Robert Niessner

Hi Gunleik,

very interesting reading. Unfortunately the download files seem to be offline. Could you please supply a download link again?

I've tried playing with all of these settings, and exporting 2K tiffs and 1K tiffs. I've hardly ever noticed any differences with DRX (wtf is that?), De bayer and De noise.

DRX = Dynamic Range Extension

If you got a clip with very bright highlights you can dial the highlights back - without getting a color cast like on video. But be aware that this setting does not work well in every situation - fire for example should be treated with care (set DRX to a low value).

DeBayer means how precise the debayering algorithm used is. The better the more CPU power it takes.

Debayering converts the Bayer grey pattern of the CMOS to the RGB image.

De Noise means that the image gets put through a denoising algorithm - it's not very effective against true noise but can help with some chroma noise and color moiree.

OLPF compensation will dial in some sharpening, better leave it to LOW or OFF and sharpen your image later, it can produce some nasty dark lines in the highlight contrast border.

Thanks Keith for the in-depth explaination. Well done and easy to understand.

I find it always difficult and timeconsuming for me to explain such a topic in English (which is not my mother language), I am glad you did this job.

Peter, I hope you now understand why your ideas won't work.

Robert I appreciate you and John's input and willingenss to share your knowledge, but your assertion (which FWICT John agrees with) pretty much directly conflicts with what Sony says about high end digital cinema CCD's. I believe I've explained my position well and have given cameras examples that illustrate the data level and workflow needed to truly capture a CCD's full dynamic range. If clipping were happening on the photosite, then the Viper and Dalsa's wide data options would offer no advantage.

 

But here are Sony's own words:

 

"A major challenge to quantization is that today's best CCDs have tremendous

dynamic range, which film photographers call "exposure latitude." The CCD in

Sony's F35 digital cinema camera can sense meaningful detail up to 800% of

nominal peak white. An exterior daytime wedding scene, for example, will retain

the texture of the lace pattern in the bride's white gown without losing the

difference between the black wool and the black silk in the groom's tuxedo."

Peter, I already know that text - but be careful quoting marketing blabla as technological fact. This is Sony marketing speech for people how don't know much about electronics. That's ok, but not a fact.

What does 800% above nominal peak white mean? It means its photosites get filled up after 8 times the amount of photons have hit them compared to the nominal peak. In other words: in best case up to 3 f-stops. As each color channel has different total fill levels this would be the maximum for the 'best' channel - normally the green channel. Other channels might clip earlier, leading to the typical video overbright wrong color look e.g. in skin tone highlights with the red to yellow to white ramp. To avoid this you clip earlier and desaturate which looks more like analog film highlights. Normally you will have a reserve of 2 f-stops to dial back. This is exactly what you get with a DSLR and RAW vs JPEG - you can dial back 2 f-stops in the highlights in the RAW image (where the JPG would already be clipped).

So Sonys marketing text just underlines what I already wrote before.

I have no doubt that you worked in sensor design, but the fact of the matter is that Sony's highend CCD's respond to a dynamic range WELL IN EXCESS of anything that's recorded to a 10-bit colorspace like HDCAM-SR 4:4:4. That is just a fact. Clipping is absolutely done during the ADC stage and not by the photosites of the F35's CCD.

 

And if you doubt this, look at the Viper. It really records raw in its FilmScribe mode and can exhibit amazing dynamic range. Same idea behind the Dalsa.

 

I feel like I'm going around in circels here.

Again: The ADC does NOT clip highlights. It converts analog voltage level to digital level. Point. No discussion. It is a technological fact. Photosites can only handle a certain amount of photons, until they are filled. That means clipping. Otherwise you would claim the CCD to have unrestricted DR.

You are mixing again different technologies. HDCAM-SR is NOT RAW. It has uncompressed color channels, but is converted from RAW by using curves and LUTs. For the sake of a better looking image some clipping is applied by using contrast enhancement.

And I never said that for those formats data processing will not clip DR. In fact, that's exactly what I said.

But RED RAW is NOT HDCAM-SR. It is RAW data coming from the sensor with NO image processing applied (except the wavelett compression and some calibration for dark current and basics).

Please take a read of how CCDs and RAW data work - it is nicely written here:

http://www.sc.eso.org/~ohainaut/ccd/CCD_proc.html

This is simply not true. Your typical CCD sensor has much more dynamic range than the recording medium allows. During the first round of analog to digital quantization, the image processor applies a knee to the upper end. This has nothing to with legal broadcast or what are normally called superwhites between 236 and 255 on the 8-bit scale. It has to do with the fact that sensor has a wider range of sensitivity than even fourteen bits would allow. But some point has to be called absoulte white and somewhere before that point a knee is applied to have the hightlights more gracefully transition to what is designated as pure whie.

