You’ll need to check the settings for MeGUI and x265. Switching from non-drop to drop frame timecode after encoding again probably would not be caused by the CFR output of Video AI.
The video you are bringing into x265 is showing Constant frame rate in MediaInfo?
I’ve not checked the MediaInfo and I’m currently back in 5.3.6. and replaced the video files. I will check again once 6.0.5. releases.
Loop button still missing from this version? Makes the preview function nearly useless when doing any preview under 10s. Really hoping for this feature to come back AND be the default behavior going forward!
I won’t ever recommend Rhea for a human face restoration. It creates a weird hallucinations of faces where there shouldn’t be ones.
It also tries to glue wrong faces to the blurry people in the background.
My workaround is Proteus for base layer and Iris at 40…50% opacity as a second layer in Premiere Pro.
Yes, it requires two separate encodings, but the final result is better then any model can achieve separately.
I’ve explained my approach in this post
It’s suitable not only for crappy videos, but for good ones too. Now I can achieve results that wouldn’t be possible earlier.
And another good thing is that I can control the influence of Iris over Proteus for different scenes too. Sometimes Iris is better - well, let’s crank up the second layer opacity!
I do the same with Iris + RheaXL but then for Iris 60% or more opacity, otherwise the stripes and artifacts produced by RheaXL will become visible. By increase gamma/contrast/colors with the source, the blury thing RheaXL sometimes does is gone or minimized.
Sometimes this gives stunning results but wide shots, faces often are still unusable, so I have to work here with Proteus, but the scene based model choose thing is super time consuming. How RheaXL keeps details is amazing, but it is Beta and very hard to use, the model needs a update
@tony.topazlabs
Is it really that difficult to implement a TensorRT version of the AION model? You just convert it and make it work in VEAI. There are many of us who prefer it over Chronos and Apollo. As it is now, it’s too slow for larger input resolutions, even on a GPU with 24GB of VRAM.
We want models for low resolution videos, models that stop making faces, eyes and monster teeth… less interfaces, decorations and clouds and more models, since you haven’t updated or released new ones for over a year.
I often hear the hardware is not sufficient, xy Cores are needed and Vram xyz. Is it technically not possible on lower Hw or “it’s too slow” and who decide it is too slow for me? I can live with a 1 fps model limited to 1080p when it is great, so I can decide use it or not.
If a 10s uploaded snippet to the Cloud takes up to 40 minutes, I call this also slow. My computer is bored, please give him something to work, a new local Light-Star^^ model, a revised Rhea, or IRIS MQ V2, but please Topaz Labs do something.
YES! Something that doesen’t produce monster faces.
will we see 5000 series native fp4 support advantages for topaz?
Has anyone noticed if the weird random stuttering in video encodes that was introduced after 5.3.6 was fixed?
Continuing the discussion from Topaz Video AI 6.0.2 + 6.0.3:
Sometimes the brutal truth is very refreshing.
If the big Cloud dream collapses, they might come crawling back
Iris LQ V1 with a high setting for Recover original details (and also high positive value for the Antialias/Deblur slider in order to still get a more defined picture).
But,yes, an update here would be great (since we kinda have an updated IrisMQ, they just called it Rhea).
Not sure what is meant by “hacked” here, but right now we’re offering early access to Starlight under our render cost.
We do expect this to be resolved as of 6.0.4, definitely let us know if you’re still seeing cases of this!
We are working on this, but will likely be after the release of RTX 5000 support for all current TensorRT models
I have a Proteus upscale of a reference clip that worked in 5.36, and jittered in later versions. In 6.0.4, the jitter is no longer happening.
What exactly does the “recover detail” slider do that is different than the “improve detail” slider"?
Recover original detail does kind of a blending with the original video, thus defeating the quite often overly smooth / „plastic-like“ look the AI models create.
Improve detail = control the strength of detail recovery and texture regeneration
Recover detail = blend back some detail from the original input to improve oversmoothing and waxy appearance
