Hi there
A couple of questions comments and issues
First of all, I’ve been trying to upscale SD video for several months now and I’ve tried every application and plug-in available
Your application blows everything else out of the water and the quality is impeccable and nothing even comes close.
That being said I understand that you have just released this application and I have noticed some issues.
I know it’s says it does not take interlaced video so I tried De-interlacing my videos and the application will not import them. The application just locks up and freezes.
When I bring in a source file that has not been De-interlaced the file loads off and is able to be processed by your application.
One major feature that I Believe it is necessary in your application… when you choose crop to fix screen and the red box shows the outline, you should be able to have a choice to position the red frame so you don’t cut off anything that’s important in the video.
When I import SD video 720×480, I’ve then choose crop to fix screen when I’m out putting at 1080 x 1920, a red frame comes on and shows the outline of the new video but I am not able to manipulate it to the position I want.
All applications that I have used in the past for this procedure has an X and Y access Point adjustment where you can relocate the cropped image.
I believe this is a necessary feature that needs to be implemented.
Also, I know your application disregards the audio portion of the video…
Maybe you can add a feature that will export the audio as a separate file before the upscaling process begins so they can be married back together at a later point after the processing and new file is completed.?
Also I primarily work on a Mac and I know you are working on a Macintosh version, any idea of a release date?
I work with a black magic GPU pro external video card and I’m hoping your Mac version will recognize this card and utilize it.
If you would like me to send the files that I have been trying to import so you could see what the issue is I would be happy to.
Just let me know how I can get them to you.
Thank you for your time and I’m looking forward to your response
Regards
Mark
This is from the scratch really a new way/idea in order to improve bad quality videos.
I tested so far a lot of small video scenes. Mostly video footage with resolution 640x480 and below.
Only with the modus “Upsampling(HQ-CG)” I get remarkable improvements. The other both modus are worsen the picture quality and therefore are not useful.
I tested all scenes several times with different resolutions/sizers/percentages.
Even I get remarkable improvements with modus “Upsampling (HQ-CG)” faces and hairs get destroyed!
So the software needs improvements at least on face recognition.
Today Video Enhance AI therefore is not usable for me. I wait, how improvements wil take place.
199€ are way too much for a not finished develped software.
HQ model as it says requires quality input with few compression artifacts and noise otherwise these defects can get amplified. HQ-CG does a good job of removing these defects but at the coast of a strong denoise so we loose the fine details. If we could get a control of the denoise amount we could get stunning outputs in some videos. The LQ model seems useless even on the worst inputs I could find from VHS or ultra low res compressed videos. It does nothing really noticeable at the best and often it gets worst.
As for the face detection I agree, there should be a face refinement like in Gigapixel AI and with a customizable threshold slider 16/32/64/128 pixels. It’s better to have a face that looks human even if the subject will necessarily look different as in reality than the subject looking like an alien or monster.
I made further tests last days so far again with bad quality video footage.
Resolution original footage 640x480 and below.
Sometimes the results are very impressive. I never thought before, that it would be possible to deblocking, denoise and sharpening such videos in such an impressive quality. But the results are very various from video to video. I guess, Topaz has to fill the AI much more with reference footage.
And again: faces and hairs get often destroyed.
ONLY the modus “Upsampling (HQ-CG)" brings sometimes such impressive results.
The modus "Upsamping (HQ) brings more sharpness but destroys faces and hairs much more.
The modus "Upsampling (LQ) works for me on no footage. In each rendered video it adds spots which were not part of original footage.
My wishes so far on improvments:
add better face and structure regocnition
implement slide controls for sharpness and noise reduction to be more flexible in order get best results
implement audio
improve the modus "Upsampling (LQ); e.g. also here treshold sliders for manual adjustments
I’ll chime in on the HQ-CG mode seeming to produce better results in many cases. It sure handles lines and edges very well, anything where aliasing would be an issue. Seems it handles such potentially jagged edges (step-stair appearance on diagonal lines) very well.
