GH4 4k 8bit vs external 4k downconverted to 1080p 10bit?

Bunk & others,
Just did some tests myself. I can say that the UHD mode downconverted in camera to 1080 10bit out is noticeably better then the 1080p mode 10bit out via HDMI. More detail in the image. I watched both shots on a 60" plasma 2ft away and took stills on my phone to compare what my eyes were seeing live. I don't have a 7Q+ to test but will be getting a Video Assist on Friday to compare internal/external/downconvert in post.
 
soarprod said:
Bunk & others,
Just did some tests myself. I can say that the UHD mode downconverted in camera to 1080 10bit out is noticeably better then the 1080p mode 10bit out via HDMI. More detail in the image. I watched both shots on a 60" plasma 2ft away and took stills on my phone to compare what my eyes were seeing live. I don't have a 7Q+ to test but will be getting a Video Assist on Friday to compare internal/external/downconvert in post.
Click to expand...
Cool, keep us posted. You might want to read this post about down converting on the Video Assist.
http://www.dvxuser.com/V6/showthrea...Assist&p=1986586270&viewfull=1#post1986586270
 
lord have mercy...

I only wanted to know a simple thing and I still dont feel any better haha.

what I wanted to know was if UHD internal 8-bit downscaled in post to 1080 was better or worse than UHD sent out to ninja 2 (resulting in a 1080p 10bit file)
 
Filmautre said:
lord have mercy...

I only wanted to know a simple thing and I still dont feel any better haha.

what I wanted to know was if UHD internal 8-bit downscaled in post to 1080 was better or worse than UHD sent out to ninja 2 (resulting in a 1080p 10bit file)
Click to expand...
I guess part of all this confusion is due to the fact that one cannot send UHD to a Ninja 2 to record it as 1080p. The Ninja 2 cannot accept a 4K signal and down convert it to HD. The Odyssey7Q+ can, which is why I mentioned it.
 
Which is the reason why I said no 8-bit file is going to grade better than a 10-bit file, regardless of resolution. But I also said I use a Ninja 2 with my GH4 and use the Gh4's HDMI downconvert function with excellent results.
 
Mitch Gross said:
I guess part of all this confusion is due to the fact that one cannot send UHD to a Ninja 2 to record it as 1080p. The Ninja 2 cannot accept a 4K signal and down convert it to HD. The Odyssey7Q+ can, which is why I mentioned it.
Click to expand...

Yes, and the whole point was that the 10-bit HD output by the GH4 when shooting (but not recording) QHD (assuming you have set the output options properly) is a great cheap option to get quality 10-bit HD from a GH4.

I speculate idly here: I have often wondered why, if we can get such a nice HD image out of HDMI in this way, why there is not an option on the GH4 to record this down sampled HD internally rather than the crappy binned HD we get internally. I would never shoot HD with the GH4 internally unless I was using the tele crop mode or for slo-mo...

The fact that we do not have this made me wonder whether the HDMI down sampled HD is actually using a dedicated chip rather than the GH4's internal processor. Seems a bit too niche a product, though... Alternatively, I guess it could be that it uses the computing power left over when it does not have to compress to h264 to do the nice scaling...
 
drokeby said:
I have often wondered why, if we can get such a nice HD image out of HDMI in this way, why there is not an option on the GH4 to record this down sampled HD internally rather than the crappy binned HD we get internally.
Click to expand...
That would be my number one wish for the GH5, even at 8 bit. But like you said scaling and compressing is probably too much asked from the processor.
 
It looks like in the end that I may be going for the PIX-E5H .... did consider the assassin but not convinced by their customer service.
 
I would imagine that for any heavy grading, the 10-bit ProRes HQ files from the Ninja will hold up better than the 8-bit 4K internal files from the GH4. I plan to test this when the right project comes along where I need 10-bit for color. Just wanted to chime in that for whoever tests this, make sure to use ProRes HQ over ProRes422. I know that makes a big difference when comparing C100 internal vs external, granted that cam can't output 10-bit like our mighty GH4's can.
 
estarkey7 said:
Which is the reason why I said no 8-bit file is going to grade better than a 10-bit file, regardless of resolution.
Click to expand...
If you're speaking esthetically, I can see how you might justify this opinion. But as you're venturing into mathematical territory, your claim is mistaken. An 8-bit file at 2X resolution contains more image detail than a 10-bit file at 1X resolution. The 8-bit 2X file can be mathematically downsampled into a 10-bit 1X file with no significant data loss other than resolution. Once the file is downsampled from 2X to 1X, its original 2X resolution cannot be recovered without significant data loss.
 
