Can report the same issue as reported by others above using a high VRAM card like the 4090. Initially get a higher speed of 1.1fps when processing with Starlight mini, but it goes back down to the normal 0.7fps in less than a minute, while using only 14GB of the 24GB VRAM capacity.
1 Like
Mine 3070 and 4070 work the same. I think itās just because of the speed miscalculation.
The initial model loading is accounted as a processing start point, which causes an unnatural fast speed at first. But we all know that the real processing is 0,7 fps (0,1 fps in case of 3070).
1 Like
still having issues with portrait rotated low resolution iphone video (with starlight mini)
same here without high Vram card, intial goes faster then itās slow as usual.
After some testing, I can once again confirm that Iām not seeing any measurable improvement in render speed on my 4090. VRAM usage also remains unchanged ā the full 24GB of my graphics card still arenāt being fully utilized, even though 100% VRAM usage is now allowed. On the bright side, using 100% VRAM no longer slows down rendering. Yayā¦
I really hope this is just a bug. Iād find it pretty disappointing if an RTX 4090 doesnāt qualify as āhigher end hardwareā.
1 Like
any plans for rotation stabilization? seems current stabilization is focused on panning (left right up down) or jittering.
Hi, 100% utilization will be supported in the next version, cause that operation may affect lower end cards, we need to be very careful.
2 Likes
Ok, I finally can test this on my 4090 build, I was restoring my old DV footage couple weeks and keep occupied by long processing time.
My footage are all 720x576, and use x2 to 1440x1152 with starlight_mini, using the 88% max GPU RAM on the previous version 7.0.0, it had a stable rate of 0.3 fps. With this update, I set 100% max GPU RAM, and it seems to be 0.6 fps for the first couple of minutes and slowly dropping back to 0.3 fps. Hmm, it does load faster when starting to export, though.
Does āImproved performance on higher end hardwareā mean that the model loads faster? Or does it mean we should be seeing faster export speeds? It seems like most people are not seeing any speed improvements.
I think thereās also a bug with choosing the resolution. Sometimes if you choose 2x,3x or 4x, then you click Original, it will still export at 2x,3x or 4x. The only way Iāve found to fix this is to switch to another model like proteus, then immediately switch back to starlight and then export without clicking anything.
1 Like
A 4090 would give 0.7-1.0 fps for such tasks while the 4070 gets about 0.3 fps.
So, yes a higher specced GPU does seriously improve rendering performance.
Can the pre-Starlight mini checks not also be used as a means to ācalibrateā the optimal % of VRAM usage based on the amount it detects? Would be more straightforward than users having to do experiments/guess work.
For me at least, Iād want a āset and forgetā option, where it sees Iāve got 24GB, and it sets itself to the highest % possible that WONāT result in spill-over to swapping with any system memory. (With a few % margin of error built-in to allow for basic OS display that may already take up a little of the VRAM.)
3 Likes
Just echoing other comments about no real change to the speeds on my 5090. Tried to process a 10-min video in Starlight Mini just for fun and it told me to come back in 3 days and showed me a speed of 0.2 fps. I will say that my GPU utilization was near 100% and my VRAM was nearly maxed out, which was nice to see. Then I went back and did a Proteus run and had the same problems as I had in the last version where there was low VRAM usage and poor GPU utilization with the latter hovering around 33%. Still not really sure what to make of that.
1 Like
Are you guys seeing better visual quality at x3? From my tests x3 usually tends to produce the better visuals (especially eyes) than x2, but itās a lot slower. This is mainly when characters are in the background. If they are in the foreground a x2 is still quite good. I may complain about speed a lot, but the quality is still phenomenal. Congrats again to the topaz team. This is awesome stuff.
3 Likes
I also have the impression and I donāt want 960p results, so I have to scale it to 1080p afterwards and this step is not model enhanced. If they gives us a 2.25x model, then we could go from 480p to 1080p
1 Like
x3 is just too slow to use directly.
Has anyone tested whether doing Proteus to 60fps first and then Starlight Mini gives better quality?
Or is the other way around betterāStarlight Mini first, then 60fps with Proteus?
I am currently rendering a file and although the āLive Exportā toggle is set, the export is not showing the live version. I have confirmed that the work is being done, as I can look at the file as it is being rendered, but this feature is not currently working (as it had in the past). Is this the case that you were referring to āThatās coming!ā eg the next edition : 7.0.1a?
1 Like
You absolutely do NOT want to interpolate your video to 60fps before using starlight. It would double your export times. In fact, some people suggested lowering fps to 25, processing with starlight, then interpolating the fps back to 30. To me thatās just too much work. Iād rather wait a few weeks/months until there are speed boosts or just chop up my file and run the export in 10 minute chunks whenever I have free time. The 5090 is an oven so I donāt want it running all the time. It generates so much heat.
2 Likes
Yep, Iām also starting out with .5, (that was exciting for several seconds) 
, then it drops to .3
1 Like
On VHS-C camcorder tapes, I capture interlaced at 640x480 @29.97 FPS, then use Starlight Mini 2x to get to 1280x960, then use Iris LQ and stabilization to scale to 1440x1080.
Starlight Mini seems to handle removal of the interlacing artifacts.
The results are quite remarkable.
2 Likes
Like you, I am using an RTX4070 and am considering upgrading to an RTX5090, but I have a similar impression of the results I can expect from the upgrade.
If upgrading to an RTX5090 would make it at least three times faster, I would upgrade immediately, but based on some of the past posts here, I donāt think thatās likely.
Another concern is the fact that while the Starlight Miniās SD to FHD enhancement is of excellent quality, the results from FHD to 4K enhancement are not as good as it could be.
I spent five days enhancing a 5-minute 1080p video to 4K at 0.0fps, and the results were disappointing.
So I have no choice but to hesitate to upgrade to an RTX5090.
2 Likes
