absolument magnifique 
I’m glad with the local rendering. My RTX4090 and RTX A6000 seem to have no trouble with it. The speed is a bit low (0.2 - 1.1 FPS) but for short clips (up to 10 minutes) it’s fine. For longer clips, I have a machine that I don’t need to work on, let’s call it my “rendering station”.
I’ve tested it with a 320x240 video. First step of the upscale with Starlight Mini and the second step with Proteus to get it to “Full HD”.
Original:
End result:
The color / contrast was changed (in Resolve) before rendering it in Topaz.
Here is my two cents: I just purchased Topaz Video AI after checking out several other options. As soon as I purchased, however, I found out that Starlight will do an even better job of rendering, but that i have to pay even more money - significant money - to render anything. I have a 35 minute VHS recording, and the “demo” 10 second render does look better that any setting on the Topaz app; however, paying another $300 in “credits” after I just paid $300 to buy the app does not seem to make sense.
Hi Erne.
Please remove your contact details as this is an unsecure public forum accessable to anyone
Thank you
You can run Starlight Mini locally. For VHS, it works as well as Starlight online.
For 640x480 VHS, use Starlight Mini 2x, then feed the resulting 1280x960 video to Proteus or Iris LQ to upscale to 1440x1080.
SLM 2x and scale to endresolution with Proteus is good compromise. What I try out now is SLm 2x SD to 960p, then with Virtualdub downscale 960p to 720p and giving TVAI this doing again SLm 2x → result is 1440p. So dual Starlight upscale, this gives me best results as i can say so far. Using a finisher for example Proteus 1x can, pending on the footage, improve the result a bit more.
The Problem is, the second Starlight upscale from 720p to 1440p happens as slow a you do 3x Starlight from 480p to 1440p…so this workflow is a torture and I don’t think I’m doing this with all my vids.
It will not run on MacOs
hoping for more roadmap updates, especially eager for any QOL changes, improved bug testing for less common videos (portrait rotated), and improved starlight mini speed (maybe even a setting for if we want more speed / less quality so we can test to find the right balance before processing a larger video?)
any plan to add starlight to benchmarks?
Are there any differences in Speed between 1x and 2x Starlight Mini upscaling? I started with 1x and then with 2x and i see no difference in fps. Both between 0.3-0-5fps. Starting at 3x, I see a difference in the FPS, that they’re decreasing. So should I always use 2x on compressed SD videos?
I think your “1x” was 960p, min resolution and 2x the same. Speeds pends on from-to resolution
Unfortunately, TVAI still has the bug where the selected upscaling factor could be not activated and the wrong output resolution is displayed. So this are two bugs that may occur. For example, I am currently doing 2x of 480p source, but 480p is displayed in the output current running job instead 960p. this is just a display bug.
But what more worse is, that sometimes if you select for example 4x, but get 2x or 3x. This only happens occasionally. I always monitor the speed to see if it corresponds to my selected factor and check the current output file for large projects. So don’t rely 100% on what you see in the output window. This bug has already been reported. I don’t understand why it hasn’t been corrected.
8 GB ?
Im guessing, educated guessing that a laptop with the suggested minimum of 8 gb nvidia card will not truly suffice as far as ,renders ,time to finish, all subjective as far as video and its length and quality.
Anybody attempting to use an 8gb nvidia gpu. on SLM?
thanks ,Mike
The SLM Sharp model in Beta has a hard 16GB VRAM requirement. From that, I would say that the future models will require even more VRAM.
I dont see why topaz hasnt just focused on cloud based and just make things cheaper for users who stay long term. Even the best hardware is not the best for this software.
Don’t say that, cloud will never be for the mass market, so the focus is right on both.
There are quite a lot of people that will never use the cloud because of privacy concerns with the images/videos.
Besides, seeing how long the waiting in the queue or the total rendering time in the cloud is I have the impression that Topaz servers aren’t really up even to the actual amount of cloud tasks and this likely will get much worse with this new subscription Studio “solution” and it’s free cloud rendering of images…
(let’s see how long that one will stay)
I have the impression that todays Starlight has deteriorated significantly in the cloud compared to May 2025 tests.
I uploaded a test video on 05/05/2025 and repeated this test on 09/17/2025 with the same source. I am shocked at how much the “natural” focus has shifted and how details and sharpness have deteriorated.
The quality of the versions from 09/17 is similar to the local SLM version.
Original source: Super 8 video, digitized in 576p
The new tests were conducted in VAI 7.2.0.2 and VAI 7.2.0.3.b Studio. There is no difference between them.
The May version is a stunning result compared to the original.
Comparisons:
- Starlight (mini) Cloud 05/05/2025 (tele + wide) 1348 x 1080
- Starlight (mini) Cloud 09/17/2025 (tele + wide) 1440 x 1152
Can anyone confirm this problem?
There is no “Starlight Mini Cloud”. The “Quality” and “Speed” online models are completely different than the Starlight Mini model, which runs locally only.
I do know that the online models did get an update to improve their results some time around May, but it appears in your case it actually gave worse results. I guess it may depend on the source?
Also, “Astra Precise” cloud is the exact same model as Starlight cloud (both the “Quality” and “Speed” models).
For my purposes, I have found that Starlight Quality, which is supposed to be the best Starlight that you can run, tends to have many problems, such as random particle hallucinations on dark or black screens, severe banding that seems to be being introduced by the model (as opposed to the video encoder), and occasionally flashes of strange and bizarre patterns in between scenes. It’s bad enough that I can’t really use it for final results. Not sure if this is fixable or not. :\
