Sony A7R IV / A7S III / A9 II to feature 8K video, as new 60MP and 36MP full frame sensor specs leak

[webrunner5]

1 hour ago, hmcindie said:

I recently did a quick grade on a fast project. They shot it with a red dragon in 6k. I did some prores files from them and then did a quick grade on top of those. One shot was scaled in 140%. Looked good (we were in an HD timeline) but I decided to replace it with the original 6k file. Snapped it right there in the shot, now the scale was something along 30% so instead of scaling HD up, we scaled 6k down. And... it looked almost the same. I switched those two shots back and forth. Almost no difference in sharpness. Funny.

What, in 12bit or 16-Bit R3D Raw?

[roxics]

I can't afford to shoot on big guns like Alexa, RED or Venice, so I'm never really getting the true potential of 4K from a 4K H.264/5 camera. 

Bring on the 8K. So long as most of these Japanese camera makers are going to continue to deliver us 8bit H.264/5 files, at least give me more resolution so I can make up the difference in post. 

 

[Shirozina]

12 minutes ago, roxics said:

I can't afford to shoot on big guns like Alexa, RED or Venice, so I'm never really getting the true potential of 4K from a 4K H.264/5 camera. 

Bring on the 8K. So long as most of these Japanese camera makers are going to continue to deliver us 8bit H.264/5 files, at least give me more resolution so I can make up the difference in post. 

 

You can't make up the DR or colour fidelity in post.

[sanveer]

1 hour ago, Shirozina said:

You can't make up the DR or colour fidelity in post.

There is a (debatable) school of thought that says that downscaling perfectly from 4k 3840 x 2160 to HD 1920 x 1080 ensures that there is more colour information and less noise, and in thus in theory, the equivalent of more bit depth. There have been lots of articles to push this theory forward. I agree with some of the viewpoints, but I have neither tried it nor an I am sure of how much the inage quality improves. 

 

(Translated this page cause it's in German). 

https://www.slashcam.de/artikel/Grundlagen/4K---Von-RAW-bis-4-2-0.html

And 

https://***URL not allowed***/best-way-downsample-4k-smartphone-footage/

 

 

[Shirozina]

13 minutes ago, sanveer said:

There is a (debatable) school of thought that says that downscaling perfectly from 4k 3840 x 2160 to HD 1920 x 1080 ensures that there is more colour information and less noise, and in thus in theory, the equivalent of more bit depth. There have been lots of articles to push this theory forward. I agree with some of the viewpoints, but I have neither tried it nor an I am sure of how much the inage quality improves. 

 

(Translated this page cause it's in German). 

https://www.slashcam.de/artikel/Grundlagen/4K---Von-RAW-bis-4-2-0.html

And 

https://***URL not allowed***/best-way-downsample-4k-smartphone-footage/

 

Yes downsampling helps and infact is essential for best quality (1:1 can't produce optimal images due to the Nyquist Theorem) but for better colour the sensor has to capture it in the first place and that data can't be interpolated by downsampling and nor can DR be increased beyond what the sensor can capture even if  downsampling and interpolation can reduce noise. 

[sanveer]

1 hour ago, Shirozina said:

Yes downsampling helps and infact is essential for best quality (1:1 can't produce optimal images due to the Nyquist Theorem) but for better colour the sensor has to capture it in the first place and that data can't be interpolated by downsampling and nor can DR be increased beyond what the sensor can capture even if  downsampling and interpolation can reduce noise. 

After looking at various smartphone HDR photos I am convinced that cameras and photos too have a lot more dynamic range than we know and can imagine. Photos start galling apart in post more due go the bit and colour depth than due to anything else. If this characteristic of the photo can be handled better, we don't have to worry about assuming that there isn't too much information to begin with. Computational Photography and AI in post should actually, in theory, be able to improve dynamic range of photos even after they've been taken. It's almost like turning B&W photos to colour. The limitations right now, IMHO, are more to to with a lack of imagination by programmers and experts in the field, than anything else. I am not sure you're right about interpolation and it's limitations. Those are more of the accepted limitations of the present times based more on unexplored methodologies more than anything else.  In theory HDR should be possible from a photo, whether physical or digital, and also up-ressed and increased in bit depth. I actually think that downsampling a 4k photo to HD will improve the noise and increase the colour information. It can then be upressed back go 4k and it might be way better than the original photo. It would obviously require a lot of processing power and great algorithms. Also it may not yield the same results and would vary from photo to photo (depending upon the quality of the original photos to begin with). There is that Sensor fusion process where information from 2 sensors is read and a composite image is created with MP count from one and noise foem the other helping out. The same thing can be done with a single image by undergoing a few steps to help improve it vastly. 

