I’m sure I’m asking too much. But it is precisely the reason why I’ve paid and I’d pay for upcoming releases.
The amount of parameters is overwhelming.
I understand you want to provide full control of output, but for occasional users this is time-consuming. If there’s any product manager there, you should understand that people do not have time to try parameters and check video tutorials. Even if I get all the parameters 6 months later I forgot all of them.
Is there a kind of “wizard” that analyzes my input video (using mediainfo, ffprobe, etc) and then asks me relevant questions to auto-configure the upscaling?
Is your video a stadium live concert? It is a solo recording?
It is a sports match? It is football? It is a tennis match?
I don’t think those type of questions/answers would benefit VAI. as to my understanding VAI analyze the picture frame by frame (well, once every 8 frames to be more exact) and refines the picture quality based on the AI engine intelligent how it would / should look the best for this type of image for this type of footage/content.
So even if you answer all those wizard questions, the AI engine should know those already based on its training module to detect the type of content it is. that is the point of AI, to do the thinking for you.
Since AI is not perfect, that is why we have manual mode as well, “to help it out”.
VAI/Topaz were smart enough also to introduce the “relative to Auto” mode (in v5.3 it is now under “Enable parameters → Dynamic”), exactly for us novice users that don’t want to spend hours finding the best settings for each video.
What it basically does (the intent workflow), is:
You export in Auto mode and you review the results
let’s say for example that you are not so satisfied with the sharpness of the video, but you are happy with the rest
you export again the source video, but this time you pick “relative to Auto” and bump up the sharpness and leave everything else at “0” ( “0” = “Auto” in that mode ).
What that means (example below) is, for sharpness you would apply “20” point extra on top of what the Auto mode estimated (e.g. if sharpness in auto was estimated by VAI auto mode to be 35, you will now apply Sharpness of 55 - Auto values + your 20). the rest of the values are left at “Auto” mode (“0” = Auto). that way you get fairly quick to the final results you want, as Auto plays well on most parts, just needs sometimes little fine tuning.
I agree with Akila. You’re asking for what the Auto and Relative to Auto already trys to do.
The fact that they don’t really work as intended for DVDs is not something that will improve fast. I’m guessing it will take years and several model versions. (And for people to constantly tell Topaz that they are not trying to upscale to 4K but just FHD.)
For the most part, any video I try that’s larger than standard DVD resolution comes out really well using the Auto parameters mode. So, I think it does work, just not for DVD and lower resolutions.
I will try to explain it in a better way so you can understand.
If what I’m asking is already there, then it is a massive failure from the UX/HCI side. It should be very clear for users without specialization in video processing. I understand that is a kind of iterative processing which @Akila explained as “re-exporting and bump”, but then translating this technique with the number of parameter combinations, and then the time to analyze results…
Are there any examples of what you can do better with very low-quality video?
I understand that you want to promote your product with spectacular results, but what I often really want to know is what is the best I could achieve with the poor-quality video I already have.
the UX was never their strong point, can’t argue with you on that one. even to this day many things I still don’t understand fully what they do
Low quality videos are very tricky task. you get mixed results. it is heavily dependent how bad the quality is and what type of content we talking about. personally I never had satisfied results on the handful of low quality videos I worked on. I just stopped bothering trying.
Most of my content is medium quality content from a Sony Video8 that was captured in a very good quality state (as much as Video8 allows over Firewire), so with those I get much better satisfactory success rate.
In that sense, that is what they are trying to do with the newly added examples to help you pick what model to use.
The best I can say is that knowing the history can help reduce the amount of models to try.
Warning: this history is long and very one-sided from my limited point of view. It might not be worth reading.
The Artemis and Dione models are old and made to do a rigid general enhancement based on general possible video source conditions.
I don’t know about Theia, but it has never produced usable results for me.
Gaia, is unique. It was made to work on new/modern cartoons with no grain/noise. I’m guessing that, because it tends to ‘bake’ noise into what should be solid colors. It can be the best model for anything, or it can give the least noticeable results. It’s the least destructive for sure.
Then came Proteus, the answer to the problem of the Artemis/Dione models. You could get similar results, but modify the parameters to account for anything that might be wrong.
Proteus Auto was made to stream-line that process of finding the best parameters.
Biggest complaints against Proteus were it tended to make far away faces into monster faces, and tended to need denoising to get the best results. Proteus 4 has reduced or removed those issues for most videos.
Iris was made to work like Proteus, but have an additional face enhancement layer added on. The answer to the monster faces. It had the added potential to work on home video sources.
The biggest complaints against it are it makes different kinds of monster faces—extra teeth, or phantom faces where none should be. (Probably mostly from super low resolution sources.)
I’m not really sure what’s different about Rhea compared to Iris.
Nyx is just for denoising. Nyx 2 over-sharpens no matter what. Don’t use it. All the other versions work good though.
…and that’s just the upscaling enhancement models. I don’t use the others. I try the frame interpolation models when they come out, but I have not agreed with the results completely yet. They’re better than interpolating TV screens though.
Reah on paper (Topaz mission statement) is the Hybrid (or combination) of both Proteus and Iris.
Proteus it’s strength is general purpose image and small faces in the background, while Iris is supposedly more powerful for the predominant face, the foreground face in the image, but not as good on none face image (or small faces) as Proteus (v4). again, based on Topaz mission statement on those AI models intent of use.
Reah on paper suppose to be “the” model, as you get the best of both worlds, the strength of Proteus (v4) algorithm to a none face part of the image and/or the small faces in the image background & Iris algorithm on the noticeable, foreground face(s) in the image. that is Topaz intent.
Rather it works well or not?
well, I don’t know. currently Reah is only supported at x4 upscaling, which is way to high for my SD content, which produce very bad image due to the high upscaling scale for my SD content.
My content can’t handle such high upscale attempt as of today. so I can’t use it nor compare it with anything else I am familiar with (other AIs).
