Ultra NR is worse than older models?

It seems from testing that UltraNR is worse than previous spatial NR options such as Better or Enhanced. I find that a mix of temporal with spatial set to better and half the values of temporal fares better in visual quality. Has anyone found similar or differing results?

I've found UltraNR to be noticeably better than the other NR options.

Though admittedly, I wasn't doing any kind of proper comparison. I just used the Analyze feature and was impressed with the results compared to what I was doing before.

I've found UltraNR to be noticeably better than the other NR options.

Though admittedly, I wasn't doing any kind of proper comparison. I just used the Analyze feature and was impressed with the results compared to what I was doing before.

Ah, for me I've been testing it on 204000 ISO video files to see if it's better with extreme noise but it seems the old ones are. I will try it with moderate noise and see if it fairs better.

It's not better than old settings.

[Rohit Gupta]

It seems from testing that UltraNR is worse than previous spatial NR options such as Better or Enhanced. I find that a mix of temporal with spatial set to better and half the values of temporal fares better in visual quality. Has anyone found similar or differing results?

If you can share your clip, we can take a look.

Ah, for me I've been testing it on 204000 ISO video files to see if it's better with extreme noise...

The very concept of 204,000 ISO frightens me to death.

Hey

Since I upgraded my OS and Resolve as well, I had to test a whole bunch of things up to now, but this is what I'm going to check very soon. If you're able to keep this thread alive, I'll post my 2 cts here soon

Ah, for me I've been testing it on 204000 ISO video files to see if it's better with extreme noise...

The very concept of 204,000 ISO frightens me to death.

Completely understandable from a professional point of view. I just love pushing cameras to their limits and seeing what usable footage I can get out of them.

204800 was an ISO I was surprised I could get anything out of and it works in a pinch for some YouTube music videos depending on topic.

It's not better than old settings.

Yup... I have found it goes as follows:

Medium noise:
Temporal 5 frame + Spatial enhanced > UltraNR > Spatial Faster

Heavy noise:
Temporal 5 frame + Spatial better > Spatial enhanced > UltraNR > Spatial faster

The only one worse I've found is spatial faster. Similar effect could be done by just blurring color and luma channels.

It seems from testing that UltraNR is worse than previous spatial NR options such as Better or Enhanced. I find that a mix of temporal with spatial set to better and half the values of temporal fares better in visual quality. Has anyone found similar or differing results?

If you can share your clip, we can take a look.

https://drive.google.com/file/d/1euQ-EVfw1JrUeqLKVo7CziVJ3B7DkUVx/view?usp=sharing
Take a look and see what works! I dub this clip the AI killer because its heavy noise also tricks Topaz AI's Nyx and Proteus to smear the image.

I've found the best for this specific image is:
Temporal:
Frames: 5
Motion: small
Luma: 50 (anything above looks like water)
Chroma: 40
Motion: (default)
Blend: 0

Spatial:
Mode: better
Luma: 25
Chroma: 25 (hard linked)
Blend: 0

It's a flat image so just use R709 color space as that is what Cinelike D2 on Lumix cameras is closest to.

'"Ah, for me I've been testing it on 204000 ISO video files..."

204,000??!! That's not video, that's output from secret-squirrel CIA cameras that can read license plates from 40 miles up and see through walls into your bedroom during a blizzard. At night. With no interior lights.

And to think I got excited in the 80's when Kodak released ASA/ISO 1000 film.

I doubt such a test is making sense. Cameras that can generate a picture at that value (like some Sonys) are doing heavy internal processing.
Rather use a camera without any processing at 3200 or 6400.

I doubt such a test is making sense. Cameras that can generate a picture at that value (like some Sonys) are doing heavy internal processing.
Rather use a camera without any processing at 3200 or 6400.

I have video NR turned all the way off in that shot within camera settings (something Sony's under the FX6 can't do). The only processing going on is the compression into the H264 10-bit 4:2:2 codec.

Ah, for me I've been testing it on 204000 ISO video files to see if it's better with extreme noise...

The very concept of 204,000 ISO frightens me to death.

AAARGH!!!! My eyes, my eyes!

'"Ah, for me I've been testing it on 204000 ISO video files..."
204,000??!! That's not video, that's output from secret-squirrel CIA cameras that can read license plates from 40 miles up...

And for those who don't know what that is...


I agree, 204,000 ISO is pretty excessive and crazy. I can understand 2000-5000 ISO, which is actually used in the real world.

Ah, for me I've been testing it on 204000 ISO video files to see if it's better with extreme noise...

The very concept of 204,000 ISO frightens me to death.

+10000

I'm more thinking for high ISO underexposed footage which I've had to deal with before when color correcting. Someone shot at something like 12,800ISO 2 stops under in S-Log3. Green cast all over and built in NR throwing off Resolve's NR.

From what I can see, usually Ultra NR does a better job with same settings, and when settings are changed to get something similar, it mushes the image more with "enhanced" mode. For spatial NR.

Manual claims: "UltraNR is an AI-based noise reduction option that provides intelligently targeted noise reduction based on the machine learning of real world video noise patterns, rather than relying on a specific mathematical formula. This mode should give by far the best results for excessively noisy footage, as well as slightly better results with normal noise reduction. It’s designed to give an optimum balance between the wanted reduction of noise and the unwanted softening of the picture."

sshot-1432.jpg (124.63 KiB) Viewed 1440 times

sshot-1431.jpg (121.38 KiB) Viewed 1440 times

Sony FX6 test files for download
by Claus Andersen | Jan 24, 2021

XAVC Intra 4K, 4096×2160, 50 fps, S-Gamut3.cine/S-Log3, 4:2:2 10 bit, 500 Mbps, shutter 180.0 deg, ISO 12800, WB 5600 K. Approx. 1GB file. – Singer/Songwriter: Gorm Bloch

https://filmplusgear.com/sony-fx6-test- ... -download/

Just wanted to thank BMD for an excellent job implementing Ultra noise reduction after finally using it to completely solve a problem that was ruining my recent project.

I have been using Enhanced Small spatial and 6x6 spatial values very successfully for years until this past week where the first minute of a two hour theatrical video was ruined with extremely ugly noise reduction artifacts. The area of the frame that was ugly was a brightly lit blue background to simulate a night sky. There wouldn’t be a problem if that was all that was happening. However half way through the clip, a fog machine filled about half the background and the fog was picking up intense white light meant to simulate the headlights of an approaching car. A large area around the bright white fog was rendered horribly.

I switched my NR Enhanced node for that first minute to be the first node in the node tree instead of the last. That made almost no difference, but I think I’ll continue putting NR node first in the future.

Desperately I tried Ultra NR with settings 12x7 as analysis of different areas suggested up to 10.7x6.7. Voilá! Success and a beautiful night sky with bright fog looking exactly as our eyes saw the scene in real life.

Gratitude!

Edit
I should point out the issue with artifacts only occurred on my 4K PQ h.265 deliverable version where 2K CDNG footage from my UM4.6K camera was upscaled to UHD to match the BMPCC4K. The HD versions for h.264 BT.1886 and sRGB deliverables did not have any artifacts around that swirling bright fog using Enhanced NR.