[webrunner5]

2 hours ago, sanveer said:

There is a (debatable) school of thought that says that downscaling perfectly from 4k 3840 x 2160 to HD 1920 x 1080 ensures that there is more colour information and less noise, and in thus in theory, the equivalent of more bit depth. There have been lots of articles to push this theory forward. I agree with some of the viewpoints, but I have neither tried it nor an I am sure of how much the inage quality improves. 

 

(Translated this page cause it's in German). 

https://www.slashcam.de/artikel/Grundlagen/4K---Von-RAW-bis-4-2-0.html

And 

https://***URL not allowed***/best-way-downsample-4k-smartphone-footage/

 

 

Andrew had an article about that on the GH4.

https://www.eoshd.com/2014/02/discovery-4k-8bit-420-panasonic-gh4-converts-1080p-10bit-444/

[Shirozina]

1 hour ago, sanveer said:

After looking at various smartphone HDR photos I am convinced that cameras and photos too have a lot more dynamic range than we know and can imagine. Photos start galling apart in post more due go the bit and colour depth than due to anything else. If this characteristic of the photo can be handled better, we don't have to worry about assuming that there isn't too much information to begin with. Computational Photography and AI in post should actually, in theory, be able to improve dynamic range of photos even after they've been taken. It's almost like turning B&W photos to colour. The limitations right now, IMHO, are more to to with a lack of imagination by programmers and experts in the field, than anything else. I am not sure you're right about interpolation and it's limitations. Those are more of the accepted limitations of the present times based more on unexplored methodologies more than anything else.  In theory HDR should be possible from a photo, whether physical or digital, and also up-ressed and increased in bit depth. I actually think that downsampling a 4k photo to HD will improve the noise and increase the colour information. It can then be upressed back go 4k and it might be way better than the original photo. It would obviously require a lot of processing power and great algorithms. Also it may not yield the same results and would vary from photo to photo (depending upon the quality of the original photos to begin with). There is that Sensor fusion process where information from 2 sensors is read and a composite image is created with MP count from one and noise foem the other helping out. The same thing can be done with a single image by undergoing a few steps to help improve it vastly. 

Digital sensors have a fixed and measurable DR which is from saturation at one end to where the signal can’t be separated from noise at the other and it doesn’t matter if it’s 8 bit or 16 bit. Also interpolation or resampling can’t recover data not captured in the first place. The  first rule of digital is garbage in and garbage out.

[roxics]

5 hours ago, Shirozina said:

You can't make up the DR or colour fidelity in post.

No. It's not going to add the secret sauce of an Alexa for instance, but it's still better to have more resolution if all they are giving you is an 8bit image. Which is what we typically get on these cheaper cameras. It means I can produce a nicer 4K image in the end. Not Alexa nice, but better than using a 100% crop of a GH4. 

[seanzzxx]

http://yedlin.net/ResDemo/

I'll just leave this here. Going 6k before 10-bit is the worst single thing they could do.

[gethin]

I dunno. I'm finding myself skeptical. Seems like the hoohar has arisen because of those 2 sensors, but we've been here before. I'll happily admit ive been wrong before too.  

 

 

[Mark Romero 2]

Personally, more interested in HDR / HLG / Rec 2020 4K than I am interested in anything 8K.

And to be honest, anything 8K is going to require the wider color gamut most likely anyway. 

[KnightsFan]

On 11/26/2018 at 11:19 AM, Shirozina said:

You can't make up the DR or colour fidelity in post.

Actually you can.

For DR, downscaling reduces noise because for each pixel in the downscaled image, you combine four signal values that are (almost always) highly correlated, and combine 4 noise values that are not correlated. Thus, your SNR will be lower and with a lower noise floor, you have more dynamic range in the shadows. You can verify this by using dynamic range testing software, or by a simple calculation-- imagine 16 adjacent pixels that each have an ideal value of 127 out of 255 (e.g. you are filming a grey card). Add a random number between -10 and 10 to each one. Calculate your standard deviation. Now, combine the pixels into groups of four, each of which has an ideal value of 508 out of 1020. Calculate the standard deviation again.

The standard deviation on the downscaled image will be lower if the random number generator is, in fact, random and evenly distributed.

(This works because in the real world, the signal of each pixel is almost always correlated its neighbors. If you are filming a random noise pattern where adjacent pixels are not correlated, you could expect to see no gain in SNR.)

 

As for color fidelity, a 4:2:0 color sampled image contains the same color information as a 4:4:4 image that is 25% of the size. Each group of 4 pixels, which had one chroma sample in the original image, is now 1 pixel with 1 chroma sample.

