Canon C100 8bit 4.2.0 AVCHD Footage and Upscaling (Topaz/Avisynth)

[dogmydog]

Hey guys,

Just wanted to share my experience on this forum and check your thoughts about it.
I'm not an engineer, though I love video and have some technical experience thanks to this amazing forum.

I have acquired a Canon C100 last year and I'm amazed by the image quality.
It is a 4k sensor downsampling a beautiful 1080p c-logged on a MPEG-4 AVC/H.264 24 Mbps LPCM VBR (4:2:0)

My only two complaints:

1. The 4.2.0 8bit color codec that, when I try to color grade, starts to fall apart.
2. No 4k.

Then I researched online about AI resizing and I found Topaz Video Enhance AI (I swear I haven't read the topic about it back then )
I have converted my recorded AVCHD to PRORES 422 4k and I think the results are better than simply grading the original MTS file.

I understand that there is no magical way to improve footage, but should I always convert the compressed 4.2.0 8 bit 1080p MTS files to 4.2.2 10bit PRORES 4k to grade?
Even if I use avisynth instead of Topaz? Or is it just jibberish?

Here are some samples for your appreaciation:

original MPEG-4 AVC/H.264 24 Mbps LPCM VBR (4:2:0)

[Attachment 59129 - Click to enlarge]
Original video file
https://drive.google.com/file/d/1A6rOKOdi_jzgsxnsfH8CU5fbQNOdyR-I/view?usp=sharing
Curves and scopes

[Attachment 59130 - Click to enlarge]

upscaled PRORES 422

[Attachment 59131 - Click to enlarge]
Original video file
https://drive.google.com/file/d/1A03Dch5LiNmocPuahq535zAHpF3Qr_JI/view?usp=sharing
Curves and scopes

[Attachment 59132 - Click to enlarge]

DxnHD 422 8bit files recorded on a external Ninja Inferno monitor

[Attachment 59135 - Click to enlarge]
Original video file
https://drive.google.com/file/d/1C9aa37NrodKFyj2fgcDs1Fj7pCJ5YYmC/view?usp=sharing
Curves and scopes

[Attachment 59134 - Click to enlarge]

Cheers from rio!

[poisondeathray]

The 4.2.0 8bit color codec that, when I try to color grade, starts to fall apart.

Did you mean in terms of artifacts ? What is "falling apart" ?

You're denoising it by using prores, but losing some signal detail as a trade off. Depending on which model you used in Topaz, that might apply some denoising as well. People convert to Prores because it's easier to edit. If the noise is limiting your ability to grade, then denoising might be beneficial for your specific workflow. The quality is techically worse.

On the other hand, if you use a dedicated denoiser workflow and lossless intermediates - you might be able to target the noise better and retain more details

You provided lossy jpg's which add additional artifacts, but if you look closely you can tell the ninja retains the most details, and the most noise (esp. chroma noise) . The range is undetermined, because you provided a jpg, not the native ninja stream. There is clipping in the hair, but that is likely from the jpg conversion. Also , in the lumetri view you have clamp signal, so you cannot tell if there are additional details in the actual stream

The color is shifted (709/601 mismatch) on the upscale - you probably just used the mts directly in Topaz, and the RGB conversion was done incorrectly. Topaz works internally in RGB, and that is potentially a problem for clipping. It does not matter that you export in prores, the clipping occurs at the RGB conversion stage before prores. To get ideal results in Topaz (or anything for that matter) you generally need to pre and post process correctly. Not only is color shifted, but the highlights are clipped in the RGB conversion - hair detail is lost and not recoverable. The correct way to do this is either legalize the range first, or use a full range conversion to RGB if using Topaz.

It's generally a mistake to clip usable data so early, and that definitely limits your grading in other ways than noise - your dynamic range is limited. Colorists generally like to have all the data so they can manipulate it in the way they see fit, instead of discarding some right off the bat

This is an apng, it should animate in most browsers, but you can see the hair detail clipping

[dogmydog]

Dear @poisondeathray, thanks for the complete answer!
I'm learning a lot from it.

