Thanks for the new beta, looks like I’m gonna downscale a bunch of videos to fit 3x upscale up to 1280 for proteus 4, will be interesting to see the results! I also went the cheeky route and would crop-in-prog to get a size small enough as to still be used by Pr4 while keeping the output at 1280, as 1280-1920 is my sweet spot at the moment.
Bug report:
It seems the “preview bar” is missing.
The “preview bar” is the white bar on the time line that shows what part of the video will be previewed. And it’s missing in this beta.
Proteus v4 4x upscale still looks quite bloby, especially in scenes with compression artifacts and high detail. E.G. A scene containing plants swaying in the wind.
Proteus v4 2x and 1x show similar blobiness. But this is probably excepted since they haven’t been fully trained.
The video I’m working with is low quality, and I am zooming in quite a lot (Using the 100% viewing option in TVAI, which is 400% on the original footage). But you can see, Proteus v3 seems to be taking the detail that does exist in the original footage and is enhancing and upscaling it. Proteus v4 is taking that detail and turning it into blobs
For couriosity the temp files never get deleted. This happened to the versions before too. Also as mentioned before the preview almost always gets stuck shortly after beginning the processing. The before/after frames do not match btw.
I tested Proteus v4 v3.2.6.0.b.
The afterimage noise has not been fixed at all.
There is a tendency for afterimages to appear in the green component, which is also the case with Proteus v3 and earlier, Gaia, etc.
no change or usable Prob-4 on 3.26 compared to the previous version of 3.25. it still removes to much details and the picture is way coated with a layer of Vaseline. it’s not usable in its current state.
when trying to use Proteus V4 the download starts then an Error is thrown asking to get the logs for support and contact the support.
Please consider this my contact. Thank You logsForSupport.tar.gz (189.1 KB)
It is almost impossible for me to compare the before and after image because either the preview gets stuck shortly after starting the processing or the before and after frame do not match.
Proteus 4 is worse than before, unlike with version 3.2.3.0.b which allows for the advanced model P4 to generate face reconstruction as low a resolution as 1080p!
P4 for version 3.2.6.0.b’s face reconstruction with 4x is pretty dismal and blotches out the eyes as well as superimpose the Alt-generated hair reconstruction as it doesn’t fit with the input video at all, it’s like someone pasted a badly drawn sketch onto the input’s hair!
It also does nothing with facial reconstruction for anything below 4x, which is per description, but at least with the previous version I’m not choking out my ram with large 4x conversions with better results to boot!
I’ll follow up with some tests/logs once I reinstall it, for now I’m using 3.2.3.0.b, unless there’s a way to have two betas installed that is!
I’m just glad it’s still a beta and not the general release, though the blurred out limbs when using “reduce jittery motions” with the stabiliser is still broken, whereas Virtuadub’s DeShaker 3.1 plugin far outclasses TVAI’s solution, I mean they were on the right track, but now it’s on the same par as power director!
It’s incredibly disappointing to say the least as trying to stabilise heavy zoom and roll is worthless as it simply doesn’t work, which means you’ll only be able to stabilise via the horizontal and vertical plane, whereas VD/DS3.1 handles it no problem, the only issue is just how long it takes, which is substantial compared with TVAI…
EDITED to add: Don’t get me wrong, I love the program that is why it seems if I’m a bit harsh, as I do not want this program’s development to go down the road of power director, or its like.
There’s a big problem : each time we change a setting after making a preview the position of the video is reset to frame 0. This makes completely unusable version to directly change and compare different settings with Proteus models.
Same when clicking back to input after making one preview.
After checking, this only happens when we set to frame number the timecode display format.
Even though it says in the description that P4 doesn’t work with resolutions that are not scaled 4x, I have had the filter actually use facial reconstruction on resolutions at only 720p!
This was, however, from version 3.2.3.0.b, and not this version, which does nothing 4x and above, so my guess is that this is just a proving ground for this model, as the degradation with the quality of this version of the model is too great.
I’m almost debating whether to skip this beta, as version 3.2.3.0.b does so much better than this one, as I’m currently using that version!
With the frame numbers as option, the moving of the frame numbers when clicking in box and random number changes is gone, which is good, but it still doesn’t select the frames properly:
On export, exporting an example of frames 13000 to 13100 exported framed 12999 to 13102 on one example test. On another, it finished a frame short. Basically you cannot trust that the frame numbers in the export window are actually the frame numbers being exported.
After clip selection of frame numbers, the left is clipped on the preview but right is not. So it doesn’t “look like” the clip applied, but when exporting it stops roughly where its supposed to, so the clip is working, but preview now suggests otherwise.
The adding of an extra frame doesn’t seem to be happening so much in this version, but on 50% of tests its now dropping a frame instead. So rather than renaming backwards to remove the additional frame, you have to rename forwards to add one.
Progressive input 704x396. Upscale to FHD and crop to 16:9.
Several relative to auto Proteus settings previewed. In tweaking a few parameters I got several frames with green errors over the video.
In the two attached images I show one frame with and without the error.