Lpowell said:
If you're speaking esthetically, I can see how you might justify this opinion. But as you're venturing into mathematical territory, your claim is mistaken. An 8-bit file at 2X resolution contains more image detail than a 10-bit file at 1X resolution. The 8-bit 2X file can be mathematically downsampled into a 10-bit 1X file with no significant data loss other than resolution. Once the file is downsampled from 2X to 1X, its original 2X resolution cannot be recovered without significant data loss.
Click to expand...

Using what downsampling algorithm?
I agree you can get better color accuracy downsampling 4K 4:2:0 to 1080p 4:2:2, but as soon as you try and grade the resulting downsampled image, it's true 8-bit heritage will undoubtedly show. Nothing is going to take an 8-bit color channel's 255 and make it have 4096 levels.
 
estarkey7 said:
Using what downsampling algorithm?
I agree you can get better color accuracy downsampling 4K 4:2:0 to 1080p 4:2:2, but as soon as you try and grade the resulting downsampled image, it's true 8-bit heritage will undoubtedly show. Nothing is going to take an 8-bit color channel's 255 and make it have 4096 levels.
Click to expand...
I agree that it is not more than HD 10-bit origination or even the same as HD 10-bit origination, but it is more than HD 8-bit origination. I refer to it as pseudo 10-bit. Remember that the UHD may be 8-bit so 255 levels, but you are also sampling four pixels down to one, so it can use the differences between the four different 255 samples and put that fractional measurement into the 4096 storage space. In a grand simplification, if one pixel reads a value of 50 and one next to it reads 51, the new measure becomes 50.5, which is more precise. Imagine a blue sky feathering slight color and density shifts and yes, this level of extra color information is enough to significantly improve banding issues.
 
Mitch Gross said:
I agree that it is not more than HD 10-bit origination or even the same as HD 10-bit origination, but it is more than HD 8-bit origination. I refer to it as pseudo 10-bit. Remember that the UHD may be 8-bit so 255 levels, but you are also sampling four pixels down to one, so it can use the differences between the four different 255 samples and put that fractional measurement into the 4096 storage space. In a grand simplification, if one pixel reads a value of 50 and one next to it reads 51, the new measure becomes 50.5, which is more precise. Imagine a blue sky feathering slight color and density shifts and yes, this level of extra color information is enough to significantly improve banding issues.
Click to expand...
Got in a bit of a debate about this over on dvinfo. On an absolutely mathematically linear gradient, using the simplest and lamest downsampling (nearest neighbor or simple decimation), it's true that there's practically zero benefit. But in reality, there can be substantial benefit. In reality, you will never encounter that mathematically linear gradient, so the variations Mitch mentions are going to be there, and the downconversion can take advantage of them. And if you use a sophisticated downconversion you can greatly, hugely reduce the banding that happens in 8-bit. Even a modest downconversion method such as bilinear can provide a notable improvement, and bicubic can do a much better job.

To explore it, I created an 8-bit gradient in UHD resolution in Photoshop. I then downconverted it using "nearest neighbor" to 1080 16-bit, and saw pretty much identical banding. I then downconverted it to 1080 using bilinear, and the banding was substantially improved. The original and downconverted are in the following PSD file.

http://fiftv.com/Gradient/Gradient-8-bit-FHD-vs-UHD-downconverted.psd

It's a PSD with two layers, so you can toggle the top layer on/off to see the difference in banding. So in this file, what you'll see are a true 8-bit 1920x1080 gradient as rendered by Photoshop, and a 3840x2160 8-bit Photoshop gradient that's been downconverted to 1920x1080, using simple bilinear downconversion. Bicubic would provide even better results.

My conclusion is that downconverting 8-bit UHD to 10-bit FHD can indeed result in retaining bit depth differences and making for smoother transitions, more shades, and better gradients, but it requries using a sophisticated downconversion method.

If you want to experiment with the source gradient (8-bit UHD), it's here.
 
