Description image

How to Export the Sharpest RED Footage (From REDCINE-X)

08.21.12 @ 1:30PM Tags : , ,

Most everyone can agree that 4K footage downscaled to 1080p looks great. Even cameras that don’t shoot 4K are often oversampling in order to output a clean 1080p file. However, when you’re shooting 4K or 5K on a RED camera and you’re outputting a 1080p file in post, it turns out not all downscaling is created equal. While RED has just released a tutorial on how to export footage from their grading software, REDCINE-X, they barely mention the Software Scaling setting, and so I did some testing myself:

First, here’s RED’s exporting tutorial:


The other day DP Timur Civan dropped by my Brooklyn apartment with a RED EPIC, as he’s a vintage glass aficionado and he wanted to test out my 35 year-old Canon K35 lenses. You might recognize this back yard from my Kickstarter video. This time, however, we were just messing around and checking out some of the crazy purple-pink flares you can get on these old lenses thanks to their slightly radioactive thorium coating1. This is handheld, ungraded stuff, and the lenses are often stopped down to the point where we’re dealing with a very deep depth-of-field (no matte box, no ND), but you can still see these unique flares (especially at the end):

<embed src="http://vimeo.com/moogaloop.swf?clip_id=47631657" type="application/x-shockwave-flash" width="400" height="300"></embed>

Brilliant, I know. We’re up for an Oscar, obviously. While outputting and downscaling the 5K files in REDCINE-X, I noticed a Software Scaling sharpness setting, which is undocumented in REDCINE, barely mentioned in the tutorial video, and also rarely mentioned on REDUSER (I posted about it here). As you can see at right there are plenty of downscaling algorithms and it’s hard to know which is best. I did an impromptu comparison between Mitchell (the default) and Lanczos3 and exported it as an animated .gif, blown up to 300%. Here it is, keep an eye on the sandal’s “adidas” lettering:

Lanczos3 is the sharper one, but more sharpness can bring with it more noise and artifacting. Lanczos3, however, is a “sharper” not “sharpest” setting so it may work in many situations. For a company that espouses sharpness so often, I was surprised that RED hadn’t documented these differences at all. However, I got a definitive answer from Local 600 DIT Thomas Wong when he recommended the solution RED would also likely recommend: “use a RED ROCKET.” There are no sharpness options with the Rocket as it is hard-coded, and one must assume that the Rocket’s hardware algorithm is better than many of these software options. Still, I found the number of options interesting, and given most of us will be outputting to 1080p even when shooting in 4K, it’s something worth experimenting with. If you’re a RED shooter and you want the sharpest 1080p output, be sure to try out more than one setting during the export!

Have any other RED shooters experimented with these settings?

[via REDUSER]

  1. According to Wikipedia, “Natural thorium decays very slowly compared to many other radioactive materials, and the alpha radiation emitted cannot penetrate human skin meaning owning and handling small amounts of thorium, such as a gas mantle, is considered safe.” That’s good to know. []

COMMENT POLICY

We’re all here for the same reason: to better ourselves as writers, directors, cinematographers, producers, photographers... whatever our creative pursuit. Criticism is valuable as long as it is constructive, but personal attacks are grounds for deletion; you don't have to agree with us to learn something. We’re all here to help each other, so thank you for adding to the conversation!

Description image 17 COMMENTS

  • Those lenses are beautiful!

  • Hardware doesen’t have to be better then software(and mostly isn’t). ideally you can program whatever algoritm you want in software It’s just a matter of performance and how many pictures you want to process per second. The only thing you can say mostly about software vs hardware is that hardware is mostly made to be realtime while software is dependen’t on the hardware driving it it. One should also note that in this case if one doesen’t know what algoritm is used by the red-rocket that it is a kind of a bad blanket statement to just say use it because it is probably the best (it probably isn’t since there is almost always a compromise if you are making a realtime card.

    • I agree to the extent that there’s no reason hardware would have a better output than software, necessarily; algorithms are algorithms no matter how they’re executed. But it wasn’t exactly a blanket statement — I quoted a DIT and then said “one must assume the algorithm is better.” Better could mean faster, cleaner… there are no options when using a ROCKET and you can bet RED went with the algorithm they felt was best. Whereas in software we’re presented with ten options and very little explanation.

      The reason I reached out to Thomas in the first place was I felt his 1080p file out of Resolve looked sharper than some of the same shots straight out of my REDCINE-X (in reference to a different project). His answer was that Resolve was using the ROCKET and that’s why it looked better, whereas I could be using different settings in software that I might not like as much. So in some situations it’s certainly possible that software COULD look better but I’m interested in real-world use. I’m posting this to see if anyone else has any insights into these algorithms or experiences to share. We can argue about blanket statements and semantics, but that doesn’t do anyone any good.

  • The lens still looks really nice!