I have also linked the video files to compare.
They are the google drive links between the screenshot and the curves (sorry about ticking the Clamp Signal, I missed the checkbox)

MTS
https://drive.google.com/file/d/1A6rOKOdi_jzgsxnsfH8CU5fbQNOdyR-I/view?usp=sharing

Upscaled Topaz
https://drive.google.com/file/d/1A03Dch5LiNmocPuahq535zAHpF3Qr_JI/view?usp=sharing

Ninja DxHD
https://drive.google.com/file/d/1C9aa37NrodKFyj2fgcDs1Fj7pCJ5YYmC/view?usp=sharing

Yes, I simply dragged the MTS into the Topaz.
I'm going to do the correct RGB conversion next time.

The 4.2.0 8bit color codec that, when I try to color grade, starts to fall apart.

Did you mean in terms of artifacts ? What is "falling apart" ?

You're denoising it by using prores, but losing some signal detail as a trade off. Depending on which model you used in Topaz, that might apply some denoising as well. People convert to Prores because it's easier to edit. If the noise is limiting your ability to grade, then denoising might be beneficial for your specific workflow. The quality is techically worse.

On the other hand, if you use a dedicated denoiser workflow and lossless intermediates - you might be able to target the noise better and retain more details

You provided lossy jpg's which add additional artifacts, but if you look closely you can tell the ninja retains the most details, and the most noise (esp. chroma noise) . The range is undetermined, because you provided a jpg, not the native ninja stream. There is clipping in the hair, but that is likely from the jpg conversion. Also , in the lumetri view you have clamp signal, so you cannot tell if there are additional details in the actual stream

The color is shifted (709/601 mismatch) on the upscale - you probably just used the mts directly in Topaz, and the RGB conversion was done incorrectly. Topaz works internally in RGB, and that is potentially a problem for clipping. It does not matter that you export in prores, the clipping occurs at the RGB conversion stage before prores. To get ideal results in Topaz (or anything for that matter) you generally need to pre and post process correctly. Not only is color shifted, but the highlights are clipped in the RGB conversion - hair detail is lost and not recoverable. The correct way to do this is either legalize the range first, or use a full range conversion to RGB if using Topaz.

It's generally a mistake to clip usable data so early, and that definitely limits your grading in other ways than noise - your dynamic range is limited. Colorists generally like to have all the data so they can manipulate it in the way they see fit, instead of discarding some right off the bat

This is an apng, it should animate in most browsers, but you can see the hair detail clipping

[dogmydog]

The 4.2.0 8bit color codec that, when I try to color grade, starts to fall apart.

Did you mean in terms of artifacts ? What is "falling apart" ?

By falling apart I mean that when I start to grade on Premiere Pro, applying LUTs, etc, I notice the introduction of blocking and loss of detail.
It seems that the detail and quality of the original MTS starts to look compressed and with less resolution.

Here is another example
https://drive.google.com/file/d/15YDJVk6omfhKoBWBudbNU3JWthYXUHUZ/view?usp=sharing

On the other hand, if you use a dedicated denoiser workflow and lossless intermediates - you might be able to target the noise better and retain more details

Sorry for my newbie question, but what do you mean by this?
What would be a dedicated denoiser workflow? I do not use proxy and always work with the original MTS files straight into premiere pro for grading and editing.

Thanks for the support!!

[poisondeathray]

On the other hand, if you use a dedicated denoiser workflow and lossless intermediates - you might be able to target the noise better and retain more details

Sorry for my newbie question, but what do you mean by this?
What would be a dedicated denoiser workflow? I do not use proxy and always work with the original MTS files straight into premiere pro for grading and editing.