Last edited:
I received my video assist, i took all test with gh4. Shoot an static shot of newspaper with tripod lock.
1.HD internal.
2.HD on Video Assit.
3.4k internal, downconverted for HD timeline.
4.4k downconverted from gh4 to VA.
All this test was on 422 10bit, without sd card on Gh4, to make sure no loss in signal.
Results.
I found 4k 10bit downcoverted to VA looks much much better, lot to detail. The text which i couldnt read on HD, was easily read on 4k downconverted, although there crop factor, but very less moire.
I also noticed shift in color external & internal with manual WB. Although downconverting from 4k to hd in post looks great but external downconverting to 10bit helps lot in color grading.
Thax
 
estarkey7 said:
Using what downsampling algorithm?
I agree you can get better color accuracy downsampling 4K 4:2:0 to 1080p 4:2:2, but as soon as you try and grade the resulting downsampled image, it's true 8-bit heritage will undoubtedly show. Nothing is going to take an 8-bit color channel's 255 and make it have 4096 levels.
Click to expand...
Here's a link to an overview of image scaling algorithms:

https://en.wikipedia.org/wiki/Image_scaling

Image sampling is a multi-dimensional measurement process, and while the mathematics are complex, it's not mysterious. When downsampling by half, the 2-bit loss in spatial resolution can literally be converted into a 2-bit increase in sample-depth, without significant data loss. With an H.264 recording, image data is converted into YUV format, with the UV chroma data downsampled at half the spatial resolution of the Y luma data. An H.264-specific UHD-to-1080p converter could create a 4:4:4 recording by preserving the UV channels intact, and combining them with a downsampled Y channel. The result would be a 4:4:4 recording with a 10-bit Y channel and full-resolution 8-bit UV channels, with virtually zero loss of image data. Alternately, you could downsample the UV channels as well, producing a 4:2:0 recording with half-resolution 10-bit UV channels.

These conversions work because in mathematics, there's no such concept as "true 8-bit heritage". Even with a non-optimal summing filter, you can produce a half-resolution 1024-step (10-bit) sample out of a 2x2 block of 256-step (8-bit) samples simply by adding the four samples together (and not dividing by 4 to produce an 8-bit average). In practice, more sophisticated downsampling techniques are used to preserve edge sharpness while minimizing aliasing artifacts.
 
Last edited:
In theory then couldn't taking 4k at 10bits and down converting give you 2k at 12 bits? Same concept but even better precision.

While 10bits is a heck of a lot better than 8 bits we cannot expect 10 bits to keep holding up well the more we increase resolution. 10bits was considered professional all the way back to SD. With HD and beyond to 4k that precision is not as fine as it once was when we dealt with SD material.

Perhaps this is one reason why 4k 10bit down converted in an external recorder can appear a little bit better on the Convergent-Design devices. You are in essence getting a down converted 12 bit sample before recording to ProRes. When the GH4 handles the down conversion it gets truncated to 10 bits since thats all the HDMI or HD-SDI can handle. Too bad none of the Convergent-Design devices support the ProRes 4444 XQ codec. That codec would keep the 4:4:4 color and 12 bits from their down conversion which would be the next best thing to RAW 2k.
 
Thomas Smet said:
In theory then couldn't taking 4k at 10bits and down converting give you 2k at 12 bits? Same concept but even better precision.

While 10bits is a heck of a lot better than 8 bits we cannot expect 10 bits to keep holding up well the more we increase resolution.
Click to expand...
Yes, you could produce a 12-bit 1080p conversion from the 10-bit 4K output of the GH4. However, noise becomes more of an issue with each increase in bit-depth. So long as you've got enough light to fill the image sensor bins with photons, 10 bits will produce comparable results regardless of the spatial resolution (assuming diffraction remains minimal).

Where bit-depth becomes crucial is in capturing a wide dynamic range. Ideally, you'd want to have one bit of digitization for each stop of DR, but with H.264 encoding techniques, a 12-stop dynamic range can fit pretty well into an 8-bit codec. With a 10-bit encoder, a 14-stop DR is reasonable.
 
Thomas Smet said:
Perhaps this is one reason why 4k 10bit down converted in an external recorder can appear a little bit better on the Convergent-Design devices. You are in essence getting a down converted 12 bit sample before recording to ProRes.
Click to expand...
Wouldn't that only work if the signal was downconverted in a 16 bit mode or higher? I couldn't find any information suggesting that is happening. Could be wrong though.
 
bunk said:
Wouldn't that only work if the signal was downconverted in a 16 bit mode or higher? I couldn't find any information suggesting that is happening. Could be wrong though.
Click to expand...
You cannot find information because we do not publish it one way or another. Secret sauce.
 
You must log in or register to reply here.
Back
Top