All right! Haha I asked in case or because we will say this park, but it would have surprised me anyway;)
again an example of bad face refinement, I hope that AI models can be trained soon not to exagerate shadows and contoures
3 Likes
Marty, sorry, but I disagree. An āintelligentā system should at least recognize that its action get worse than the input it is processing, and leave it as it is when it cannot improve.
2 Likes
Yes. I have had more than a few instances where Iāve managed to enhance video to a very acceptable quality, only to find that peopleās faces were turned into either blotchy messes or smoothed down to look like anime illustrations.
It would be nice if there was an option to turn face recognition/enhancement off. There are many instances where it does more harm than good.
It would be very nice if the face āenhancementā feature could be switched off.
2 Likes
Have you found that face distortation is dependent on the quality of the source material ?
I sadly have to add to the choir that 3.0 is a BIG step down. Iām running an rtx 4090 and while performance has improved for me, it can no longer preview properly. It constantly lags behind, i.e no moving picture in preview, then a jump. Once DONE processing, the preview starts working and runs for a long time after. Where is compare? Much less options than before, at least i canāt find a way to get the same granularity, and worst of all, the results are noticeably worse :(. I think it would be Better to keep this a separate āquick processā product, and use the existing one as the more stable and feature rich also going forward. My upgrades end in a month and Iām pretty upset that probably at that time you wonāt have fixed this to an acceptable level. Iāll probably need to stick with the older version for a year or two until you get this functionalā¦
3 Likes
Of course, the quality of the input material is always a factor, but since they are now giving faces āspecialā treatment, this makes for a different situation. The problems with āenhancingā faces is usually at its worst when the source image is low resolution, and the faces are small within the context of the image.
Closeups or large images of faces tend to look good; or at least as good a pre-facial enhancement faces looked. - The problem comes with small faces. It appears that the enhancement algorithm finds little, indistinct faces and overdoes the enhancement, which tends to create speckles and blobs where the facial features should be.
As I mentioned in a previous post, Iād appreciate it if they would provide a way to turn face enhancement on or off, OR provide a slider to control the face enhancement sensitivity level.
2 Likes
I find the Video AI 3.xx interface very frustrating. Things have changed too much and I canāt easily accomplish things that were easy with the v 2.xx versions.
How the hell do you mark the frames you want to process to stills ?
The Brackets for Start and Stop are gone. When I use the Trim, I canāt figure out where the beginning frame or the end frame are.
1 Then when I try to process it, either nothing happens or the output ends up somewhere other than intended. Nor could I find any instructions anywhere
2 Worse yet, prior versions put the output next to where the orig video appeared, now when I look at the Preferences, it looks like I can only set 1 permanent output folder. This is ridiculous.
Why canāt the Interface change minimally, so we donāt have to learn some totally new language. Change the engine as much as necessary but donāt mess w the interface.
I give the developers an F for failing to keep things easy, and especially for failing to provide a Manual, or a video with instructions. There is a video showing how to handle multiple videos at once, but I first need to know how to learn the new FYF%$%# interface.
ADDED LATER - Never mind some of the comments, I noticed that right at the start before importing anything there is NEW USER link and some things are explained. However selecting START and END Frames is not explained, as far as I am concerned, but the Export Location is still totally confusing to me.
6 Likes
āā¦ Closeups or large images of faces tend to look good; or at least as good a pre-facial enhancement faces looked. - The problem comes with small faces. ā¦ā
That is exactly what I try to document with my examples shown.
The videos Iām actually processing are originally produced by professionals for TV but squeezed by whatever (270,000 frames into 1,800,000 kB) into progressive 1280x720 format, and mp4 container. Off course itās a challenge to rebuild original resolution and enlarge again, even more challenging when speed matters (0.3 spf mean up to 22 hours).
TOPAZ has proven AI competence in their photo section where subject detection and face refinement can be done successfully. Well, single image processing @ 11sec/image cannot be applied 1 by 1, but at least TVAI should avoid to invent speckles and blobs when it detects small faces.
1 Like
How do i export multiple videos without making them all have the same settings? I tried adding 20 videos each with their own output settings but it will only export the selected video. If I select all it just makes them have the same settings. undoing all the work I did. Am I doing something wrong?
My issues with 3.0.2 are:
- Panning and zooming during preview causes it to stop responding and show the hourglass cursor of death. I have to kill it off and restart it.
- Processing 2K to 4K using Proteus takes 0.74 spf, whereas in TVEAI 2.6.4 it takes 0.52.