I mean use dedicated denoiser such as neat video, or preprocess in avisynth etc... They will probably do a better job than relying on a codec such as prores to denoise and deblock, because that' s their specific job. But some of the models in Topaz might do a good job too with denoising, it depends on what model was used and how the model was trained. It can be a delicate balance between over denoising and losing fine details

The ninja is the closest to the "original" signal, and it keeps the highlights. The tradeoff for retaining more details, is retaining more of the sensor noise.

[dogmydog]

On the other hand, if you use a dedicated denoiser workflow and lossless intermediates - you might be able to target the noise better and retain more details

Sorry for my newbie question, but what do you mean by this?
What would be a dedicated denoiser workflow? I do not use proxy and always work with the original MTS files straight into premiere pro for grading and editing.

I mean use dedicated denoiser such as neat video, or preprocess in avisynth etc... They will probably do a better job than relying on a codec such as prores to denoise and deblock, because that' s their specific job. But some of the models in Topaz might do a good job too with denoising, it depends on what model was used and how the model was trained. It can be a delicate balance between over denoising and losing fine details

The ninja is the closest to the "original" signal, and it keeps the highlights. The tradeoff for retaining more details, is retaining more of the sensor noise.

Fantastic! Thank you for this.
If it's not too much trouble, can you please recommend a guide on how I can retain/improve my 4.2.0 8bit footage to grade?
My whole point here is to get better results grading the MPEG-4 AVC/H.264 24 Mbps LPCM VBR (4:2:0) recorded in camera without exporting a mushy h264 mp4 at the end.

[dogmydog]

Hey guys,

Sorry to bring this topic back up, but I still wonder if there is anyway of improving the Canon C100 8 bit c-log native AVCHD format to something else using Avisynth or any other software based method in post, so I don't lose much information on the color grading process.

Here are again the original files:
MTS (8bit 4.2.0 recorded in camera)
https://drive.google.com/file/d/1A6rOKOdi_jzgsxnsfH8CU5fbQNOdyR-I/view?usp=sharing

Upscaled Topaz (MTS upscaled to 4k Prores)
https://drive.google.com/file/d/1A03Dch5LiNmocPuahq535zAHpF3Qr_JI/view?usp=sharing

Ninja DxHD (8bit 4.2.2 externally recorded)
https://drive.google.com/file/d/1C9aa37NrodKFyj2fgcDs1Fj7pCJ5YYmC/view?usp=sharing

I understand that converting to PRORES to colorgrade is bad because I'm losing information right on the start.
However, if I simply drag the MTS files into Premiere Pro and start color correcting, with LUTs etc, I start to introduce blocking and loose resolution.

Can anyone help me with more insights?
Thanks again,

[Selur]

a. Have you thought about using some lossless format instead of ProRes? (I guess Premiere Pro should support some lossless formats.)
b. Have you tried other resizers than the one from Topaz? (nnedi3, Waifu2x with upconv_7_photo as model, SRMD, RealSR with DF2K_JPEG as model, AiUpscale or some GLS-Resizer are the first things that come to mind)
c. I don't get why color changes should introduce blocking, this should only happen if you export sot some lossy format with a too low bitrate,.. (sure color clipping, gamma changes can remove details, but introduce blocking?)

Cu Selur

[dogmydog]

a. Have you thought about using some lossless format instead of ProRes? (I guess Premiere Pro should support some lossless formats.)
b. Have you tried other resizers than the one from Topaz? (nnedi3, Waifu2x with upconv_7_photo as model, SRMD, RealSR with DF2K_JPEG as model, AiUpscale or some GLS-Resizer are the first things that come to mind)
c. I don't get why color changes should introduce blocking, this should only happen if you export sot some lossy format with a too low bitrate,.. (sure color clipping, gamma changes can remove details, but introduce blocking?)

