YouTube video experts don’t understand why flat/log footage on 8-bit cameras is a bad idea

UPDATE: There is now a short video that scratches the surface of what’s explained here if you don’t feel like reading everything below. If you’re interested, give it a watch: “No, This Doesn’t Look Filmic – Shooting log, flat, and LUTs all suck”

Part of being a professional in any field is understanding where mistakes are easy to make and what little details make a big difference in the end result. If you’ve skimmed this blog, you already know that I’m a big obnoxious noise-maker when it comes to shooting down the bad knowledge of “shooting flat” or “shooting log” on most cameras. The reason is simple: basic math and basic understanding of how color is stored says it’s a bad idea. Unfortunately, there is a huge body of YouTube video work out there filled with bad advice. These tend to use phrases like “the film look” and “more dynamic range” and “cinematic shooting” and “make your video look filmic.” The vast majority of these advice videos are poor, and because they’re doing a video to show off how great of an idea it is, the demonstrations often look poor too.

I ran into such a video recently and dropped the usual short nugget about not shooting flat or log on 8-bit cameras and how professionals don’t do that. The person who posted the video responded to my comment and it provided a wonderful opportunity to expound on the subject further. This starts with the video description and each successive comment in the chain. This post’s secondary purpose is to archive what was said. I hope you learn from it, in any case.

The following frame grabs are from the video. Though they will have suffered some degradation from YouTube compression and can’t be viewed as equivalent to what you’d get straight out of the camera, they are DEFINITELY what you can expect “FILMIC PRO LogV2” iPhone footage to look like after it’s been uploaded to YouTube!

The Conversation Begins

Description: “Whoa. FiLMiC Pro LogV2 is here! We’ve been testing it for a few weeks, and…It’s pretty great. Up to 12 stops of dynamic range on the latest iPhone XS and new higher bit rates nearing 140Mbps. It will even breathe new life into your older iPhones (like iPhone SE and 6S). Watch the video to learn more!”

Me: Or just shoot without log nonsense because it doesn’t play nice with 8-bit output files, and get the image right in-camera. You know, like an actual videographer that actually knows how to do video work would do.

Video poster: If you have a log option then you can’t match that in camera. The DR just isn’t there. So shooters that know what they’re doing use it. This log was created to work in 8bit very similar to how Canon Log was designed for 8bit video. They both work great and actually have many of the characteristics of 10bit. For example, the blue sky holds perfectly in grading – no banding that often happens with other 8bit log footage or just 8bit footage in general. You should give it a try.

Me: Dynamic range isn’t that important, especially when we’re talking about a maximum of half a stop obtained by sacrificing color difference information, introducing quantization errors in the corrected footage, and further losing subtle color gradients to macroblock losses in the H.264 AVC compression algorithm. I have run multiple tests on footage shot with a variety of supposedly better picture profiles on 8-bit footage from neutral to “flat” to log and the flat/log footage is always worse: it breaks faster, it has clear loss of find detail even before any grading is done which is even worse after correction, and the dynamic range increase is not enough to noticeably improve the overall look of the footage.

I’m concerned that you said it’s similar to 10-bit footage. There is a fundamental misunderstanding of what 10-bit depth does demonstrated in that statement. It’s a matter of precision. There is no benefit to 10-bit footage that is not being graded or is only going to be minimally graded because the quantization errors introduced across small changes are so small that they result in no visible side effects. When the stored color curves are significantly different from the desired color curves, there are much larger quantization errors introduced in the push back to normal colors and these are responsible for the ugly color issues (especially the plastic-looking skin tones) in corrected flat/log 8-bit footage. 10-bit precision eliminates these quantization errors because they store (relative to 8-bit color depth) a fractional component with 2^-2 precision that causes the curve change calculations to make much better quantization (rounding) decisions even if the “push” constitutes two full stops worth of difference.

You can get away with weird color profiles for such footage due to the added precision, but 8-bit color is already truncated to the same precision as the destination color depth, so further rounding operations during correction work is almost exactly the same as if the 10-bit rounding is simply floor rounded (truncated) also. Multiple operations can compound the problem further. Look up how a number like 1.1 is stored in IEEE 754 floating-point and why the historic FDIV bug in the first-gen Pentium CPUs was such a big deal for a better understanding of the issue of numeric precision problems in computing. The same principles affect log-to-standard correction of 8-bit footage.

I get a response and I’m not impressed

Video poster: Thanks for the detailed analysis! For most DPs dynamic range (including highlight roll off) and color science are the most important things. That’s really the primary reason the ARRI Alexa has been the king of movies & TV. Even before they had a 4K camera. Now here we’re obviously not talking about that kind of performance or quality, but the new LogV2 increases the DR significantly – 2 1/2 stops equivalent on an XS Max. 2 stops on most other devices. It also has better highlight roll off. So these things alone make it worth using and will give footage a more cinematic feel (something we almost always want).

