0.0.1b doesn’f fully utilize things, but it runs faster
0.0.2 and 0.0.3 both max out vram and considerable system RAM – which i am fine with – but the problem is that they both run slower fps than the 0.0.1b for pretty much everyone. There’s no improvement here. DO you guys have any suggestions or thoughts on how to get this in the othert direction… Go ahead and make the thing run full tilt, that’s what it’s there for.. but at a performance loss? Is this an issue being worked on?
I’m using 4090 based machines and 0.0.1b is still far outpacing the newer betas to the point where there’s no point in testing the newer betas because we’re talking about spending 12-15 hours on a 5 minute video vs 6-8 hours… There’s no incentive to even want to test the rest of the features when the hit is so dramatic (and it’s consuming more power to do less)
The weird thing is that I got 0.7fps on 0.0.1b the first time i used it now it seems to max out on this machine (also 4090) at 0.4 … and i’m running windows with basically every everything killed that isn’t necessary except topaz.
I even overclocked the gpu here and no help getting back up.. weird
I had up to 0.5fps with a 470 card now I have 0.2fps/0.3fps and doesen’t matter which Beta is use (i have all of them installed) so I think they did something with the models?! It is the same model folder for all beta versions when you don’t change it. Perhaps 0.6fps was to fast and they nerved it down to half speed…for faster speed pay cloud
Something fishy. It said it downloaded the model instantly, then proceeded to just give errors when I tried to use it. I repaired with the installer. No change. I uninstalled it but had it keep the user data. No change. Had to reinstall without saving user data to get it to work again. Man, I forgot how annoying the default options are on this program.
The output resolution is still random. Just tried doing 2X and it did 3X. Tried again on another video. Same thing. Changed it to 1X with square pixels and it did that correctly.
There’s no way to get it to consistently render to the correct output resolution still.
I am struggling to work out how Starlight decides in the output size of the video. I am upscaling standard def PAL clips - 720x576 or deinterlaced versions which are 768x576 and I am not getting the sizes I am asking for - I can type in a preferred width - say 1920, or preffered scaling - 250% or just choose 1920x1080 HD and I get 2304x1728. If I choose 1280 I get 1280x960 (which is correct for something that is 4x3 in the first place).
I have not seen an answer on the forums although others raised it. But what are the rules? Can Starlight only scale at fixed sizes?
Results are very good BTW.
Hi, there is certain minimum resolution we could process based one algorithm. Our engineer team is working on showing users the explicit target resolution.
Just getting started with this beta and Starlight mini and wow you guys have built something impressive! We have many old wedding and graduation videos that would benefit from this greatly though it seems it will come at very high GPU time.
On my 4090 I notice that the GPU is running at 100%, (in nvidia-smi, little in task manager) but is only using 130-150 watts. My guess is that the load is on the tensor cores only? Knowing the bottle would be a big help in choosing my next GPU.
same here mystical things happens. At my first trys I get only 960p (we talked about this), and then few day later I set 1080p again and get 1080, so I was happy but it didn’t last long, now when I set 1440x1080 I get 1920x1440, that’s strange
Since I downloaded an update (I think?) for the Starlight Mini model earlier — about 93MB — the model no longer works. No matter what kind of video I try to upscale with the model, I always get the same error message:
“An error occurred”
There was an issue when attempting to export video footage.
Under “See export error,” I only get:
“Please contact support.”
Collect the logs: Help > Logging > Get Logs For Support
Contact: Help > Get Support
There is enough free disk space available, and even videos that worked before are no longer processing. What’s going on?
ok I did some resolution tests. I does 2x 3x 4x…only
if you have for example 640x480 and set resolution to 1440x1080 it auto takes 3x, which gives you 1920x1440 (and 2x is 1280x960)
1440x1080 is not possible with a 640x480 source, when you have 480x360, then you can get 1440x1080.
Please change this: When the Starlight mini model is selected, only the scale factors should be made selectable and resolution must get calculated and shown but greyed out. Then in the output section, the correct resolution must be shown
I have tested this, you are right, but the difference is minimal. But the speed difference is large. In my opinion, i prefer speed. The “lower” quality of Starlight Beta 1 is awesome enough!
Original (4x upscaled bicubic):
The difference is not on the image. It’s on the temporal consistency. But we will expose the option to you guys. You will be able to choose less VRAM for Starlight-min
Yes, let’s say thank you to Kevin & team, sure, we nag around, it’s so slow, and there is this, and there is that, but all in all we are happy. I am just finishing a 48 hours upscale of an old 24 min NTSC VHS, and you made that possible, thanks again, even at 0.2 fps
Thanks, this information is very usefull, we know now more whats behind. Ability to set ammount of VRAM/Ram usage or for Enduser let’s say can set the speed doing upscale faster or slower, would be very nice.
