprezzed to 900px

Web

675px uprezzed to 900px

Web

809px uprezzed to 900px

Web

855px uprezzed to 900px

Web

What do you see? At which point is the image acceptable, if at all? I’ve found the fairest way to judge is to sit at the ideal distance from your monitor and view the images at 100% without pixel peeping.

Factors that affect resolution either way

Comparing vector images gives us absolute theoretical results. In the real world, video suffers from many issues that reduce its resolution:

Because of the compounding effects of the above, a 4K image is not really a 4K image. This means, you can ‘get away with higher levels of uprezzing.

Also, let’s not forget one another elephant in the room – and that is the interpolation algorithm used. There are many out there, and depending on the software and methodology used, you might achieve different results. I have not applied any sharpening to the above images, but that affects results, too.

My observations

What do I see? I find the images look acceptable at an uprez ratio of 1.5x. I have also observed this while uprezzing 720p to 1080p, which I did for my feature film.

Take note, though, you can tell the difference at even 1.33x or less, especially with the text part. I must put this down to the font and the uprez algorithm used. If the viewing distance is sufficiently far away, an uprez rate of 2x is also acceptable – but not to me. That’s what it all comes down to. We as image makers must take responsibility for our images.

Here are some tidbits about 1.5x, the magic number:

  • 1080p/720p = 1.5
  • 2880 x 1620 / 1920 x 1080 = 1.5 (The former is the resolution of the Arri Alexa)
  • UltraHD 3840 x 2160 / 2432 x 1366 = 1.5 (The latter is the resolution of the BMCC)
  • 4096 / 2880 = 1.4 (The former is 4K, the latter is the Arri Alexa). This is why it was acceptable to ‘blow up’ Skyfall for theatrical distribution in 4K.

What do you think? What’s your magic uprez number?

7 replies on “How Much Can You Uprez, or, What is the Limit of Image Interpolation?”

  1. Your test material already has jaggies clearly visible. Unfortunately, you exported your original test material from Illustrator WITHOUT using the anti-aliasing function. You should either select art-optimized or type-optimized when saving something like this out, otherwise you’re starting with a resolution-deficient image, meaning EVERYTHING downstream from that will look crappy, as you have demonstrated here.

    Try this again, but using anti-aliasing, and you will come to completely different conclusions. For instance, you probably CAN uprez by 200% or more if your artwork doesn’t contain too many diagonal lines.

    Also try a photograph or rastor-based image within Photoshop. You’ll find some photos can go up 250-300% and still be presentable, depending on how you use them.

      1. Okay, but you said to use Illustrator in your instructions. And this discussion about rastor imagery for video is useless if you’re starting out with bad rastor imagery to begin with. That is merely what I was pointing out, how to get the best rastor imagery from your Illustrator files.

  2. kaiz3rwork  Nothing can take it up to 2 and be acceptable. A doubling of the resolution is visually perceptible, and the eye cannot be fooled. But like I said, billions of people watch 2x TV without knowing or caring!

  3. Interesting !
    What about tools or plugins for uprez ? Like magic bullet etc… Does it can throw the 1.5 to 2 ?

Comments are closed.