Regarding 10bit that was stated as “characteristics” of it, which is true in our experience with the example mentioned. It’s obviously not 4:2:2 10bit color space, but like Canon Log which was specifically designed for 8bit video, this new LogV2 performs exceedingly well in this environment. Very low noise and compression artifacts also. Would say the new higher bit rates help here, too

We don’t pretend to understand exactly how all this is done now with computational imaging etc. as we’re filmmakers not video engineers or app developers. But the results speak for themselves.

The best part is this is all just getting started. Can’t wait to see where computational imaging in photography and video can go.
I can respect the cordial reply and I can definitely get behind how exciting of an era this is for people who want to make great video. The tools and education have never been more accessible. None of that has anything to do with my main assertion: flat/log footage on 8-bit camera gear is always a bad choice. As explained in my previous comment, mathematics agrees with me: you can’t magically fit >8 bits of data in to 8 bits, then stretch that compressed data back out (using curves to change one set of color curves to match another) and have it still retain 8 bits of precision after that set of calculations is applied and rounded to still fit in an 8-bit space.

Let me now argue against the points made directly here, and then I’ll post my response comment. I’m skipping most of the post because “has the characteristics of 10-bit” is total bullshit; simple math says that a 256-range color channel can’t represent anything close to what a 1024-range color channel can represent, so that entire line of reasoning is off the table. The note about importance of highlight roll-off is correct, but the ballet images earlier in this article demonstrate yucky highlight roll-off, so I’m not sure where the knowledge ends and the marketing blindness begins on that one. Chroma subsampling isn’t implied to be 4:2:2 by them, so that’s an easily dismissed mention. “Very low noise and artifacts” makes me seriously question if they’ve ever looked at how their footage falls apart, even once! As for not being a video engineer or app developer, neither am I. I just happen to have done more than my fair share of research.

Proving them wrong using their own footage.

“For most DPs dynamic range (including highlight roll off) and color science are the most important things” – first of all, where are you getting that information? Do you know most DPs? Do you actually understand highlight roll-off and color science? Your own video pushing this log shooting app for an iPhone shoots down everything you’re saying! Here, take a closer look at some more stuff from the video:

Xyla test chart

I can see 10 and 11 in “Natural.” They won’t show on a limited range HDMI display due to black levels 0-15 clipping, but they’re DEFINITELY there. Leaving 10, 11, 12 off is intentionally deceptive. Flat styles give you 0.5 stop more dynamic range at best, and their “proof” of 2.5 stops more DR proves them wrong.

This is my favorite test: use Levels in GIMP to crank up contrast, brightness, and bring shadows into midtones. Here, the truth comes into full view. On the left is a 2x zoom of the right side of the FiLMiC PRO LogV2 Xyla dynamic range test chart above, shown exactly as it came from the YouTube video, while the right is my heavily boosted version that pushes the data to its limits and shamelessly reveals those limits. There are some major things that stand out here:

  1. “Natural” has non-clipped black data all the way to the “12” slot, so Natural has 12 stops of dynamic range despite the misdirection that it’s only 9 stops.
  2. The log image on the bottom has brightened shadows, but even at number 8, it’s falling apart. The full damage of dynamic range compression caused by 8-bit log footage is obvious: the “natural” stop bars have more consistent (and therefore useful) color down to at least stop 9 and possibly even stop 10, but the log footage already has inconsistent color at 8 and by 10 it’s jumping all over the place. Even down to 11, the “natural” setting has more consistent and desirable behavior when pushed than the log one does.
  3. At stop 12, there’s basically zero useful information in either profile other than “clipped black” or “one value above clipped black.” Despite this, the log footage is clearly a worse choice at stop 12: it’s so noisy and so mangled by the H.264 AVC compression’s macroblocks that it’s actually revealing garbage data as if it’s legitimate. This is why flat/log profiles on 8-bit cameras (and even to some extent on ALL cameras) don’t actually help: the lowest blacks are full of noise and those profiles result in amplification of the noise.
  4. It’s not part of the stop bars, but…do you see how the solid black background around the bars is a dark grey on “natural” and a more medium grey on the log profile? It’s one thing to show near-black as a crumbling noisy region that’s almost completely invisible, but the total-black blacks being brighter on the log profile shows that it doesn’t even use all the bits appropriately all the way down to solid black. It’s wasting bandwidth, so to speak.

That by itself is pretty damning evidence against the log profile they’re marketing, but I got really curious after calling out the missing “10 11 12” and I wondered: “is there anything PAST 12 that just can’t be seen with the naked eye?” I popped the capture of the Xyla chart back open in IrfanView and did a gamma correction as high as it would go to see what would happen. Remember my assertion that you only gain 1/2 stop max of dynamic range by using log profiles, the one that they denied by saying they got 2.5 stops more DR instead of 0.5? Remember that as you look at the following image.

