Dear all,
IRIS does a really good job imho. But it does filter out most freckles, even with high “recover” rate. Maybe you may train its filtering?
Thank you very much!
Best regards,
Andy
Dear all,
IRIS does a really good job imho. But it does filter out most freckles, even with high “recover” rate. Maybe you may train its filtering?
Thank you very much!
Best regards,
Andy
That’s the biggest problem I see right now with Iris–loss of detail regardless of settings. Sharpening can often be improved by upsampling using Artemis and Unsharp mask effects in an NLE, but loss of detail can’t be compensated. On a few videos where this was an important issue, I made two outputs using different models and composited them as layers in Premiere with tracked masks facilitating the details where they were lost. Very time consuming.
Iris seems to do a very good job so far (not tested much) but I see something very weird.
Same video, 640x480 same settings, upscaled to 200%, Iris Auto
Mac Mini M1 8GB RAM 6.1 FPS
Mac Mini M2 16GB RAM 1.7 FPS (!)
Anyone noticed something similar or it’s just me?
[edit]
Trying to find what is wrong I noticed a big difference in system activity.
There are two processes running the encoding: Topaz Video AI and ffmpeg
In M1 (with higher FPS) the Topaz process takes ~6% CPU and ffmpeg ~140% and in M2 (with the lower FPS) the Topaz process takes ~95% and ffmpeg ~40%
And because ffmpeg does the actual work it’s ‘logical’ to have such reduced performance.
Any clue?
Stargate SG-1 upscales have never been acceptable to me…until now. Iris does a terrific job on the first three seasons (which were shot on 16mm film). I had used Proteus previously but if I got the overall image clean and sharp there were artifacts that looked like halos around people’s eyes. Iris has fixed that!
Now, to see how it works on Star Trek Voyager!
Also trying to give some live to the first three seasons of Stargate SG-1 with mixed results so far.
Are you working with the DVD release or the BluRay?
The BluRay. Here are my current Iris Settings (still refining them a bit)…
Relative to Auto:
Revert Compression: 0
Recover Details: 10
Sharpen: 5
Reduce Noise: 0
Dehalo: 2
Anti-Alias/Deblur: 7
Add Noise: 0
Recover Additional Detail: 0 (you may want to up this for S01E01, mainly for the opening shot in the gate room).
I don’t add grain, but that’s a personal preference sort of thing. That should give you a starting point.
I output to MP4 containers using H265 Main10 @ 32Mbits/sec (approx 8 to 10GB file output)
If you are starting from the DVDs, my reccomendation is to do two passes. Try Iris for the first pass and Artemis for the second. I’m working with Voyager DVDs and I go to 960p lossless on the first pass, and then 1080p H265 on the second.
I was working with the NTSC DVD’s but not to long ago changed to the Bluray edition after a friend lent me them.
The DVD’s are a pure pain to make the episodes to look clean and sharp, the Bluray’s are much easier however they also have problems and the most noticeable ones are color banding.
The settings I’m using do not differ much from yours…
Manual:
Revert Compression: 1
Recover Details: 10
Sharpen: 0
Reduce Noise: 1
Dehalo: 0
Anti-Alias/Deblur: 4
Add Noise: 0
Recover Additional Detail: 8
I’m working with the CLI so I can pipe Avisynth scripts and the script I’m using have a grain plugin to help cover almost all the color banding.
Output to FFV1 lossless codec 422 10bits and later encode to H265 10bits.
I’m trying Proteus after Iris to give more live to the final result and so far I’m loving it.
Here’s some screenshots from before and after.
Frame 7050 from Brief Candle
Source…
Iris…
Proteus after Iris…
Thanks for sharing your settings
I also agree with the workflow comment. Prior to Iris, I used Proteus in most situations, occasionally using Artemis for sources that responded better to that model. For my past two encodes, each with a mid-to-high quality source, Artemis gave better results than Proteus. In each case of using Iris, I found that I had to use an extreme setting in manual mode that I would have never contemplated with Proteus. In the first case, I had to really crank up the recover detail and recover original detail north of 70 to get the desired results. Concerning the “petroleum smear” that another post complained of, I encountered the same thing in a second encode and found that it was due to noise. I progressively increased the remove noise slider, and this disappeared from all frames once it hit a value of 70 (it was a very noisy high-definition source); I had to compensate for image softening by ratcheting the recover original detail slider to 90. Still, I got the results that I wanted. I would have never considered these settings in Proteus.