Cu Selur

Dear Selur, thanks for the quick response!

a. I have thought about DnxHD, since it is windows friendly for a longer time. Would that help me not losing details on the early conversion?
b. Thanks for the examples, I will definately look them up! But for now, resizing is not my main goal, the main goal is being able to apply LUTs, colorgrade without my exports looking like this (her face details look worse then the rest, they look mushy and not like the background):
https://drive.google.com/file/d/15YDJVk6omfhKoBWBudbNU3JWthYXUHUZ/view?usp=sharing

When I shooted 14bit magic lantern raw video on my Canon 50D I was able to grade all I wanted and when I exported to the final format (h264 8bit 420 5mbps), the end result would looke cleaner than the one I just linked.
c. Yes, my question is: if I color correct/grade my MTS file, do I lose more details, colors get more clipped, etc. than if I convert to an lossless intermediary codec such as DnxHD for example?

Thank you for the patience and sorry if these are all newbie questions

[Selur]

I have thought about DnxHD, since it is windows friendly for a longer time. Would that help me not losing details on the early conversion?

Last I checked DNxHD was not lossless and it doesn't look like that has changed:

Avid DNxHD ("Digital Nonlinear Extensible High Definition") is a lossy high-definition video post-production codec ...

see: https://en.wikipedia.org/wiki/Avid_DNxHD

her face details look worse then the rest, they look mushy and not like the background

a. not sure what the video looked like before compression but when upscaling some constrast sharpening might help and when downscaling dpid might be worth a look.
b. seems more like a bitrate issue on encoding, adjusting adaptive quantization might help a bit, but other than that you would need more bit rate to keep the detail level. This is easy to check, if you export to a lossless format and the details are not then it's not a problem of the encoder. If the lossless output has the details, then it's a problem with the encoder, encoder settings. It might help to lightly remove some grain etc. to lower the amount of details in a controlled fashion to help the encoder with its job trying to preserve details.

if I color correct/grade my MTS file, do I lose more details, colors get more clipped, etc. than if I convert to an lossless intermediary codec such as DnxHD for example?

No. If you load the source color grade and export (with encoder&settings X) or load the source export to some lossless format and then encode that output (with the same encoder&settings X) the output should be identical. Intermediate steps are okay as long as they are lossless.
If converting from a lossless to a lossy format removed to many details only options are:
a. try better suited output format in case that is possible
b. try another (better) encoder and/or adjust the encoder settings
c. remove details in a controlled way to help the encoder (combined with b.)

Cu Selur

[dogmydog]

I have thought about DnxHD, since it is windows friendly for a longer time. Would that help me not losing details on the early conversion?

Last I checked DNxHD was not lossless and it doesn't look like that has changed:

Avid DNxHD ("Digital Nonlinear Extensible High Definition") is a lossy high-definition video post-production codec ...

see: https://en.wikipedia.org/wiki/Avid_DNxHD

her face details look worse then the rest, they look mushy and not like the background

a. not sure what the video looked like before compression but when upscaling some constrast sharpening might help and when downscaling dpid might be worth a look.
b. seems more like a bitrate issue on encoding, adjusting adaptive quantization might help a bit, but other than that you would need more bit rate to keep the detail level. This is easy to check, if you export to a lossless format and the details are not then it's not a problem of the encoder. If the lossless output has the details, then it's a problem with the encoder, encoder settings. It might help to lightly remove some grain etc. to lower the amount of details in a controlled fashion to help the encoder with its job trying to preserve details.

if I color correct/grade my MTS file, do I lose more details, colors get more clipped, etc. than if I convert to an lossless intermediary codec such as DnxHD for example?

No. If you load the source color grade and export (with encoder&settings X) or load the source export to some lossless format and then encode that output (with the same encoder&settings X) the output should be identical. Intermediate steps are okay as long as they are lossless.
If converting from a lossless to a lossy format removed to many details only options are:
a. try better suited output format in case that is possible
b. try another (better) encoder and/or adjust the encoder settings
c. remove details in a controlled way to help the encoder (combined with b.)

Cu Selur

This is VERY insightful, I really appreciate the time spent on replying to this topic.
I'm going to keep studying but your answer has clearly widen my perspectives. Thank you