Theia detail creates noise (quality problem)

I don’t think this problem has been reported before.

It is about Theia detail creating noise in the output file.

Theia detail can clean noise nicely but it also creates noise in the output file.

Tip for viewing images :
Right-click the image, copy the image address and open it on new browser tab/window.

Edit : the source video was processed in 100% output size (no upscale) for denoise/deblock.

Source

As you will see, it is just a plain black background.

screenshot of the original noise

Original noise viewed by neat video

Theia detail settings

topazsettingsPNG

Theia detail preview

Outputs

Depending on the output format, the original noise is replaced a different kind of noise.

Mp4, mov and jpg output have rectangular shapes

PNG and TIF display a grid.

Conclusion

I think it would be good if the Topaz team could check the quality of the output with a third party tool like neat video.

The result is not consistent and changes depending on the output format.

Noise is created by the AI.

1 Like

Hi,
So your JPG and MP4 (h264) outputs both have compression artifacts since those are lossy compression formats (I assume you didn’t use CRF 0 and even then blah blah complexity subsampling conversions).

However your PNG and TIFF output do show a clear pattern that for now I’ll call noise but I wanted to be sure. I ran experiment to see exactly what is introduced during noise reduction with a 100% lossless input.

I can’t post images yet but here’s a OneDrive link, you’ll need to download the original for the PNG
https://1drv.ms/u/s!AghXT-yfmnmUoBlRf46LofRCiYCf?e=g8iLUM

Source

Video
ID                                       : 0
Format                                   : RGB
Codec ID                                 : ULRG
Codec ID/Info                            : Ut Video Lossless Codec
Codec ID/Hint                            : Ut Video
Duration                                 : 1 s 0 ms
Bit rate                                 : 151 Mb/s
Width                                    : 1 920 pixels
Height                                   : 1 080 pixels
Display aspect ratio                     : 16:9
Frame rate                               : 24.000 FPS
Color space                              : RGB
Bit depth                                : 8 bits
Bits/(Pixel*Frame)                       : 3.040
Time code of first frame                 : 00:00:00:00 / 00:00:00:00
Time code source                         : Adobe tc_A / Adobe tc_O
Stream size                              : 18.0 MiB (99%)

This was created with the HD bars and tones in Premiere. I wanted to verify the output was lossless so I extracted one frame from the video and verified that for each color bar each pixel was the exact same. I then ran this video through with your settings and took a PNG from one frame (all had similar changes).

The above image shows some noise was created by VEAI and what I think might be an attempt to sharpen a little (even though sharpen was set to 0). This makes sense since VEAI isn’t just simple spatial and temporal noise reduction as that’s already present in software for a while now. I believe it’s trying to avoid detail loss.

I think if you take a video that already had noise reduction applied to it and then you analyzed both temporal noise and color frequency noise it would just show that output as noise VEAI didn’t perfectly change your obvious (to humans) black video to just RGB0,0,0 for all pixels in all frames so there is still difference both temporal and by color frequencies. How would Neat video know that? This is why the noise reduction is a human judgement, otherwise there would be options.

@ctdeneen Your reply is certainly interesting. Neat video shows noise on different layers of a video. My screenshots show the Y layers but there are other layers too. Neat video does identify temporal and spatial noise.

We’ve got plans to update Theia soon (it hasn’t been updated in quite a while) and will keep this in mind. The focus has mainly been on Artemis / Dione updates and a new FPS conversion model which will be out soon.

1 Like