I get 0.3 fps for SD-HD Starlight mini encodings on a 4070 with 12 GB so your speed seems about OK.
2 Likes
Thanks
I love Starlight Mini, but I am getting weird eyeball-popping in people (seems to happen when they are a middle-distance away from the camera). Itās similar to the way that Iris redefines and seems to slightly enlarge the eyes on faces. Itās just distracting enough to be annoying. Iām hoping this issue gets resolved in future updates!
1 Like
V7 crashes on anything.
When you can start it, it crashes when opening a folder or drag a video into it.
If you come this far, it crashes on preview and so on.
I give up.
Windows 11 Pro, 64 GB RAM, Ryzen 5950x, RTX 4080 Super, no overclocking
Hello, Iām from Germany, and Topaz Video AI wouldnāt start after the update. Finally, using Windows compatibility mode (Windows 8) and running the program as an administrator helped amazingly.
Topaz doesnāt start or crashes if you arenāt online in the beginning to verify your account. That was a reason for me to change from wlan to lan.
1 Like
Starlight should restart rendering automatically as soon as launched. It will even patch the two runs together a produce a complete final.
To be reasonable, there is a real limit on just how much can be done with foliage (or other similarly complex and repeating image components) in DS9 episodes from seasons 1 through 4. These seasons were mastered in Sony D2 format, which recorded analog COMPOSITE video. Starting with season 5 they switched to DigiBeta which recorded digital component video.
As a result, even if you watched the original broadcast masters, that foliage would look like garbage. The only way to fix that would be to rescan the 35mm film originals. AI upscaling can do amazing things, but it canāt do everything (yet).
1 Like
Hello, dianac.
Yeah, TVAi should restart is the unknown, unfortunately.
Some users has reported that this feature doesnāt function as designed for them. 
1 Like
I only had it happen to me once. When I relaunch TVAI it resumed instantly. Didnāt know that didnāt work for everyone! 
What was your file type? This happened to me with a wmv. Itās very frustrating. I was going to try to see if itās possible to edit some code to get it to work. If topaz is using ffmpeg, ffmpeg supports wmv input and output, so Iām not sure why wmv files crash the starlight exports.
Iām totally fine if it exports to .mp4, but I want it to be able to read and process .wmv files.
Hello there. I am new to this software. Using Starlight Mini to upscale old 360p / 480p videos to 1080p. The results are mesmerizing, yet the process is painfully slow. Using an RTX 4090 I get around .3 to .5 fps convertion. Is there any advantage moving to an RTX 5090? Or is the software still not optimized to warrant that upgrade?
Hey Topaz Labs!
Idea!
Starlight mini is not really a professional grade productā¦.
Itās kind of made for the general publicā¦. For us idiots like me.
Not for most of the video professionals that I see on this forum. I mean, I use Video AI a lot, but I rarely tap more than about 1/10 of its potential.
Seems to me you were missing out on marketing this product to the general public at large.
I think you should offer a complete cut down version of Topaz Video AI. Call it āLiteā or something cuteā¦..with the starlight mini model as the only model offered.
Simplify the interface a bitā¦.. maybe add in a super easy editor for finished rendersā¦..mention what you can do with the āProfessionalā version of Topaz Video AIā¦ give a clear and easy way to just send it to your Starlight servers if they donāt want to mess with itā¦..and provide clear upgrade paths for those who want to move to āProfessional.ā
And sell it for $59
I think the normal, everyday Joe who will completely balk at the $300 licensing fee for the current professional versionā¦. and may be intimidated like crazy by the complexity of the existing programā¦..
Would jump on this for $59. And it would increase your advertising for the usage of your emerging starlight server based program.
Especially when they see the results!
2 Likes
I asked the exact same question a while back and several folks answered that it is really no advantage to move to a 5090 at this point.
And the software has not been optimized for that model.
They also pointed out that even if you do ultimately move and they do optimize the software somewhatā¦
Starlight mini is still going to beā¦ā¦sā¦.lā¦oā¦w
But you are correct, the results are incredible!
Iām thinking about just building a low-cost machine that all it does is render starlight mini material 24 hours a day.
Itās that good.
Iām going to hold off on that thought though to see if they come out with a solid Mac version as I also have video AI on a later model Mac.
1 Like
Has anyone tried this on H100, RTX Pro 6000 or multi (2, 3 or 4x 5090) hosted servers by the hour to try and get max performance per hour for one off jobs?
When will the new version be released?
Has version 7.0.2 been released?
2 Likes
My main problem is that the AI generates a pretty repeatable pattern similar to texture patterns in older video games which irritates me often even more than the original soft unrecognizable mish mash.
Thank you very much, mate! Will use that same idea of having an equipment dedicated to rendering 24/7.
Hopefully, it will be optimized for fasters speeds on high performance hardware.
1 Like
Am I the only one who isnāt getting particularly good results from Starlight Mini? I find that it enhances about 98% of most videos brilliantly but the remaining 2% will typically be a deal-breaker. It completely mangles most text, for example, no matter how well-defined the original is. It also creates brand new faces in low-detail scenes, meaning that characters will randomly change into completely different people and back again. I must have tried it on 20 or 30 low-res videos at this point and only a couple have been worth keeping (and even they had to be downscaled and treated with grain to minimise flaws that SM had introduced).
It seems to me that the model just isnāt quite there yet. The basic approach is game-changing but the implementation has a long way to go before it can provide consistently useable output.
4 Likes
This is known as āmonster faceā. TL hasnāt been able to eliminate it on their other models, so I wouldnāt hold out hope. Iād like to see them consider a manual intervention (from the user), who selects keyframes in the video and a source face and have the AI scale the face to that scene without distorting it.
2 Likes
