Blackmagic Forum

The results are pretty nice, but it's extremely slow :/ Even with 720p videos.

It would be cool to have this on part with what Topaz Labs Video Enhance AI can do now.
They optimized it a lot, and it's so much faster than before (on top of removing the compression artifacts).

Videoneth wrote:The results are pretty nice, but it's extremely slow :/ Even with 720p videos.

It would be cool to have this on part with what Topaz Labs Video Enhance AI can do now.
They optimized it a lot, and it's so much faster than before (on top of removing the compression artifacts).

Two very different tools, different management of Gpu. If you want more speed you should render out clip (I hate when I wrote it, but for something slow should be like this).
The guys of Veia did a lots of work and a lots of training, between photo and videos, it’s a lots of effort and time, I don’t know if Blackmagic Design guys can do that at same time with DR dev.
We could ask an OFX from Veia team like they did for Ph plug-ins.


Inviato dal mio iPhone utilizzando Tapatalk

I'm aware, but the goal of these tools are pretty similar
I only used Topaz's tool as comparison because they are the only other serious company (imo) doing it too.

In this case, the training as not much to do with how the models are used by the GPU, since they are already trained.

I just hope that the new x2 Superscale will be worked on because that's something I'm using constantly (the old 2x superscale, it's "okay", but better than not activating it on my clips).

But we can't go fare with renders at 2-3 fps on a RTX 3090. I don't wan't to see this thing just rot like the face detection tool for example (I wrote about it in the past, it could be very useful, but it doesn't work well at all).

2-3 fps is certainly slow. If I apply 2x Super Scale to all my 25fps Pro Res 422 HD clips in a project after grading, which takes no time at all, and change the timeline to UHD, my playback speed slows to 13 fps. Renders to h265 UHD using the Nvidia encoder run at a similar speed. This I put down to my hardware.
Even with 50 fps MP4 clips I see the affect of applying 2x Super Scale almost immediately.

My own brief test on 1080 clips showed zero improvement over the original 2x method, but processing slowed from real time to a hair-pulling 2 fps with my 12 GB RTX 3060.

In my last post I was referring to the original Super Scale. As to the Enhanced version I absolutely agree as my findings are the same.

I had very pleasing tests with new algorithm.

1080p 10bit Sony S-Log

2x Enhanced + Color correction

Render in place -> shows 7fps on my MacBook M1 Pro Max

And I can see significant improvement over original 2x Super Scale in terms of quality


I tried comparing to Topaz Video Enhance AI and it's a mixed bag of results, Topaz tends to change colors and on some less common objects completely mush them, however some parts of the image looks stunningly good compared to DR 2x Super Scale.

Hard Is Easy wrote:Render in place -> shows 7fps on my MacBook M1 Pro Max

Yeah that's better than with my rtx 3090 at 2-3 fps lol :/

I wish I could use it because the results with the new 2x are way better than with the original SuperScale, at least for the type of videos I'm using.

I have also been experimenting with the new Enhanced 2X SS.

In side by side comparisons I can persuade myself that I see an improvement over the Standard 2X SS.

But it is very slow for me. Typically I get 10fps using Enhanced SS, or 50fps using Standard, when upscaling 720x480 to 1280x720, with my 3080 video card. (Speeds vary according to subject matter, size, etc. ).

This makes it very difficult to justify usage.

Also I don't see the relevance of comparisons being made here ( and elsewhere across the Resolve Board) to Topaz Video Upscaler. This software is not in any way a NLE or a Coloriser. It only upscales (so it would be an extra process, if used), and it has always been very slow.

I understand that Topaz have been making great efforts to speed it up, but a search online could only turn up a detailed review from 1 year ago that reported:

"If you use NVIDIA GTX 1080, zooming in from 480P SD to 1080P HD MP4 H.264 will take about 0.5 seconds per frame, which is about 15 times the length of the original video. To zoom in from HD to 4K, each frame usually takes 1 to 2 seconds, which is about 50 times the length of the original video."

OK, a GTX1080 is an old video card, but 0.5fps for 480p input is unfeasibly slow.

I hope that DaVinci can improve the speed of the Enhanced 2X SS process.

IB

Ivor_Bigunn wrote:I understand that Topaz have been making great efforts to speed it up, but a search online could only turn up a detailed review from 1 year ago that reported:

"If you use NVIDIA GTX 1080, zooming in from 480P SD to 1080P HD MP4 H.264 will take about 0.5 seconds per frame, which is about 15 times the length of the original video. To zoom in from HD to 4K, each frame usually takes 1 to 2 seconds, which is about 50 times the length of the original video."