Half a stop later, we’ve seen the light

Dynamic range chart with stop #13 visible
Oh look, it’s dynamic range the 13th. Jason’s calling, it’s for you.

That’s right! There are some tiny bits of light squeaking through the non-numbered stop #13 on the log profile but not on the “natural” one. It’s clearly not enough to usefully pass through the full stop difference between 12 and 13, but it’s squeaking by in tiny amounts anyway, so that’s probably 1/2 stop of light. Gee, where have I heard “log footage only gives you a maximum of 1/2 stop of added dynamic range” before? Well, I’ll be honest: I originally got all the answers I’d ever want on this subject from an amazing and detailed article on flat/log picture profiles that pulled very few punches when it comes to explanations; I HIGHLY recommend reading it and looking at the excellent graphs and demo images to get a comprehensive understanding of what goes on behind the lens with flatter picture profiles.

While you’re still looking at the image above, take special notice of what the grey area that is supposed to be black looks like. There is inevitably going to be a little bit of light spill from the heavy white side and it’s supposed to taper off as the darkness of the stops increases because that’s just how light works. The “natural” black area has a nice even circle around the brightest stops and it tapers off nicely as it slowly approaches 4. Contrast (pun intended) this with the log profile and you’ll notice that the log profile’s not-quite-black area looks horrible. Not only does it look over-exposed, but there are quality issues all over the place! The smudgy black holes are from compression artifacts being amplified as hard as they can be amplified. It sort of looks like this fancy log profile is mostly just exposing everything higher from the outset . I’ll let you decide what’s going on, but I’m going to say right now that the “natural” profile is the clear winner in this petty little fight.

But wait! We’re not done yet! I’ve completely blown the “FiLMiC PRO LogV2” video in question out of the water, yet I have a few more images to show you to seal the deal, plus my response to their response. Here’s a set of images they displayed of a creek in “natural,” LogV2, and LogV2 with a LUT applied to grade it. Look at these images and pay special attention to differences between them:

If you clicked through the three full images and really paid attention, you probably noticed the same thing I did: the sunlight accent on the left side has been all but lost in the log footage! Not only that, but the stream, ground, and glare on the leaves have been brightened so much that the image has lost any sense of depth it had in the “natural” version. It’s…boring and flat now. The log footage has more shadow detail, but the leaves betray the real reason: the log footage was exposed higher than the “natural” footage. It’s not better because it was shot in a log profile, it’s better because the competition was improperly exposed!

My response to their response

Me: 4:2:2 is chroma subsampling and isn’t really an issue with 8-bit/10-bit compared to compressed color curves. How did you determine that you’re getting 2.5 stops of added dynamic range? What was the dynamic range before? Is the dynamic range in question absolute dynamic range or is it usable dynamic range? The video implies that it’s a guess.

I have looked closely at your 1080p ungraded/graded footage in this video (particularly around 4:15) and I find that there are some serious issues with it; ironically, they’re in the places your arrows point: the noise and macroblock compression banding in the blacks before and after grading are unacceptable to me, as is the highlight banding on the ballerina’s right side and the blocky “plastic” look on her face where a smooth gradient should be. The footage immediately after that’s shot “natural” looks drastically better, though the skin tones need to be a bit desaturated. That’s the thing: it’s 100x better to remove that color difference information in post than to throw it away for more dynamic range and be unable to restore it without heavy secondary corrections. Anyone who wants to confirm what I’m saying can skip between 4:15, 4:20, and 4:35 and decide for themselves.

At 1:45, the natural shot clearly has better color, more detail, and less glare, but it was also underexposed relative to the log footage; since it’s not a high-contrast scene, you could have easily exposed the natural footage a bit higher to retain more of the stream detail. Also notice on the left at 1:35, you’ve got a shining bit of sunlight; that shining bit of sunlight is largely lost after the log grade is applied at 1:41; in fact, the lowered contrast is less attractive because there is no visual depth.

At 8:45, the doll image from the phone is more washed out and obviously lacking in fine detail, most noticeable in the eye detail and in the sharpness of the stripes on the dress. The color lacks a punch that’s visible on the right. I am aware that some of this is because it’s a phone. The results aren’t bad for a phone, but the results are nowhere near equivalent. In the end, I suppose what matters is that the person watching the finished product likes it, so do what works for you. I like a healthy dose of contrast and color, but the washed out desaturated look seems to be all the rage these days, so who am I to argue with the trend?

What should you take from all of this?

tl;dr: Don’t shoot flat without 10-bit color and a professional-grade workflow or you’re trashing your footage, regardless of what some “expert” on YouTube says. It doesn’t look like film, it just looks like amateur hour.

Supplemental information about dynamic range measurement with Xyla charts: Is it really possible to measure a camera’s dynamic range?

2 Replies to “YouTube video experts don’t understand why flat/log footage on 8-bit cameras is a bad idea”

Leave a Reply

Your email address will not be published. Required fields are marked *