  • Thanks Moore and Xavier… I like them too!

  • I’m by no means a master of image resampling, but my understanding is that algorithms that appear sharpest typically introduce aliasing or haloing. Lanczos, for instance, causes haloing because the kernel goes negative toward the edges of the sampling window. This is where the perceived sharpness comes from, but it can cause problems with extremely high contrast edges.

    Filters with small windows (or that don’t put enough weight on surrounding pixels) also introduce aliasing; the simplest example is if you imagine quartering an images resolution by throwing out every other pixel (vertically and horizontally). This will produce a perfectly sharp image but will obviously introduce aliasing. Trying to do the same to get factors other than 1/4 is even more unpleasant because you’re not throwing out every other pixel but instead some other proportion which leaves straight lines weirdly kinked.

    Lanczos3 is, I believe, generally regarded as the “best” compromise between softness and artifacts, but in the end the only way to get the best results is to select a sampling algorithm based on what the original signal looks like.

    • I think you can make out some haloing around the adidas symbol in this example, which might be why RED uses Mitchell as the default.

  • I think one of the best reference charts showing the tradeoffs for common algorithms is the last 6 images in http://wiki.blender.org/index.php/Doc:2.6/Manual/Render/Oversampling_%28Antialiasing%29

    Obviously this doesn’t translate perfectly to down-sampling real video because a) the samples are all point samples (so there is no “softness” in the raw data) and b) they are arranged in a special pattern designed to be be down-sampled as opposed to a grid of pixels that may or may not line up at all well with the destination pixels, but it still does a pretty good job of showing how different algorithms are suited to different situations.

    For instance, the perfectly straight lines and the stained glass both look great with the “sharper” filters, but the slightly-slanted lines and the checkerboard pattern only look acceptable with the “softer” ones. If you look at the chart you can get a reasonable intuition for why this is the case; ones like “tent” and “box” produce a lot of aliasing because they abruptly stop looking at samples that are far away, and Catmull-Rom and Mitchell-Netravali produce moire both from aliasing and haloing because they put negative weight on samples as they get far away.

  • I’ve used Lanczos3 sharpening for a number of years now in other software, and the ringing is indeed a problem, especially as you increase the number of taps (Lanczos4 is notably worse). I prefer Spline-based resizing when possible, though sometimes find a softer resize (say a bicubic or gauss) to be preferrable in the temporal sense, because its softening gives the false impression of smoother motion, similar to how longer shutter speeds make the motion smoother. To each their own!

  • john jeffreys on 08.21.12 @ 11:50PM

    I WISH I COULD ACTUALLY AFFORD A RED SO I CAN ACTUALLY EDIT FOOTAGE FROM ONE

  • Recommending the ROCKET is understandable, but its not wise from my experience. Most RED users I know use the ROCKET for Proxy work, not for release stuff.

    The point is that the ROCKET has a completely different deBayer than Graeme Natress cool software deBayer. As a result, the RED ROCKET output shows off significant rectangular artefacts in the dark blacks, which can be easily revealed by pushing the signal with very very strong gain. Compare those results to the software only deBayered image – side by side.

    That noise level will create you about 70% larger files in native resolution when converting to a constant quality codec. Vice versa, if you use a constant bit rate codec (e.g. for distribution) those artefacts will lead to a less efficient compression and thereby reduce the overall image quality.

    Even a BluRay file will be about 10-15% larger due to the noise artefacts, or vise versa have a reduced quality when needing to fix into a given data rate.

    Given that many around aim for e.g. web distribution, data rate is important to preserve. In the age of web distribution, the best quality at a given data rate should be the holy grail.

    Per se, the RED ROCKET has a build-in sharpening in place, which makes all its outputs look sharper than the software deBayer. But one can easily apply sharpening in post, to a custom desired level, via software. Still then we prefer the software path over the RED ROCKET because of the overall image quality.

    I am not saying that the RED ROCKET is bad, but its different. And its damn cool for creating dailies and proxies when desired. Thats what it is made for and where I would recommend it for.

    From our experience you will get similar sharp looking images with less noise level when using the software only path. On a recent PC that will not take longer than with the ROCKET.

  • Lliam Worthington on 08.22.12 @ 11:26AM

    Fantastic thread. Thank you Koo. (And other posters) Sorry that I can’t contribute. It’s just something I’ve been wondering about.

  • I heard on one of the fxphd podcasts that on Red software scaling is better quality than the red rocket. The advantage of the rocekt is purely speed.

  • great submit, very informative. I’m wondering why the other specialists of this sector don’t notice this.
    You should proceed your writing. I’m sure, you have a great readers’ base already!

  • It’s awesome in favor of me to have a website, which is beneficial in support of my know-how.
    thanks admin

Description image2 pingbacks

LEAVE A COMMENT