I should add that I don’t see Iris as being limited to low-to-medium quality sources. The two sources that I’ve used since Iris’ release, higher quality sources, responded best to Iris. Artemis, Proteus, Gaia, Thea weren’t even a close second.
Here’s an example of my first Iris encode, using Wings footage from 1976. This was a case of requiring a lot of recover detail/recover original detail. My processing path was de-interlacing and convert to 1:1 square pixel in Hybrid before processing with Iris in TVAI. Within TVAI, this was a one-pass operation, upscaling from SD to 1920x1080, with fine grain added.
original interlaced footage…
TVAI using Iris and fine grain…
I’m pleased with the results. The Iris settings used for this were…
revert compression 10
recover details 63
sharpen 4
reduce noise 2
dehalo 0
anti-alias/deblur 3
recover original details 90
and fine grain (amount 1 and size 1)
What are the Proteus settings you’re using after Iris? I thought I could get away using Proteus to sharpen up my Iris renders, but it still tends do unforgivable stuff with people’s faces. Then again, the videos I’m working with are probably in much worse condition than your DVD example.
For the moment, the only reliable method I’ve found is maxing sharpness on Iris renders, gently sharpening through upsampling with Artemis or Gaia to HD, and an Unsharp mask effect when color correcting in Premiere.
Almost none…
Manual:
Revert Compression: 0
Recover Details: 0
Sharpen: 0
Reduce Noise: 0
Dehalo: 0
Anti-Alias/Deblur: 10
Add Noise: 0
Recover Additional Detail: 20
But I forgot to mention before that I’ve downscaled the Iris pass from 1080p to 576p (1024x576) and upscaled 200% (2048x1152) with Proteus. It’s faster this way and the quality is better.
There are to many ways to skin a rabbit… If it works for you and you like the results…
Hi Andy! Can you share the input file you are using here? 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.
Is your memory usage set to 100% ( Topaz Video AI > Preferences)?
Don’t set the memory to 100% on Apple silicon for SD upscales as this will reduce performance.
See e.g. here:
Hey there! I am investigating this user’s case and for this reason, I do need them to raise their memory usage. Thanks!
ffmpeg seems to be not really optimized for M2 and/or multiple cores at really high counts.
Handbrake which heavily relies on ffmpeg as well e.g. here is only slightly faster on the M2 Ultra (24 cores CPU and 60 cores GPU) than on a M1 Pro (10c CPU / 16c GPU) when doing a single encode.
Only when doing 2 or more encodes at once the M2 Ultra begins to show its strength, being over double the speed of M1 Pro. So I guess ffmpeg somehow is limited to a max number of cores for a single task?
I tried both scenarios in M1 8GB and M2 16GB minis.
10% memory 1.8 FPS on M1 and 1.1 FPS on M2
100% memory 2.2 FPS on M1 and 1.7 FPS on M2
I still don’t understand why M2 is less capable.
And this happens not only with Iris but with every model
Is there any chance I have a malfunctioning M2? It’s brand new and I bought it to improve the encoding speeds just to see them worse…
1.7 FPS is normal for an upscale 3X for a video of 640x480 pixels?
Also, I see a big difference between Windows and Macos
I had similar speeds of ~6.5 FPS on versions before 3.0 (2.6.4) in a windows Ryzen 9 3900X 32GB RAM and Nvidia RTX2080 and M1 8GB and now M1 (and M2) are much slower.
I understand that Video AI is a complete rewrite but I would expect similar results between win/macos when they were similar before.
We are looking into this now and I will update you ASAP, thank you for sharing these details with us
Hi MarkWalterMD. I found out that my particular “petroleum jelly” smear problem with IRIS is related to my Nvidia GeForce RTX video card. The smear disappeared when I changed my Processor Settings to CPU and this works even with Reduce Noise set to zero. Thanks for your post because it inspired me to investigate my problem further (after setting sharpen to 100 and recover original detail to 100 didn’t fully remove the smear from my IRIS tests).
Hey, has anybody else noticed that Iris doesn’t seem to follow the bit rate setting on output? I can’t get more than ~22Mbps even when the output is set to 40Mbps. Other models behave as expected…raising the bit rate results is less compression and bigger files.
The bit rate is being set correctly in the FFMPEG command, but seems Iris is throttling the output.