On this, I have yet to see any improvement to faces in Proteus that the older models cannot beat, and I have spent a very long time trying Proteus. It can get close if you try very hard to customise the settings, but Proteus seems to be less of a tool for upscaling, and more of a tool for cleaning up existing video. It works very well as a simple alternative for some more complicated pre-processing for example.
Add a little bit of de-halo in Proteus to gain back the ability to recover details. 
The problem you mention with faces may lie in your pre-preparation. It could also be due to attempting to rescale by too big of a magnification in one pass. - Note: To do multiple passes sucessfully, you really do need to be using a low loss or lossless CODEC.
2 Likes
Can you elaborate on how adding dehalo helps bring back more detail?
Nope. I have been using this application for years now, and I have tried hundreds of inputs and variations of inputs and outputs, sometimes trying 20+ in a single session of testing.
I can confirm quite convincingly that Proteus doesn’t do a good job with faces. I have already had this discussion with Development as well as I assume its from not working on the models for a bit, so I am hoping it gets better - but right now, all version of Proteus will not perform as well on faces as older models, including in V3 - with the caveat that this is limited to 90’s DVD sources and not say hand cam and really bad footage as I have not tested these.
1 Like
the early Artemies modes were good in recovering details, but du to the attempt to make it look “more natural” again Artemies lost this ability to a major degree. But the early Artemies is still included in the de-halo and strong de-halo modes, so adding these helps. The Proteus option “recover details” does in fact not recover any details but makes existing ones more visible.
I actually tried this - using these models specifically because of a similar comment - and they do work ok, but I still go back to Artemis V8 prior to the large change that screwed it all up. I do routinely check the newer versions to see if I can get a similar output, but no success so far.
Proteus does, however, tend to work well for the CGI sections of footage on the DVD’s. Its ability to clean the footage up without needing to recover details means I can usually use it for that purpose - just not people and faces.
Astana,
I think you have misunderstood my reply. There was nothing personal in my reply.
All I was saying was that TVAI 3 works best if you don’t try to reuse the techniques you may have used in the previous versions. And, it takes patience to become familiar with this version and get it to work for you. - I had a lot of problems with 3.x until I came to that realization.
And I did say that it is a young, new product and it still has a few glitches. If you would prefer to go back to using the previous version until a few more of the problems are fixed, then you are not alone. A lot of others have already done that. (And perhaps I would do that, too - if I had any time constraints. But I don’t; I’m retired and mainly do this for my own amusement.)
If you are having a problem with a specific issue, post it here. Many of the other forum subscribers (including me) may be able to help. Or, just contact Topaz support; they are very good and happy to help…
2 Likes
I accept
Older versions of AI were aggressively trying to build models.
Current AI version aims to be clean
I hear what you are saying, but recently I have had more success by getting my video very clean at original size and then rescaling it upwards in small increments. Also, be careful not to turn up detail/sharpness too much initially, you can do more with them when you have have scaled up a bit. detail cand do too much until there is enough room for more detail. Too much sharpening, in low res, (too early) is not especially helpful either.
1 Like
I disagree. I think they are currently putting more emphasis on the GUI and it’s peer handshake with FFmpeg. There are stability issues and little glitches at many levels that must get fixed. I’m certain they will be putting an emphasis on perfecting the models soon after that.
1 Like
I understand what you are saying, I wasn’t going to go into the extensive testing I have done, but your answers seem like they are directed to someone with little to no experience with trying these models. Everything you have suggested I have done hundreds of times - I am well aware of how inputs and preprocessing the footage can affect the output and also how to use Proteus. I use it for the CGI elements currently because for non faces it does well.
The point I was making is that if you are wanting to upscale faces and retain good level of detail, steer clear of Proteus and for the moment, any V3 version of this program. You will get better results, for far less time spent using the older models.
This is not implying that the current ones are terrible, or that you cannot use them, but if your spending weeks upscaling and want the best results, I wouldn’t be wasting time on Proteus until they start working on the models again to improve it.
It is good in that you can tweak a lot of settings, but almost everything it does can be done outside of TVEAI and the one slider bard that should affect quality of upscale itself, does a bad job of it. You would be better served spending time on adjusting the input and using the older models that do a better job currently - until they can work on models properly again.
1 Like
In my case, I did the same with about 10 frames (90 second video).
When taking samples, I would take 2 bright pictures, 2 dark pictures, 6 intermediate brightness pictures, and so on.
However, some scenes may have anomalous values, so this was taken into consideration.
If samples were taken at about one frame every 10 seconds and the median value used, etc., it would be faster than detecting in all frames and the effect of the AI filter would be more uniform.
1 Like
Agreed.
The changelog still says “Quality mismatch between 2.6.4. and 3.0.x for some models.”, which I can only assume means that 3.0.x has worse quality in some situations than 2.6.4 (otherwise they could remove this line or replace it by “quality improved”).
It seems that the program settings specify to use the CPU and not the GPU. Check it.
oProteus, or not. I have found that working from SD dimensions and having small background faces is a nightmare un less you enlarge it slightly giving the enhancement algorithm the need space to add some detail to the face. It might not be the original face; not when SD resolution give so very little information to work with and not enough room to asd the detail detail necessary necessary to describe a face.
To be specific, there are some faces that are simply to darned small to do much of anything with, and there are semi-close small faces it can recognize and do something with. The bottom linr is that SD res is so low, it is actually miraculous that ican add detail of anything even to enlarge it to FHD with detail that was only hinted at in the original image. If you feel the need to use the older versions of enhancements that you can then feed to TVAI 3 and get better results.
Gleaning as much info as you can from your posts, along with the others. But in my case, which is for upscaling old 80s 4:3 VHS sources from 480 to 1080, would you say that is too big a leap for a single pass?
Thanks in advance.
It was the 2x frame rate increase I cared about, the downscale just seemed a handy way to reduce the data for that (and it does support it).
I did try Apollo and just the frame-rate conversion, but it did poorly - okay, miles better than chronos(fast), but still not usable. Maybe it doesn’t like sharp source material?
Nope, sounds like a good plan. Se my post here for my going to 960p from 288p:
Also if it’s VHS you may need to fix the Interlace (which is an option).
1 Like
Thanks mate! 
just watched the two back to back and we’ll impressed.
