I’m about to start a new project, and at the time I was confident that the end product would be acceptable using the Iris enhancement and scaling. To prepare for the project, I used a Prelinger archive movie to do my testing. This was on November 21, 2023. The original movie was 320 × 240 pixels and I scaled it up to 1920 × 1280 and the results were excellent. Here is the Mediainfo file from the resultant video.
videoai : Enhanced using iris-1; mode: auto; revert compression at 0; recover details at 0; sharpen at 0; reduce noise at 0; dehalo at 0; anti-alias/deblur at 0; and recover original detail at 20. Changed resolution to 1920x1080
I’m currently on TVAI 4.2.1 and regardless of what setting I use, I can’t come close to the converted movie from November. I tried using the previous Iris model versions but with no luck. This isn’t the first project where I’ve found that previously saved setting using the Iris filter have come nowhere near the original Iris filter in quality.
I’ve uploaded a zip archive with the original movie that I first fixed in Premiere Pro “FreedomH1956_edit_1_tvai.mov” and the upscaled Iris movie “FreedomH1956_edit_1_iris1_1280.mov”.
In the upscaled movie, you can see details in the gentleman’s jacket and there are no discernible outlines on people’s faces or edges. A nearly perfect upscale and enhancement in my books, hence the confidence that I would be able to tackle my new project.
If anyone can add some insight on any settings that would match the previously processed movie, I would be grateful for the suggestions. I would also like a Topaz engineer to view the movie and perhaps tell me why the same filters and the same setting don’t match, or even come close in that matter. It’s very disappointing that I’m unable to reproduce consistent results across versions of TVAI. Freedom Highway (1956).zip (55.4 MB)
This is a pretty common experience for me, with a number of apps. The current version of my video editing app produces output with slightly different color balance compared to the previous one, so I keep both in case I need to go back and revise past work.
Yes you can.
When you install you have the option to name the directory install location.
Amend that and add version number.
IE: "C:\Program Files\Topaz Labs LLC\Topaz Video AI_3.5.4"
And you can have any version you want.
@windspear , my guess that the problem is not different TVAI versions that have different AIa even though they share the same version, no I don’t think that is the cause. my explanation of what might had happened is the “Auto” prediction values had changed with certain versions.
If you noticed, you used Auto mode. in Auto mode, TVAI tries to estimate what would be the best values every 8 frames for the intake source (in older versions it was once every 20 frames).
what might had happened is that TVAI when versions went up, the Auto prediction algorithm had changed (improved based on Topaz) and now it sets different values than how it was before.
The way for you to confirm or dispel this theory is by installing the older version which worked fine for you, run the upscale in Manual mode, press estimate button at the top of the values in manual mode and then run the upscale.
if that would upscale as you predicted (close enough), try running the same manual setting on the newer version and see if it yields the same results. If it does, then you know it is the Auto prediction settings that had changed, not the AI enhancement itself.
You can compare the results by using the Video-Compare tool between the 2 exports. just make sure you export them at the same codec & quality as much as possible to be fair on the comparing task.
Thanks, I’ll try your suggestions. I download Video-compare and tried it out, and it does an excellent job. Up to this point, I’ve been uploading my videos to Frame.io and using the comparison mode. Video-compare is a lot faster.
And on the Mac it’s even easier.
Just rename the old version before you install the new one (best via DMG).
I’ve had four different TVAI versions installed for testing at times…
You’re right. The manual estimates are different between versions of TVAI. I installed version 3.30, which introduced the first version of Iris. I initially converted six hours of old home movies back in June, but was unable to match the results using newer versions of TVAI. I used the Manual Estimate in 3.30 to find the settings that Auto had used and did a screenshot for reference. I then installed version 3.54 and found the Manual Estimate were much different than version 3.30’s settings. Adjusting the settings in 3.54 I was able to reproduce results that I could produce in version 3.30.
I then installed version 4.04 which was the version that I used to process the Freedom Highway videos that I referenced in my original post. I then checked the Auto settings using Manual Estimate and recorded those. I upgraded to TVAI to version 4.1.2 and noticed the Auto settings Manual Estimate were again different. Using the Manual Estimate settings from version 4.04 I was able to reproduce the quality of my previous enhancements.
From now on, I’ll be recording the Manual Estimate settings if I use Auto in the future. In fact, I may just get into the habit of recording my settings for any project. That way, in the future if I need to reprocess video on newer versions of TVAI, I’ll always have the original settings to fall back on.
Thanks for the heads-up on this issue.
If you’re using Auto, you’ll want to take snapshots of Manual estimates at multiple points. Because Auto settings get reestimated every eight frames or so. If fully repeatable results are your goal, you’ll probably be better off using Manual rather than Auto or Relative to Auto.