OK, a GTX1080 is an old video card, but 0.5fps for 480p input is unfeasibly slow.

I hope that DaVinci can improve the speed of the Enhanced 2X SS process.

IB

Going from 480P to 1080P is running about 20-35 fps for me on the latest version, with my 3080Ti.
That makes the Topaz solution very usable.

I second the improvement in the latest Topaz. It is now much faster than the previous version. This is from the update from this week.

Super scale 2x Enhanced plays at about 6fps without caching enabled. Renders fairly quickly with Smart Cache enabled. Mac Studio Max 32GB

Thanks for the updates about Topaz Upscale.

IB

Peter Cave wrote:Super scale 2x Enhanced plays at about 6fps without caching enabled. Renders fairly quickly with Smart Cache enabled. Mac Studio Max 32GB

6 fps is not "fairly quickly" lol.
At the end, the cache had to be rendered too, and any changes will make it do it again, so there is no time gained.

6 fps for a 15 mins video at 50 fps (what I generally edit) is too slow, it's not usable.
I don't complain because it's a beta, and I'm sure they will work on it, if it's ends up with another thing that will never get updated, and usable only for very very short videos.

DavidVogt wrote:I second the improvement in the latest Topaz. It is now much faster than the previous version. This is from the update from this week.

Even the previous version were faster, even if someone add interpolation (30 to 50 fps for example) +upscale + Enhancement.

I can't imagine Optical flow (speed warp) + the new 2x

I also looked at Topaz. But just workflow wise it's muuuuuuuch easier to stay inside resolve for me. Apart from speed, how do you evaluate upscale quality between Topaz and Davinci?

There is also a new upscaler OFX plugin from RevisionFX called REZup.

Just tried the new superscale 2x in resolve...weird half-frame upscale. anybody else having this?

Robert Niessner wrote:There is also a new upscaler OFX plugin from RevisionFX called REZup.

165 USD?!?!? That's mad

Mads Johansen wrote:

Robert Niessner wrote:There is also a new upscaler OFX plugin from RevisionFX called REZup.

165 USD?!?!? That's mad

The one from Nucoda/Filmworkz is more costly.

I have now played around with the demo of the REZup OFX plugin from RevisionFX.

Boy, is that slow - I can do a 3x superscale from SD 576p50i to 1080p50 in Resolve 17.4 in realtime including Neural Engine deinterlacing with my RTX A5000 - but with the RevisionFX plugin if I switch to GPU it is in fact slower than using the CPU which is rendering at 2 fps.

And it does not do much - I can't get any good results and it seems to be far from doing any AI enhancement magic. De-ringing also seems to do not much. Maybe the PAL SD stuff I am feeding is too low quality for that.

Using the new Super Scale 2x Enhanced with default 0.5 sharpen and 0.5 noise reduction on a 47’ graded video. DaVinci Resolve 18.5b4. The original footage (2K CinemaDNG and 4K BRAW Q5) is on a 1920x804 timeline as is. My h.265 PQ output is defined as 3840x1608. Render speed is 8 fps on my Mac Pro 2019 48 GB. Render required just over 3 hours.

I feel the memory pressure (yellow in Activity Monitor) as I’m using almost no CPU but using 40 of a maximum 48 GB of memory with swap and compressed memory. I really need to add 192 GB RAM. Using 80% of my Radeon Pro Vega II 32 GB GPU.

I have had success with two previous renders like this (PQ and HLG) for 2 hours of video. Patience required.

Robert Niessner wrote:
And it does not do much - I can't get any good results and it seems to be far from doing any AI enhancement magic. De-ringing also seems to do not much. Maybe the PAL SD stuff I am feeding is too low quality for that.

If you still have a Topaz trial/paid have a look at the new iris upscaler with face refinement. It's not meant for high quality camera footage, more like typical low bitrate internet video. I"m sure there's a lot to fix with it but for first public release quite impressive. Also unlike the new Enhanced Resolve 2x upscale which is 100% GPU, it puts a high load on CPU and lower use of GPU so ideal for many. For people who want speed though, this is also not for them, probably the slowest Topaz upscale model, although I"m mostly at 100% CPU so i'm bottlenecked

I can live with a 2x Enhanced upscale requiring almost four hours per one hour of video. My longest videos are typically up to 2.5 hours runtime so that could mean up to 10 hours for the render. So if I start the job in the late evening, it’s ready by early morning. I know that may be intolerable for those with great time pressures, but I’m fortunate not to have those concerns.

Deliverables in upscaled 4K h.265 only take 30% more storage space than 2K h.264 so that penalty to deliver twice the resolution is worth the effort.