 

Also, none of these chips record in RAW w/ capital letters. AYK, RAW is a lossy Red codec which records 12-bit mosaiced color. That's nowhere near what's coming off a CCD sensor.

 

My ? is how much image processing is done on-chip by the CMOS sensors used in a typical camera that bakes-in WB, gain, knee, etc.. An example of such a camera would be Sony EX-1/3 for the chip design and the F35 for single chip Bayer pattern design. You guys seem to think that the CMOS sensors in cameras like these are limited to just a basic straight line quantization w/o knee, gain or any image correction or adjustment applied. I'm just not sure if that's the case.

Sorry, but you seem to have a wrong impression of that technology.

I have worked 3 years in developement of highend CCD scanners for Vexcel Corporation (now Microsoft Photogrammetry) and I surely know one or two things about sensor and data as you might imagine.

RAW is raw data from the sensor. And nothing else. RED RAW is compressed RAW data. And nothing else. Please do not argue around those facts with me.

The dynamic range of a sensor is no problem for 14-bit ADCs. Clipping is done by the sensor itself, as soon as the wells ar 100% filled. Sensors record light linearly (our eyes do it logarithmic), resulting in a very dark RAW image with noise levels equally distributed at all levels. For viewing purposes you apply a gamma correction curve so brightness levels look correct. That means you spread the dark parts of the image and that is, where the high bit-depth is needed. The highlights are no problem because they even get compressed.

Now the noise in the darks gets also boosted and that's the reason why there is visible noise in the darks of digital video but almost no visible noise in the highlights. Basically the more noiseless a sensor is, the better is its dynamic range.

You can compress any dynamic range into any end format - question is how dull it will look and when will you face banding.

For a contrastier look on TV monitors you have the sacrify DR and clip levels by applying a s-curved contrast enhancement, you fit your DR to the DR of you output medium.

So for instant good looking pictures all cameras do this DR limiting within their image processor before recording to tape. The RED and other RAW cameras do NOT do this. They record the full DR of the sensor.

Do you know if CMOS cameras that bake-in user selected image adjustments, perform gain, knee and white balance calculations on chip? Or are they typically done off-chip with the sensor "only" doing light gathering and the initial analog to digital conversion?

 

I would think at least knee and gain would be done during the CMOS sensor's initial on-chip quantization, BIDKFS.

Peter, neither CMOS nor CCD is baking anything like knee or wb into the signal. They all record RAW. This RAW data is sent to an image processing device where those adjusments will be done. The concept of knee, superwhite and superblack is only for broadcast formats, it is basically a manipulated gamma curve. The reason for superwhite and superblack is merely because broadcasted ntsc/pal signals couldn't work with that data and had to be cropped off otherwise causing all sort of interferences because they were too hot.

When you watch footage from DV it is already cropped in the highlights from RAW - even with superwhites. If you can look at RAW data even from consumer cameras you will see they have a much better dynamic range than you would think from the compressed DV footage.

It makes no sense to me that you would argue for having basic image adjustments, like white balance, be done at the post production stage where data quality is lower. Wouldn't you want white balance calculation to be done at the sensor level during the analog to digital quantization? That is when data is truly off the sensor.

Peter, that is the point where you got the RAW approach wrong.

The data quaility in post is NOT lower, the only thing happend to the raw data is the wavelett compression to keep data rate manageable.

You could work with WB filter if you want as the sensor is lightbalanced to daylight. It will affect noise levels in the blue channel. But that's it.

RAW data workflow is not comparable to other digital formats with baked in corrections and heavy compression like HDV.

Re-whitebalancing HDV for example is very bad.

White balance compensation in post is performed upon an image that has undergone three stages of processing: 1) the analog to digital quantization at the sensor level, 2) compression to .R3D and 3) demosicing.

 

I can understand you arguing against trying to lock down a look in-camera (even if I tend to disagree). But basic image adjustments, like white balance, I see no good reason to leave for post. Leaving everything for post seems more like a philosophy than a technical advantage.

As someone doing a lot of postprocessing and VFX I can assure you that RAW is a godsend for me - it leaves all possibilities open for tweaking and getting the most out of it. That doesn't mean you should be a careless shooter. A good DP will provide the solid basis for the best RAW footage which you can manipulate to anything you want. Ideally you won't have to tweak anything, but you will have all freedom of choice if you change your mind later. It is very easy to set raw wb in post, but can be hard to correct baked in improper wb.

