Thanks will test this later today!
There are several cases in TVAI where RAM or VRAM are not properly weighted at process start.
So the error occurs when one is blocked and the OS can no longer handle it.
I have mentioned in the past a potential memory leak or relocation issue. This should not happen. It either happens immediately or not, but many cases lead to a crash after hours of work.
Indeed, many internal x4 related items that exceed the maximum stream of 8K lead to an unexpected (bad) termination. TVAI should have an internal test before starting a process, what do you think?
Tony you’ve been ignoring messages about the Artemis Aliasing/Moire model.
This fix worked for me - I can now run RXL at 2x on a 1440x1080 source.
About the RXL model, looks good so far but a little heavy on the over-smoothing again, I think the default preserve detail needs to be cranked up a little bit.
It’s in C:\ProgramData\Topaz Labs LLC\Topaz Video AI ALPHA\models
It’s a hidden folder by default so just copy paste that path into the folder location field and hit enter.
1 Like
I see yes sorry for RXL 1 it is clear first Alpha that has RXL and Hyperion Alpha 3 also very clear the 3th Alpha release that has Hyperion . So it is a versioning system of models
1 Like
still did not answer the question. why x4? and not let’s say x1 or x2. what is so special in x4 we see almost ever new AI initially released only in that multiplier (x4)? I am not talking about versioning. I am talking about upscale support.
1 Like
Because supersampling follows a logarithmic progression, or even “supering” a sample.
The x4 is used since many years in all visual aspects because of its visual efficiency. It’s neither, nor too much (ressource consuming). It has nothing special with AI : See the most usual use : antialiasing, it is x4 since 20 years, and only changed starting with Nvidia which used AI to solve the “problem” quicker.
Only anysotropic filtering has been pushed ahead beacuse of its 3D isometric nature than ruined image if not used.
Moreover as Moore’s is almost unchanged since 50 years, this multiplier remains the same because of same cost/performance ratio.
Some more complexed algo. follow pure mathematic methods : linear regresion, Fourier suite, Laplace interpolation aso …
3 Likes
Have tried various 720p sources (1080p gives just an error @ x4 on a 4090) and unfortunately haven’t found a single case where Rhea (or even worse RXL) wouldn’t look artifical by overemphasizing existing- or adding imagined details.
2 Likes
I did watch the monitoring of MSI afterburner while starting RXL.
GPU = 4090
Vram Memory usage jumps to 18 GB (1080p input).
Then it wants to start processing (wattage usage goes up a few %) then it does drop the error.
Then I stick to my previous answer they use nowadays x4 because it is possible with the Nvidia cards the x4 is a way to increase the frame 4 times internally and do a more sophiticated but slower enhancement then when this is done in the old way for legacy nvidia cards. that is why it says it will geneerate a lot of resources these models x4 are primarily to look for details in the frame then enhance . The later models look for details enhancement with requires 4X lots of resources than scale back to the desired output. So in the future with even stronger cards and 8 K becomes default we might see 8 X but for now 4 X internally and 4 K output as a max is just fine
I think I know what you mean, but it all really feels irrelevant to me when I try 150%, 200%, 250% and almost always 150% produces the best result. I then re-scale using classic tools and maybe some light sharpening if appropriate (lapsharp brings out the best in Proteus, I find).
And anyway sometimes you have to aim for 100% from a smaller intermediate source, because lots of streaming-quality material is far worse than what was acceptable on DVD and digital TV.
One of my favourite tricks lately is running Gaia before Proteus, if the source was ‘MQ’ or ‘HQ’ at the source resolution (I hate those terms and need more scientific ones). Always have to run test comparisons, though. TVAI has increased my workload not decreased it, but only in an amazing way.
My low quality sources, say 480p, and highly compressed … they’ll go up to 150% or 200%, but the bad stuff doesn’t get much further than 720p often.
I’ve just discovered my LQ stuff really loves Artemis (strong or medium, YMMV) and I’m throwing out a lot of Proteus work I did a few months back. It wasn’t great anyway. But it’s fascinating how a 1 megabit/s video can need it, whereas a 2 megabit/s video (probably made when ADSL2 was replacing ADSL1) can cope with a much stronger model.
I do not speak for Topaz, but I am pretty confident that work on those models has stopped over two years ago. The work has moved in favor of models that are more versatile—like Proteus, Iris and Rhea.
The one-size-fits-all design of the Artemis models limits them to a pass or fail use case. Of all the DVDs I have enhanced (over 1000 episodes of shows and such), 0% of them come out better using any of the Artemis models, than if I use Proteus. Because, if Proteus comes out bad, I adjust the settings until I find the most optimal output. Not perfect, but easily better than any of the the Artemis outputs.
If Topaz has come to the same conclusion about the Artemis models, your request sounds like you want them to work on regressing the progress they have made.
1 Like
I also noticed loss of fidelity. With Rhea, @nipun.nath was able to increase the preserved detail level during that alpha. RXL, while better, still needs an adjustment in that category.
2 Likes
Thanks for the updated model, it actually works now!
Tony generously provided a fix, just download the new model in this post!
I’m REALLY diggin’ this model!
Further testing revealed that they may have solved the problem with patterns being properly displayed while in motion, for example if someone has mesh patterned leggings and the video featured the subject running, then the mesh patterns that naturally flow along with the runner’s movements unbroken along the contours of the leg’s muscles in motion.
Whereas without the model, the patterns while still moving along with the runner’s legs would still follow the contours, but the pattern itself will break and artifact!
I’m gonna do more tests, but it looks like they’re on the right path!
2 Likes
Can’t wait for my current encoding to end to follow your steps 
Texture of clothes should be a pain to train, there’s such a there is such a variety possible.
On this other side, arms and legs movements should be easy. Training models on legs/arms is done sincem any years.
What about cars, windows, waves, buildings, grass, fields ?
1 Like
I’m sorry but I gotta say RXL for that looks substantially worse, not better. Not talking about the text either.
1 Like
