It’s no secret the main issue with the Sony FS7 is noise performance. In Ryan Connolly’s side-by-side comparison of the FS7 and Canon C300, he notes the superior color reproduction (which has been recently compared to that of the ARRI Amira), but also the crippling noise performance at native ISO. As an FS7 user since March 2015, I’ve had my fair share of noise problems across several shoots. Up until the recent release of firmware V3, my workaround had been to overexpose the sensor by up to a stop – made easy with Cine-EI mode. As Alister Chapman wrote in his blog post “Understanding Sony’s Slog3”, the problem isn’t that Slog3 is noisier, but rather that it applies gain to the knee of the gamma curve to increase recorded shadow detail (simulating the industry-standard Cineon log curve). If this is true, what is it about the FS7 that’s causing noise issues and how is Sony's fix?
What is digital noise?
Digital noise can be broken down into two main categories – chrominance noise and luminance noise. Chrominance noise patterns depend on the sensor setup – CMOS sensors tend to record information in YCbCr color space (RAW is an RGB exception), meaning it stores the luma value, and the color difference information between red and blue. These three channels can then be broken down into spatial frequencies – simplified for our purposes to “low, medium, and high”.
How does noise reduction typically work?
Since the human eye is more sensitive to changes in brightness over changes in color, rudimentary noise reduction (NR) algorithms start by gently blurring Cr and Cb channels, leaving the Y channel sharp keep the appearing image sharp. This becomes a problem when work required color accuracy (i.e. blue/green screen) because you blur the line between where one color begins and another ends.
Modern NR methods include wavelet based frequency suppression (think EQ for images), and profile-based NR. Neat Video uses the latter, utilizing advanced algorithms to pinpoint sensor-specific spatial frequencies.
Is the FS7 actually “noisy”?
Numerous blog posts, forum threads, and camera tests would say no, however real world experience says yes. I’ve found the camera to perform greatly when given a proper thick negative (i.e. when I have access to a nice lighting package and get time to do heavy preproduction – not every shoot operates this way.)
On Cue93’s most recent production, a music video promo for TIME ENOUGH, I had to balance for the motivating source in the room: projection onto an 8’x16’ mesh wall. During pre-production tests, I read a reflectance reading of f/4 at 180˚ 2000ISO. This was fine for me and I planned to light a tad brighter than that. For ambience, I hung a skirted 6x6’ Full Grid overhead and shot two daylight-balanced Source4 750w fixtures into it, controlled via a DMX lighting console. Our key light was a secondary projector bouncing off a 4x4’ shiny bead board.
What I didn’t take into account in pre-production was the material of the projection surface – it was nowhere close to the real set. I lost about a stop of light, leaving me at an f/2.8. I also didn’t consider the projection footage – f/2.8 was my aperture at the BRIGHTEST point (with most of the footage being on the darker side). This was a major goof on my end, but I needed to fix it. The answer was to plan to consistently underexpose by a stop and try to keep an f/2 at 4000EI. This ended up working well enough, but I was left with predictably noisy images. After several painstaking passes through Resolve, Neat Video, and Gorilla Grain, we ended up with something passable (actually, I rather enjoy it). This is a great example of a DP messing up and the technology falling apart, but what about projects where everything is going right?
For Storylab DC’s HelloWallet promo piece, I had a nice K5600 lighting package and decent grip package. Even after grading, I noticed a surprising amount of noise in correctly exposed parts of the image – namely the blue shirt and green of the window. So what gives?
What were your test settings?
The following test images were all performed with a Rokinon 85mm lens between T5.6 and T11, with a 650w Fresnel slightly above the camera pointed at the Neat Video YCrCb chart. The pink squares on the chart are to measure Cr (with higher luminance squares on the right) and the green squares are to measure Cb. The chart is purposely out of focus to eliminate any printing resolution bias. The dollar is a classic 1st AC trick to test sharpness – print money has incredible detail due to all the security features. A Dropbox folder containing all the raw images from the test can be found here. The camera was in CineEI mode shooting DCI 4K XAVC-I. The only post-processing done was Sony's Slog3 to REC709 LUT and Lumetri exposure compensation inside Premiere.
How does the FS7’s in-camera noise reduction perform?
My initial findings showed most of the noise on the FS7 is in the high/mid frequencies of the Cb channel (which would explain the blue/green noise from the HelloWallet promo). In the following Cb comparison of images at 4000EI, you can see that NR low doesn’t really do much. I found a slight decrease in high frequency noise, but only a bit; it was largely undetectable at EI values lower than 4000EI. The NR middle setting starts to remove larger amounts of high frequency noise, leading to less “fast” noise jumping around the screen - this was discernible at native ISO but undetectable lower.
My main target was the NR high mode – how damaging is it and what are the benefits? As you can see from the images below, its very effective even at 1600EI.
After seeing this, my next question was: “how does it affect sharpness”? As you can see from these close-ups pulled at 200%, there’s no noticeable difference in sharpness. When punched in crazy amounts, there is a small change (pay attention to the MDCCLXXVI at the base of the pyramid), but this isn’t practical to me.
In the following look at a 16000EI chart, you see how much the high NR setting actually does to high and mid-frequency noise. You can also see how it doesn’t really do anything to low frequency noise, but that’s often the most difficult noise to alter.
Charts are cool but any pictures of actual environments?
Funny you ask… I shot a quick clip of our gear room, lit only by the practical lamp pointed at the wall (this image is in Slog3 to make the lowlight noise stand out). It’s obvious to me how much the high NR does to darker areas of the image, but I see no discernible change in the highlights. Also below is a comparison between Neat Video’s NR algorithm and the in-camera NR. I am not surprised that the NeatVideo version removed a larger quantity of higher and mid frequency noise, as well as beginning to alter the low frequency noise. I will definitely be using Neat Video when I can plan for it, but again, the in-camera NR is a huge improvement over nothing.
What is the practical application of all this engineer jargon?
When pushing the camera further than a stop under, I will probably switch the noise reduction to high, but for most work, I will leave it to low. I’m still going to rate the camera at 1600EI, and Neat Video is still vastly superior to the in-camera noise reduction, but I can't always dictate the workflow. It’s comforting to know that if I have to push the FS7 farther than I’m comfortable, the noise reduction is not destructive to sharpness or color accuracy.
What does Sony have to fix with V3.1?
Honestly, I wanted to write an article finding that the noise reduction was flawed in some way and that I had the fix for Sony. Instead, I found firmware V3 was actually very effective and well designed. That said, I do have two suggested fixes for V3.1:
- Noise Reduction setting should be stored in XML metadata. If I have noise reduction on high and I then try to use a pre-existing noise profile for the camera at said exposure index value, I will actually end up severely damaging the quality of the image. A simple extra line in the XML would clear this up for colorists – even though I know most don’t ever actually open the metadata... Along the same line:
- I would like to see an option added to display the NR profile on-screen or in the status menu. In testing, I often recorded the wrong NR setting because I just forgot to change it – if it’s not on screen OR in the metadata, how do I know what profile is active other than scrolling through the menu? This was incredibly frustrating in testing, but I could also see this being another headache for camera operators. Even though we don’t need more on-screen clutter, adding it to the status menu could be a good compromise.
The new V3 firmware update has effectively unburdened the FS7 from previously crippling noise without many drawbacks. Sony has finally hit a home run for a feature that pairs great with Slog3 gamma. All images in this article are available for download so you can give it all a look yourself. Also available for download are the custom Neat Video profiles I’ve built for Slog3 at different EI values – easy to install and they save you from doing a lot of this work yourself! Please feel free to share this article, send feedback, and get in touch!