In the case of white balance, all it does is add another stage of processing to post, and its never going to look better than if the math was done in-camera.

It won't look worse either. Because the camera can't do anything to RAW what you couldn't in post.

The extra stage in post is not relevant as the data has to processed anyway.

UPS, didn't see there were already a bunch of answers...

It's important to remember that before cameras like the RED came along, video cameras generally didn't respond to infrared at all and neither did movie film, so there was never any need to engineer an IR blocking characteristic into filters.

There exist a lot of video cameras which have trouble with infra red. Just point your remote control into the optics and watch your display.

A lot of Canon and Sony cameras show the IR response to some extend. Especially the Sony EX1/3 are very sensitive to IR, resulting in brown colors on dark gray or black cloths when light with tungsten lights. DSLR seem to have better protection, e.g. my Canon EOS400D does not show any problems with IR. CCDs and CMOS sensors always needed IR cutoff filter.

Dear Cinematographers,

I have some questions regarding film grain.

I am already quite successfully offering denoising for RED and other digital sourced footage as a service.

Lately I have done a degraining job in town for a company which shot a documentary on S16mm / 50 ASA - which turned out to be very grainy due to bad light conditions (cloudy winter day in the morning). The denoised footage turned out very well (better than I thought) and the company is going to shoot on film again instead of using HDCAM because grain is no issue anymore for them.

I can show before and after samples if someone is interested.

So here are my questions:

1) Has the type of telecine process any influence on the graininess of the DI?

2) If yes, which is the best process, which the worst? And how much is the difference cost wise?

3) If you look at your own footage - would you wish it to be less grainy or are you ok with it at most of the time?

4) Would you consider the removal of grain as bad for the 'look' or do you think cleaner would be better because the release print will add grain back anyway?

5) Do you see a demand for degraining film?

And the final question:

Are there any DPs here who own the rights to their 35mm footage and are willing to give me some short DIs (2k/4k DPX preferred) for testing purposes?

If I am allowed to use that footage for my showcase I would in return do a degraining job of a few minutes of footage for free.

Amount of grain removal can be chosen, so fear no plastic look ;-)

Thanks in advance for your insights and opinions.

Cheers,

Robert

Some new examples:

First one is a 120fps 2k shot of a sunset.

I tried some new tricks to create a HDR like image, ISO-160 and ISO-1280 combined, both denoised.

Please click the preview images to see the full HD versions.

Source ISO 160 - directly out of RedAlert:

Source ISO 1280 - directly out of RedAlert:

Denoised HDR (ISO 160 + 1280) + unsharpmask set to 300/0.5/12:

Second one is a campfire with a 2-bank kino and some lights for the shrubbery in the back (which was quite weak).

Please click the preview images to see the full HD versions.

Source ISO 160 - directly out of RedAlert:

Source ISO 1280 - directly out of RedAlert:

Denoised HDR (ISO 160 + 1280) + unsharpmask set to 300/0.5/12:

It would be great to see these images in montion....did I miss the link or do you have a demo disc?

Hi Chris,

you can download before and after examples in my posting #2 - the links just below the images.

It is a RAR-Archive with DPX files inside. You will need an editing suite which supports DPX or

any compositing software like Combustion, After Effects, ...

If you have no possibility to work with those DPX files, just tell me what you would prefer and for

which plattform (PC/Mac). I would suggest a h.264 encoded MP4 file because of the quality/size ratio.

Matt, I strongly advice you to have your assistant testing the workflow beforehand and put some thoughts into media management and backup solutions. You know - Murphy's laws always strike you most when you don't expect it.

A laptop with two external raid-1 drives will be a good, mobile solution. I tend also to burn some extra blurays of the raw data for an additional backup with an optical media.

It's also helpful to have some thoughts into folder structure for the footage, it will make your live easier. Have a look at the free software clipfinder which makes working with raw a breeze.

It seems like they overdid the effect and put the distance between cameras too wide.

I think 3D can be a nice feature but it has to be stup carefully not to strain the watcher.

A friend of mine here in Graz, Austria is working on 3D since years and he showed me some really impressive footage e.g. a flight over a mountain in 3 - it's so real that you feel dragged in.

Hey Mike,

You can call me Robert ;)

I'm pretty impressed, though I think that a bit less sharpening/denoising would look better - in the car interior shot there's an artificial black line between the man's coat and shirt post-sharpening that wasn't there in the first frame.

That is clearly the result of the sharpening process - common sharpening works by enhancing local contrast. When digitally projected it will be less visible because the projector, lens and screen work as bluring devices. When filmed out on 35mm it will be less visible too.