[Shirozina]

38 minutes ago, KnightsFan said:

Actually you can.

For DR, downscaling reduces noise because for each pixel in the downscaled image, you combine four signal values that are (almost always) highly correlated, and combine 4 noise values that are not correlated. Thus, your SNR will be lower and with a lower noise floor, you have more dynamic range in the shadows. You can verify this by using dynamic range testing software, or by a simple calculation-- imagine 16 adjacent pixels that each have an ideal value of 127 out of 255 (e.g. you are filming a grey card). Add a random number between -10 and 10 to each one. Calculate your standard deviation. Now, combine the pixels into groups of four, each of which has an ideal value of 508 out of 1020. Calculate the standard deviation again.

The standard deviation on the downscaled image will be lower if the random number generator is, in fact, random and evenly distributed.

(This works because in the real world, the signal of each pixel is almost always correlated its neighbors. If you are filming a random noise pattern where adjacent pixels are not correlated, you could expect to see no gain in SNR.)

 

As for color fidelity, a 4:2:0 color sampled image contains the same color information as a 4:4:4 image that is 25% of the size. Each group of 4 pixels, which had one chroma sample in the original image, is now 1 pixel with 1 chroma sample.

In the context of the original post I was responding to you can't make up the difference between and Alexa's  DR and colour capture buy subsampling an 8k 'consumer grade' sensor. 

[KnightsFan]

12 minutes ago, Shirozina said:

In the context of the original post I was responding to you can't make up the difference between and Alexa's  DR and colour capture buy subsampling an 8k 'consumer grade' sensor. 

Oh, ok, my mistake. I thought you were responding to the second paragraph you quoted, about 8k making up for 8 bit h264.

[Gordon Zernich]

On 11/24/2018 at 2:01 PM, chrisE said:

Read my post. I don't need it, but it could be handy. Of course, all the (wannabe?) "pro's" out there who shoot everything perfect, don't need to crop or reframe in post. ?

What I wanted to say if you don't understand it, is that it can be useful to have a higher resolution to crop/re-frame.

excellent point; thank you

[Mokara]

On ‎11‎/‎23‎/‎2018 at 4:45 PM, Andrew Reid said:

Sony has developed a duo of 8K capable full frame sensors, which will soon be released in two Sony cameras.

Read the full article

Got the specs sheets @androidlad?

Having a sensor that can do these things is one thing, having a processor capable of handling the data flow is quite another.

On ‎11‎/‎26‎/‎2018 at 10:28 AM, Shirozina said:

Yes downsampling helps and infact is essential for best quality (1:1 can't produce optimal images due to the Nyquist Theorem) but for better colour the sensor has to capture it in the first place and that data can't be interpolated by downsampling and nor can DR be increased beyond what the sensor can capture even if  downsampling and interpolation can reduce noise. 

You will get more accurate color from downsampling because the composite pixel is based on more information than a single physical pixel. It will also increase the bit depth of the composite pixel (assuming the original data was 8 bit, it would convert it to 10 bit). That does not mean increased dynamic range however, but it would result in more accurate color and luminosity. Shooting in 8K with a beyer filter in place means that you should be able to resolve true color at 4K resolution (assuming you are using a RAW feed of course) since each composite pixel would be receiving input from two green pixels and single red and blue pixels.

[Shirozina]

11 hours ago, Mokara said:

Having a sensor that can do these things is one thing, having a processor capable of handling the data flow is quite another.

You will get more accurate color from downsampling because the composite pixel is based on more information than a single physical pixel. It will also increase the bit depth of the composite pixel (assuming the original data was 8 bit, it would convert it to 10 bit). That does not mean increased dynamic range however, but it would result in more accurate color and luminosity. Shooting in 8K with a beyer filter in place means that you should be able to resolve true color at 4K resolution (assuming you are using a RAW feed of course) since each composite pixel would be receiving input from two green pixels and single red and blue pixels.

The CFA density dictates the colour fidelity and accuracy. Increased subsampling will help but I repeat it can't invent colours that were not captured. By this I'm talking the very slightest differences in hues needed to get critical colours in skin tones right. I shoot 8k stills in 14bit RAW but I don't get a miraculous improvement in colour by sunsampling to 2k and certainly not anything approaching the quality in colour from a good MF digital back which isn't trading sensitivity for colour fidelity with weaker CFA density.  

[androidlad]

IMX435AQJ datasheet, I'd be surprised if A7S III didn't use this sensor/variants of this sensor.

[Andrew Reid]

Did this data sheet come from a hack or something? It says version 0.3 and has a placeholder [Product information] showing that it is incomplete.

If true, 3840 x 1080 at 480fps would be very tasty.