Clip 101 Pilot_edit.mp4 and log file uploaded. This is from an old series called VR5 with Lori Singer, virtual reality in… 1995. We’re still not there:-)
Mac doesn’t appear to be the only one effected. Upscaling on Windows shows tiling artifacts.
Notes: OS: Windows 11 GPU used for processing: RTX 4090, driver 531.79
The issue is most noticeable on the first frame of the video, but it’s still noticeable on other frames
The videos I tested are 1920x1080 with a 4x upscale with Proteus v4 on auto
3.2.7.0b, I still get artifacts with a strong sense of afterimage.
It tends to appear in the green component, so I am wondering if something is wrong with the color processing when learning.
In the case of live-action video, artifacts may be less likely to appear because the colors are not as extreme as in animation.
However, it cannot be said that artifacts do not appear in, for example, strong lights in live music performances or special processing in sci-fi dramas, and we hope that this will be thoroughly verified and addressed.
Proteus v4 appears to use information over time, presumably in attempt to increase detail and temporal stability.
Looking at some live action videos with objects moving in front of a background, there is a little bit of ghosting similar to what you’re seeing. But there’s also a transition period. When an object passes in front of another, “temporal history” is lost on the background object, and that can be seen as that part of the background having different “enhancement” characteristics compared to other parts of the background. Over the course of a few frames it falls in line with the rest of the background.
Proteus v3 shows similar stuff, but it’s not as pronounced as Proteus v4
Although, this could be related to me using the “Auto” setting which adjusts settings over time.
Proteus should be doing noise removal etc. based on the previous frame’s information even before V3. I don’t remember where, but I remember seeing Topaz make an official statement to that effect.
As you say, it also occurs in V3, but not as pronounced as in V4.
As for my test, I did not use the automatic configuration.
In the CLI, I used “tvai_up=device=-2:model=prob-? :scale=?:compression=0.2:details=0.2:blur=0:noise=0.1:halo=0:preblur=0”.
Proteus v4 still appears to have a tendency to make a video softer and blobier than Proteus v3, especially when working with lower quality videos.
Higher quality videos are still softer, but it’s not as bad.
Here’s some examples:
Note: Original on the left, enhanced on the right
Note 2: Also, all these images where taken from 120 frames into the processed video.
Note 3: Auto was used in all situations.
Watermark on grass 2x upscale:
Note: The camera and watermark is stationary. Proteus v3:
Trees 4x upscale (There were no tiling artifacts in this one):
Note: The camera is stationary and the trees are swaying a little.
This ones a bit difficult to compare. Proteus v3 has more detail, but a lot of it is clearly imagined and wrong. But from a “distance” or not zoomed in, it looks fine. Proteus v3:
On a positive note: Proteus v4 continues to be significantly better than Proteus v3 at dealing with and hiding compression artifacts from the original video.
I think the reason for the loss of tonality is that the image is converted to a JPG according to forum specifications.
When viewed on the NLE, the tone jumps are not as apparent as they are in the JPG displayed on the forum.
In Proteus v4, the noise is reduced, but the detail is greatly diminished.
The image quality is like an oil painting, as if a large amount of Gaussian blur had been applied.