It totally depends on where you watch it. For example some of the cheaper, latest Samsung computer 24" LCDs have some kind of digital sharpening/local contrast enhancing, where any footage looks artifically noisy (God knows what their engineers were thinking), while on my older DELL 24" LCD it looks perfect.

And the girl's dark hair doesn't have much detail in it after processing, probably because it was buried in the noise of the toe of the curve to begin with.

This is more a problem of motion blur (she's leaving the car), because in my first post there is another frame from the same scene where her hair is quite detailed.

I also think the skin tones after denoising look too plastic and smooth to me, whereas leaving a bit of noise in the image would have made it more natural - after all we're all used to shooting 500T film with acceptable results, which are not grainless.

Did you inspect the 1080p examples 1:1 on a full HD monitor or on one with smaller rez?

I tried to avoid plastic skin and removed noise to my personal taste. But I can keep any amount you want - or if you prefer - regrain it to match a certain film stock.

About noise and RED: I have spent a whole lot of time in researching denoising techniques and as a result I am now proudly offering my footage denoising service. You can have a look at the results you can expect in this thread, which I will constantly update with more examples.

The RED ONE can be pushed very far with my postprocessing techniques, I have provided examples of ISO 1280 and ISO 2000 which I can get very clean without losing lots of details.

@ David Mullen

David, your examples show the typical blue channel problem of the RED. If you are interested I can give you a shot on those examples if you can provide me with the RED RAWs. I guess I can get rid of most of the noise in the darks.

Finally here are two moving examples in DPX format:

First one is the same shot inside the car from above,

just some frames with motion of the talents so you

can see how much detail I can preserve.

Source + sharpening:

Denoised + sharpening:

Car inside - 1080p DPX Original footage (processed through RedLine) - 452 MB, RAR archive

Car inside - 1080p DPX Denoised footage - 367 MB, RAR archive

The second one is a car rigged night shot - car driving out

of a tunnel - no extra light except the original street lights.

This one was very tricky and took me a lot of time to get

rid of the random moving blue blobs from the blue channel

(which had almost no detail in the shadows - just ugly

wavelett compressed noise).

Please click the preview images to see the full HD versions.

Source + sharpening:

Denoised + sharpening + some extra magic:

tunnel ride - 1080p DPX Original footage (processed through RedLine) - 458 MB, RAR archive

tunnel ride - 1080p DPX Denoised footage - 362 MB, RAR archive

To extract on OS-X you could use the free and simple UnrarX

Please use WinRAR v3.6 (or any compatible unarchiver) or later to extract on Windows.

Don't hesitate to contact me for saving your valuable shots - it

might cost less than you would expect. :lol:

You can PM me here at Cinematography.com or just write me an email.

Usually you can expect an answer on the same business day

depending on your location (I am located in central europe).

The problem:

You've got that essential shot under low light conditions

e.g. with your RED ONE but results came out noisier than you

can bare (especially in the blue channels)?

I am sure a lot of you have already faced this ugly problem.

The solution:

But don't worry, there is a lot I can do for you to save

that special shot and to spare you a re-shooting!

I am offering my postproduction skills in denoising/degraining your

film/RED footage (or any other HD source, even HDV can be improved).

This is a shot made with the RED ONE and Zeiss Primes at a lake where sun was nearly gone and the only lightsource is the fire place in front of the actors.

The shot has been processed in RedAlert with OLPF compensation set to maximum, chroma denoise set to OFF, ISO = 800, curve ramped up in the shadows, scaled down to 1080p. Originally you won't see much in the shot - it is really dark.

Please click the preview images to see the full HD versions.

Source - directly out of RedAlert:

Source + AFX filter sharpen set to 100:

Denoised + AFX filter sharpen set to 100:

This is another example which I put to extremes:

This is a shot made at night, just some kinos outside - shot with ISO 320 - I pushed it up to ISO 2000 just to see how far I can go.

The shot has been processed in RedAlert with OLPF compensation set to maximum, chroma denoise set to OFF, ISO = 2000, curve ramped up in the shadows, scaled down to 1080p.

Please click the preview images to see the full HD versions.

Source ISO 320 - directly out of RedAlert:

Source ISO 2000 - directly out of RedAlert:

Source ISO 2000 + unsharpmask set to 300/0.5/12:

Denoised ISO 2000 + unsharpmask set to 300/0.5/12:

I think it's quite impressive what the RED ONE can give in very low light conditions - after some special postpro treatment.

Don't hesitate to contact me for saving your valuable shots - it

might cost less than you would expect. :lol:

You can PM me here at Cinematography.com or just write me an email.

Usually you can expect an answer on the same business day

depending on your location (I am located in central europe).