This is really technical stuff — how much do you know about pattern training models?
Or did you mean to ask why 4x was chosen instead of 2x? The answer for that is probably as easy as people have higher expectations as the modern displays continue to improve versus the archival material they need to use.
In my experience the current state of the models isn’t really satisfactory for x4 upscaling. It will just look too artificial/weird in many places.
My favorite application is for noisy but otherwise high quality material where TVAI really brings out the details or 2x upscaling from good or better 1080p sources which can give a nice uplift in quality.
For my low quality content I guess I will have to wait for some more years.
I’m getting the Topaz watermark. Wanna update my beta license ? 
I tested the alpha, but nothing works on my side, even with 4x upscale it fails all the time. To review this alpha version for a next update of this one more stable.
This is the same error I get no matter what I try with RXL.
2024-08-30 09:46:20 13604 CRITICAL: Model couldn’t be run for outputName [“out”]
2024-08-30 09:46:20 13604 CRITICAL: Unable to run model with index 0 it had error: no backend available
2024-08-30 09-46-20.179 Thread: 20764 Info FF Process Output: 17
2024-08-30 09:46:20 13604 CRITICAL: ONNX problem: Run: Non-zero status code returned while running DmlFusedNode_2_8 node. Name:‘DmlFusedNode_2_8’ Status Message: C:__w\1\s\onnxruntime\core\providers\dml\DmlExecutionProvider\src\DmlGraphFusionHelper.cpp(1060)\onnxruntime.dll!00007FFE3896A5A9: (caller: 00007FFE389F174F) Exception(2) tid(3524) 887A0001 The application made a call that is invalid. Either the parameters of the call or the state of some object was incorrect.
Enable the D3D debug layer in order to see details via debug messages
This is just my guess: I think they are doing 4X models because of the block artifacts people kept finding on models like Apollo while running them on 4K. Maybe it has a way of avoiding those artifacts.
I am not convinced that is the reason.
if that AI was aimed for High quality source content it could be argued that this might be the case. but we saw it with Iris, & Reah which are for LQ / MQ. hardly any users I know that have LQ/MQ at SD bump it up to 1440p or above (x4).
There is some tech’ reason behind it, on why x4 is easier to implement. and I am intrigued to know why?
============[IGNORE THE BELOW RANT, ANGRY RANT NULLIFIED]==============
It really pisses me off that the alpha does NOT respect the custom directories set for previews, as in it defaults to the C-drive, where most use SSDs, and SSDs have a FINITE WRITE, and is bricked when it has been exceeded!
And considering the fact that I JUST exceeded several gigs of data from only 3 previews, this is extremely infuriating as I cannot use the alpha as I don’t want to BRICK my SSD!
Every time I set the directory, it STILL USES THE C-DRIVE even when specifically having preview storage turned off!
EDIT: I have no such issues with the general release or beta, curiously enough…
===========[IGNORE THE ABOVE RANT, ANGRY RANT NULLIFIED]============
^I fixed the issue, forget about the angry rant, it has been nullified by my solution, as I simply C&P’d the default folder to my desired directory, and deleted the default!^
The great news is that while the new RXL is magnificent, (Outstanding work from the dev/research team, you guys hit it out of the park!) it will not work with videos with vertical framed vids as in: 9:16 for example, yet works fine when rotated 90 degrees to 16:9, strangely enough, no matter what codec I use…
Oh yeah, logs incoming I’ll have to set things up though…
Hopefully @tony.topazlabs is reading this thread so he’s aware half of us can’t run the RXL model due to the errors every time!
Correct, x4 is useless for me. my material reacts well at x2 , but going with x4, is already pushing it.
I don’t mean that, I mean we literally can’t run the model due to errors!
I am less concerned about that, I am sure it would be sorted very soon.
Hopefully they release 1x and 2x variants once they get the model refined.
which brings me back to my original question. what is so special about x4 that they always release AIs 1st at that multiplier? what x4 has that x2 or x1 doesn’t? what is the secret of x4?
That’s way beyond my knowledge, machine learning, model training, something something…
@nipun.nath Could probably tell us in couple sentences though!
I’m here! Logs for these crashes would be super helpful:
@Imo already submitted logs and we have the same system specs so that should help but I can as well…
Log submitted!
Thanks! It looks like the 4x upscaling of that clip requires too much VRAM for this GPU.
RXL will scale to 4x size internally even when set to another output mode, so during processing it is actually trying to render a 5760x4320 output and running out of memory.
Updated: pasted wrong resolution at first
Another option that could help is to use this version of the rxl-1.json file. This one has the largest block size removed and may reduce VRAM utilization:
rxl-1.json (11.4 KB)
Even on a 24GB 4090 card? I guess I need to just try 2x and 1x then!