NVIDIA RTX A4000, Intel Xeon Silver 4216, Windows 11
With each iteration it got a little better, but now with V3, itās actually worse than the V1. I canāt do anything with V3, and I donāt want to reinstall V2. I think the best solution is to extend free upgrades until end users see measurable improvement in video quality and processing speed. We really need a concession from Topaz for this. Iām sure end users, like myself, would appreciate it.
9 Likes
I put my work with ve on hold too. Still wonder why the basic functionality canāt be restored in such a long period of time. Sometimes I have the feel that the main priority relies on the support off all sorts of third party tools: lightroom, photoshop, ffmpg, avisynth where at the same time very basic functions which are fundamental for the application, such as a proper working preview or selecting individual frames, fall short under the table for the sake of those tools. 
3 Likes
Hi. I canāt provide logs Iām afraid as Iāve uninstalled it, since it was completely unusable for me. It was happening almost every video I tried, plus Iām having the same issues others are wilth lack of features, previews not working well, and results taking so long itās not feasible.
Can I make a suggestion please? Apologies if Iām repeating something already covered. I like this software and I want it to succeed - Iām even prepared to help try and trouble-shoot and provide feedback. But you cannot expect people to pay for software that is so clearly not ready, knowing they will waste their ā1 years free updatesā while they canāt even use it. I held off on purchasing until v3 was out because v2, missing some offline reliability, is no good for me. The way things stand I therefore have no choice but to get a full refund and come back in a year and see if it works properly yet, if no other solution presents itself before then.
The alternative is to offer everyone like me an extension to the āfree updatesā offer while you sort this out, which I think would be very reasonable. If I was offered an extra year of free updates, or as many months as it take to reach a stable v3 that has reasonable feature parity with v2.6.4, then I would not refund, Iād hang around patiently and help provide feedback and wait for it to get better.
Obviously itās up to you what you do, but I feel thereās a lot of goodwill being blown with this release, and at present I feel like Iām paying you to beta test your software, which isnāt reasonable. Youāre obviously under a lot of pressure trying to get the new version back to where 2.6.4 was. Offering loyal customers free updates while you do it would go a very long way in my opinion.
8 Likes
Itās been a big setback but I hope that they can improve. I hope it doesnāt take too long to get back to the same quality and speed as 2.6.4. Then obviously I want things to improve even more.
So far thereās been a huge delay with version 3 and nothing in that version is really any better than 2.6.4, but if they needed to rewrite the code in order to make the future product better I am totally fine with that, I just hope it doesnāt take too long to get there. Other programs are adding face enhancement for video and Topaz is falling behind.
Iād like to see huge speed increases too that use the full capabilities of my GPU.
6 Likes
I agree with your view. I hope Topaz will extend the update period for those of us who purchased 2.6.4 and have since then re-upped, purchasing another yearās updates.
VAI 3.0.2 is simply not a release product, for all the specifics others have articulated.
At this point, Iāve spent ~ 150 hours working with VAI 3.x.x. It was spectacularly successful on one video of a SpaceX launch that I took with a cheap point & shoot. Itās been unusable on seven others - DVD scale-up, DVcam video improvement, and other sources. The shortcomings have been so numerous and varied that I despair of getting ahead of them.
Iāve decided to return to VEAI 2.6.4. On the whole, it gives me better results, itās faster, and itās more stable than VAI 3.
I hope VAI 3 will eventually be the stable, innovative, and fabulous product we all want. Meantime, Iām aware that my update period is ticking away, and I donāt have a release product. Topaz needs to address this soon.
13 Likes
Right on! I think Iād prefer the facial recognition capability from V2.x to what is very problematic in V3.0.2. - I can tune Proteus to do a VERY acceptable SD to FHD enhancement; that is until it detects a small face. - Frankly Iād prefer not having that feature to having a feature that makes a mess or forces using a lower enhancement level to avoid the face problem.
AI is a miracle for image enhancement. However, it needs two things to do the job properly. First are recognition samples, which it uses to identify the parts of the image. Next, it needs ārulesā to keep the recognition/enhancement process in perspective. When you turn ādetailsā up or āsharpenā up, the AI tries harder to see/enhance the image.
Presently, the face recognition is too aggressive in comparison to other recognition objects, which makes for that dialing in a good image that gets spoiled when it finds (and overdoes) a small face.
2 Likes
I havenāt found any face recognition/improving option in V3.x yet. Where it is? I also heavenāt heard about this special feature added to video ai yet. Think it was only included into the photo applications?
1 Like
Agreed. As far as my testing has found, setting the recover details slider all the way to 100 just darkens a lot of edges of most everything.
1 Like