But then it seems to handle textures less well in some situations, meaning it smooths out fine details somewhat between such lines and edges, making some surfaces look plastic. The HQ mode seems to keep textures better, but seems to have aliasing-issues with lines and edges.
Suggestion:
Some sort of combination of the two modes might do wonders. Maybe a slider to add in textures and fine details from the HQ mode into an otherwise clean/smooth HQ-CG upscaled video would solve this. Would make it possible to have nice upscaled edges without aliasing issues, and keep more detail where they might otherwise get smoothed out.
A slider for sharpening might also be good.
Regarding the sharpening; as it is now I get pretty good results by upscaling to about 1.5X the target resolution, and then apply some sharpening and scale down to the target resolution in Premiere/AE or something.
I further tested several footage.
Often the deblocking and denoising are very impressive.
Also smoothen and straighten of edges are nearly perfect.
In often cases the denoising and smoothen of the footage is too much and the video looks too unnatural on the end, cause fine structures are gone.
Here it needs some adjustment sliders in order to increase/decrease
denosing
deblocking
sharpening
In other cases of e.g. grass the tool adds unnatural structures into the grass.
Is a next update so far with first improvements foreseeable?
I have tried it with various inputs and so far I am very impressed with the results. One area where it doesn’t really work all that well is digitalized analog video. I have quite a few videos from analog sources which show all the signs of analog footage (wobbly image, scanlines, combing). If the product had a way of upscaling these just as convincing as it can do upscaling of videos with lots of compression artifacts that would be a real game changer for people doing restoration of old footage.
I made the same experience. I tried with different settings some old Hi8 video clips, which I digitalzied years ago.
The improvements are on a very small level. Details are not really improved. Also edges are not stabilized after rendering.
Same with old VHS footage.
Obvouisly at least DV footage is required as minimum requirement in order to get improvements.
I received this morning a Topaz Newsletter, that a new version for Video Enhance AI is available.
Is this true?
If yes: which features was changed and which features was added?
Some additional info on how the program handles certain video is really needed for professional users. Mainly:
The program will now open interlaced video. What deinterlacing algorithm is used? This actually isn’t true… most DVD’s are actaually progressive but flagged as interlaced. It would have been nice if Topaz had communicated here.
The program will open anamorphic video (i.e., DVD) and desqueeze 720x480 or 720x576 according to the metadata (for example 720x576 at 1.78 = 1024x576). What sort of algorithm is used for the desqueezing?
In looking in the tldb directory, I can see there are 1x, 2x and 4x training data. How does the program determine what data is used? For example, if I scale 1024x576 to 1920x1080, that’s 188%. Is it using the 2x model data? At what percent does it decide on which training data? If the final result is 202%, does that mean it switches to 4x?
For final resolutions that are not 2x or 4x, what scaling algorithm is used to get to the final resolution? For example, if I scale 480x360 to 1920x810, that’s 2.25x. Does that mean it uses the 4x training data, then downscales the result, or the 2x training data, then upscales the result? If so, what downscaling or upscaling algorithm does it use (lanczos, spline36, spline64, etc.)?
03/22/2018 - Questions still not answered. Why won’t Topaz communicate?
It’s clear that Nvidia has more market than AMD, but with RDNA arquitecture that tendency is changing, so AMD market is growing.
Just for curiosity it would be good to know why AMD GPU isn’t supported. Do Nvidia and Intel implement something that AMD doesn’t support? Is this just for Video Enhance or is something general of the AI engine in all the Topaz apps?
I’m sure there are many things being worked on to improve Video Enhance AI. So, maybe the concerns I will express here will be non existent after additional versions come out.
After processing a clip with Video Enhance AI, I’m not able to do any further processing of this clip in my video editing software. I’m using Power Director 16. A Video Enhanced AI clip can be placed into the timeline just as any other clip, but I can’t use any of PD 16’s tools to do any of the many things it allows me to do to an unprocessed clip. I can’t even edit the length of the clip. At this point, anything I need to do to a clip needs to be done before it even gets imported into Power Director 16.