The reason I’m upscaling in the first place is that my camera A is usually using the B4 Fujinon 20x7.8BRM zoom so my maximum capture resolution is 2K. 2x Enhanced is a wonderful ‘free’ option to go to quality 4K deliverables using that lens compared to spending $12,000 on a 4K Fujinon 20-120mm PL mount Super35 lens since I never seem able to afford that.

Thank you to the Resolve team at BMD.

Here's a free upscaler (Waifu2x) if you're interested in trying (I've have had good results with tweaked settings & engine choice depending on source). Yea, the GUI is a little dated & scattered, but it does it's job.

Waifu2x-Extension-GUI

Photo/Video/GIF enlargement and Video frame interpolation using machine learning
Supports AMD / Nvidia / Intel GPU

https://github.com/AaronFeng753/Waifu2x-Extension-GUI


Oh, another decent free one for Still Images is Upscayl:

https://github.com/upscayl/upscayl

Sarasota wrote:Here's a free upscaler (Waifu2x) if you're interested in trying (I've have had good results with tweaked settings & engine choice depending on source). Yea, the GUI is a little dated & scattered, but it does it's job.

Waifu2x-Extension-GUI

Photo/Video/GIF enlargement and Video frame interpolation using machine learning
Supports AMD / Nvidia / Intel GPU

https://github.com/AaronFeng753/Waifu2x-Extension-GUI


Oh, another decent free one for Still Images is Upscayl:

https://github.com/upscayl/upscayl

That guy's popup 'Nags' accompanied by audio nags to give him money is next level annoying considering it's doing it minutes after I installed. On the plus side, he's extremely active with the releases

CougerJoe wrote:

Sarasota wrote:Here's a free upscaler (Waifu2x) if you're interested in trying (I've have had good results with tweaked settings & engine choice depending on source). Yea, the GUI is a little dated & scattered, but it does it's job.

Waifu2x-Extension-GUI

Photo/Video/GIF enlargement and Video frame interpolation using machine learning
Supports AMD / Nvidia / Intel GPU

https://github.com/AaronFeng753/Waifu2x-Extension-GUI


Oh, another decent free one for Still Images is Upscayl:

https://github.com/upscayl/upscayl

That guy's popup 'Nags' accompanied by audio nags to give him money is next level annoying considering it's doing it minutes after I installed. On the plus side, he's extremely active with the releases

Agree... That nag screen is a bit much (cough, ok - next level). I gave it a whirl and it'll have it's uses over here (I signed up / donated to kill the popup, plus he earned it in my books). You can disable the notification sound in the "additional settings" (right column).

Sarasota wrote:

CougerJoe wrote:

Sarasota wrote:Here's a free upscaler (Waifu2x) if you're interested in trying (I've have had good results with tweaked settings & engine choice depending on source). Yea, the GUI is a little dated & scattered, but it does it's job.

Waifu2x-Extension-GUI

Photo/Video/GIF enlargement and Video frame interpolation using machine learning
Supports AMD / Nvidia / Intel GPU

https://github.com/AaronFeng753/Waifu2x-Extension-GUI


Oh, another decent free one for Still Images is Upscayl:

https://github.com/upscayl/upscayl

That guy's popup 'Nags' accompanied by audio nags to give him money is next level annoying considering it's doing it minutes after I installed. On the plus side, he's extremely active with the releases

Agree... That nag screen is a bit much (cough, ok - next level). I gave it a whirl and it'll have it's uses over here (I signed up / donated to kill the popup, plus he earned it in my books). You can disable the notification sound in the "additional settings" (right column).

So, Upscayl is just a frontend for https://github.com/xinntao/Real-ESRGAN- ... tag/v0.2.0

Can someone remind me to do the things https://videoprocessing.ai/benchmarks/v ... alers.html wants, in order to get datapoints on how DR is doing?

Mads Johansen wrote:So, Upscayl is just a frontend for https://github.com/xinntao/Real-ESRGAN- ... tag/v0.2.0

Can someone remind me to do the things https://videoprocessing.ai/benchmarks/v ... alers.html wants, in order to get datapoints on how DR is doing?

Ooooh, thank you for that link (much appreciated).

The strange thing I noticed is that if I render the project, with even only one clip having SS-2x-Enhanced enabled, the entire project's render time goes out the window (just over 1 minute -> 45-55 minutes) and DRS is unresponsive to requests to stop rendering.
However, if I Render-In-Place that one clip, the processing time is, say 2 minutes, then project render time is back to a minute.

Seems like something odd is going on here.