Whatever model I use in the new beta software, Topaz beta is making my GPU heat much more than the regular Topaz video app. On my A6000 PRO Blackwell, it is at least +8/10° Celsius (compared to the normal app).
Anyone else with other nvidia GPUs noticed the same ?
Hello Michael, Go to Topazlabs.com and sign in (click on the little guy in the top right hand corner).
Scroll down to My Products. If you see that your product or products are active for 2026, that tells you that it is going to be renewed next year. If not, I think you go into Manage Subsription and do changes in there. Hope this helps.
Right. It makes sense to have two models. One that uses cuda for nvidia cards and a version that doesn’t for mac/amd/intel etc.
Since this newest beta came out, I haven’t used this Proteus version til today because I’ve been testing the Sharp and Mini. I’ve been using Proteus in v7.02, so I checked out Proteus in this beta release and it seems (unless my eyes are playing games) that Proteus’ quality has gotten even better in this beta. 
Yes, i noticed the same (RTX 4090), in other Versions (and other software) i get about 84°C, with starlight sharp i (last week) got 90°C!, the card lowered it’s seed a bit to compensate. (We had high temperatures here). I then placed a fan next to the case (the glass wall of the case was also very hot), and then the GPU slowly became cooler (and faster again).
M2 Ultra (24 Cores, 128Gb RAM), SL(M) for 6 min video (720x480, scaled x4, to 2880x1920) was showing 0.4 FPS, then 0.3 FPS, it took 1d55m42s to render the 6 min file BUT rendered 1 MIN ONLY in that time period. Also preview is broken, one side shows very small and other side scaled up.
An other Speed and Quality test with the new Starlight Mini:
(RTX 4090)
Input video: 640x384 391 frames (15 sec):
Starlight Mini (Old): 7min 29s (=449s)
Starlight Mini (new): 5min 36s (=336s)
output as TIFF, i looked at the timestamps of the output frames (New starlight needs longer to load)
=> New starlight Mini is about 1.3x faster.
An other observation:
The quality depends heavy on the input resolution (should be divisible by 16?)
My first attempt was 640x360, the quality was horror (blurred, double lines, seems as if two frames were blended together, then i upscaled the input height so 384, the quality was much better.
Starlight-Sharp:
In the options menu is “Input resolution”, but this is wrong, it gives a minimum resolution of 960 width!
You can also use the free software MSI Afterburner. It lets you tweak the fan curve of your GPU ventilators, and the default curve is often enough to keep your card under 80°-82° even under heavy load. It is very handy and easy to use 
So they nerved down SLm quality for more speed? Is this the case? That’s exactly what I was afraid of. It would be ok if we can switch between more speed and better quality then it’s a benefit when you can choose, but that’s not the case. So the good quality is only available in the cloud then?.. 
Dakota did mention this early in this thread…
“The intention for this beta is to test optimizations. If you read the post you’ll see we’re aware of the quality difference.”
It’s bad when the card runs hot, because then cooling isn’t sufficient, but it’s desirable TVAI max out our GPU. Check also case fans, can make a big difference, the hundreds watts have to leave the case. And yes I also recommend Afterburner to everyone anyway.
additional I installed an extra slot fan under my gpu, which is available on Amazon or elsewhere It sucks the hot air from the GPU directly outside.
I had a couple mistakes in my first results chart. I tried updating the snapshot and accidently deleted my post. Here is the updated results chart for Nvidia, AMD, and Apple devices.
Alledgedly this is not the case - but the bad quality only due to a preliminary implementation of the model.
It also has to be that way as with the current “quality” of SL mini you could as well use AI Arty (ugh..)
I ran some additional tests on a Radeon 7900 XTX, and overall the build runs without issues on this hardware. As expected, the visual quality isn’t great at this stage, but since the current version is mainly about confirming functionality, that’s fine.
One issue I noticed is that the time remaining counter consistently underestimates processing time. Refining the calculation logic would make it more reliable for users and help with workflow planning.
Performance on my system generally ranged between 4.5–5.5 fps, with occasional spikes up to 10–11 fps. It would be helpful to know whether Starlight’s performance is affected by scene complexity. For example, if simpler or more static scenes (with fewer motion vectors) allow for higher throughput. If that’s the case, exposing this information could help users better understand performance variations.
At the moment, Starlight cannot be used as a second enhancement step. Adding this capability would improve flexibility, especially for workflows where users want to chain multiple models. Currently, my preferred approach is Nyx for denoising, followed by Proteus, Iris, or Rhea for detail enhancement. I’ll be testing Nyx XL next to see how it integrates.
Suggestions:
Would love to see Topaz plans for 2-3 years about Video AI? For era when Nvidia 6000x would be released
Would be there quality like 10-20 time better, like today if we compare what we have was have at 2 years ago?
Is this still actual?
Did buy a sub for GPAI an TVAI yesterday.
I just need to dig for a good footage on my system to test the models.
I imagine Starlight to be more like Proteus 4, in that it will give better results when you give it the original source, since it has more to work with.
My fully loaded M3 Max neither SM nor SS work for me in 7.2.0.2.b. SS is greyed out and SM spins up kicks in my fans to the max but doesn’t seem like it’s processing and then I have to shut down my M3 Mac because I’m afraid it will fry. Doing nothing!
So, with the shifting of everything Starlight related to a “Pro” license and the sudden move to a subscription only model:
Will AMD and Apple users really be left in the cold with only the quality-wise unusable early Beta (more an Alpha as it seems) while owners of Nvidia GPUs have a nicely working model?
It’s not as if the Mac licenses were cheaper than Windows ones and Starlight was promised and also introduced before the licensing change, so I think that we Mac/AMD users definitely should get a fully working perpetual Starlight mini as well.
Otherwise this would be a major betrayal of those that bought TVAI perpetual licenses because of the promise of SL mini.
