Experiments with New Iris Model

Yeah, and the audio functions are acting funky now too.

Hi again. I just fixed my smearing problem by updating the driver for my Nvidia card. Just as well because having Processor set to CPU took 10 times as long to do any IRIS Export.

1 Like

That’s interesting. I wouldn’t have guessed the use of GPU vs. CPU would have resulted in quality differences rather than encoding speed differences. I’ve been using my nVidia graphics card for my encodes and haven’t tried using the CPU. Have other people been having this problem as well?

Hi MarkWaltermd. My smearing problem turned out to be a problem with the driver for my Nvidia RTX2060 video card. As soon as I upgraded the driver to the latest version, I finally got to see how amazing IRIS looks to everybody else. (Changing my Topaz Processor option to CPU was just something I tried out of desperation when nothing else worked ).

1 Like

Are you certain that this is not file-specific? If your file has a max bitrate of 22 and you choose 40, the output will be 22.

If this is not what is occurring, please share the input file and the Topaz Video AI exported file with me.

You can securely submit your files(s) to our Dropbox using the link below. Please be sure to send me a note to let me know you sent something.

Submit File to Dropbox

Can you share more about this?

@ida.topazlabs I tried the Iris model on a late 70’s music video. The various/hilarious added/enhanced eyes aren’t going to work for me but are good for a chuckle.
Other details added/enhanced like the hair, teeth and microphone are impressive though! :slight_smile:

If you like I could upload this source video to your dropbox.

Thanks, Sheldon.

edit: I should mention that the unnatural eyes occur throughout and that I realise this is tough footage considering the lack of detail and makeup and bangs/fringe over the eyes :wink:

Try rising the recover original details to a really high value, e.g. 85 or even higher.

And then also up the deblur value to something beyond +40 to still actually get more sharpness /details in other areas.

2 Likes

I am upscaling from 1080p to 2160p. Source is compressed with AVC, 24fps, 1080p @ 19.2 Mbit/sec. Outputting to HEVC, 24fps, 2160p @ 32Mbit/sec. However, the output file is actually only ~17 Mbits/sec. I have tried replacing the default profile (medium) in the command line with the ‘slow’ profile, but with no difference.

Forgive me if I am mistaken, but doesn’t FFMPEG decompress the data and then recompress it? Even if changing the codec was the only change, if told to use 32 Mbps it will output a ~30Mbps output. Since I am quadrupling the number of pixels, the output should at least be slightly larger than the input, even allowing for the greater compression of HEVC. Instead the 4K version ends up being 15-20% smaller than the 1080p original.

I am running an upscale right now in v3.3.3 to make sure it hasn’t gone away in the new build and will post the files to Dropbox when done.

Thanks Joachim. For this case that doesn’t help. The problem is poorly placed/sized and/or weird-looking eyes being added to the video.

Here is another example>>

I’d like to give this footage to someone who can use it to improve Iris or some other model.
I’m willing to wait for as long as it takes before upscaling my video collection.
– Sheldon.

HA! This one gave me a chuckle too… also those lips! Wow, AI enhancement for sure! Yes, that would be great if you could share the input file as well.

If you reduce the sliders are the results still this wild?

You can securely submit your files(s) to our Dropbox using the link below. Please be sure to send me a note to let me know you sent something.

Submit File to Dropbox

Here is a sample with sliders at 0 and previous ‘estimate’ slider settings. It is different. Maybe a little better too.

and a couple others for fun>>


Also the microphone flip flops from being spherical to flat topped.

I’ll upload the file to your dropbox.

1 Like

You are mostly spot on, but, there are a few things to consider.

The Iris model is quite aggressive when it comes to noise removal.

The compression algorithms are more effective when there is no noise. The encoder cannot give a bitrate higher than necessary if there is not much information to encode. So, if the source video has a higher amount of noise, the bitrate could be high, mostly because of the noise.

However, if the Iris model removes the noise in the output video, (even with 2x resolution, or 4x number of pixels) the compression algorithm would be much more effective and the bitrate would be lower, without compromising the quality.

I think you are correct about noise removal. The source material is quite noisy and, while it is a 1080p Blu-ray rip, also is not very sharp to begin with. So large areas of color, once all the noise is gone, probably compresses down as one shade instead of dozens.

I ran a few minutes of the source through Proteus with ‘reduce noise’ set to max and got similar output bit rates.

Thanks for setting me straight! :slight_smile:

1 Like

Can you provide me a short clip as well?
Short videos can be uploaded here when zipped.

OMG! Those “ears” :rofl:

1 Like

Since its release, I’ve had very mixed results with Iris. On two videos which were originally 1080i, I achieved spectacular results with Iris that couldn’t be matched by the other models for progressive input (I had de-interlaced prior to TVAI). This was with minimal upscaling.

I also attempted to use Iris on 2 sources, both 480p, in which Iris performed terribly due to over-softening/blurring. This over-softening effect could not be corrected through aggressive use of “recover details,” recover original detail," “sharpen,” or the “anti-aliasing/deblur” slider. This includes closeup scenes in which faces were prominent; no matter the setting, the untouched source at the original resolution looked better than anything Iris could offer.

So, in my limited experience with Iris, the results can be outstanding or outstandingly poor, depending on how Iris responds to the source material. When it performs poorly, it is due to extreme softening/blurring with loss of textures and edge definition. Can this model be tweaked to limit the softening effect? Thanks.

2 Likes

“Iris” is “Artemis Low” little brother. that is how I find this AI to be.

Except that I would have never considered using Artemis low on the 1080i sources that got exceptional results with Iris. Unquestioningly, Artemis low would have ruined it. Iris destroyed the 480p sources, in the same way that Artemis low can.

and you are not supposed to use Iris on a high quality video (1080). that was explicitly stated by Topaz.
it is not meant for that.

1 Like