RTX 4080 S 16 GB VRAM set to 88%, 4 min FramePack/Minimax video took 6 h with 0, 3 fps.
For me starlight mini local is by far the best model right now, thx a lot for the update!
1 Like
The ‘darkening’ in your example is really quite different than the problem shown in other examples. But you be you in not believing of course. I have seen both (it darkening and it not darkening with the exact same video) on MY OWN COMPUTER so i guess i know what i’m talking about but… again… you be you in not believing what other people are saying.
It is not a problem the Topaz devs should be concerned about, those paying for the program do not have this problem.
You have a 4090 with this correct?
so it uses around 36GB VRAM at most?
Will Starlight Mini or Topaz Video AI support NVIDIA DGX Spark in the future?
Correct
Why are people saying VRAM percentage affects quality? I thought lowering it would only affect performance/speed, not the actual quality of the model. So does that mean anyone with 16GB cards are getting lesser quality than those with 24GB+ cards? Seems unfair as not everyone can spend 2K on a card just to get the maximum quality from the AI models.
1 Like
hello Tomas, not at most, at least I can’t confirm it, I systematically use the minimum target resolution to gain time. I guess it could take more VRAM if I choose x2 or x3 ?
I agree as well. I think of it as the “Recover” function for low quality images in Gigapixel.
1 Like
Actually, it’s just a theory that i started — no one from Topaz can really confirm it without spilling the tea 
. But here’s the vibe: the model itself always works the same way, like the logic and structure behind it don’t change. What does seem to change is how much it’s able to “understand” or “hold in mind” about the video before it starts upscaling.
Lower VRAM doesn’t make the quality worse by default, but it limits the size of the context window and maybe how many internal layers or temporal references the model can fully use at once. So instead of seeing the big picture, it might have to cut corners, reset more often, or process shorter frame sequences — and that can affect things like consistency, detail retention, and texture stability. Especially in scenes with fast motion or complex patterns.
So yeah, people with more VRAM might get smoother or more coherent results just because the model can stretch out more. It’s not about Topaz “locking” quality behind a paywall — it’s just how heavy and memory-hungry these models are. Still, the Mini version of Starlight actually holds up surprisingly well, even with less VRAM.
6 Likes
You guys with 5090’s.
What resolution videos are you processing with starlight mini? I see you mention variations of 0.2-0.9fps.
Meanwhile, I process at 1.0 fps with my 5080. How is that possible?
1 Like
I think a better test would be to splice a few seconds of the closeup at the beginning of that walking clip and see if letting the model have a look at that makes the rest of the video that follows it look better.
This reminds me of a question someone asked a while back about the potential for uploading images of a person’s face so the enhancement models could use it to improve video clips. The Topaz reply cited the legal and ethical issues that might arise from the potential misuse of the app to create “deep fake” videos rather than technical reasons why this couldn’t be done.
The starlight credits are outrageous, I’ve already payed for topaz video ai 2 times now and now they’re pushing to charge hundreds of dollars for like 30 min of cloud processing… i get stuck waiting for everything even after shelling out hundreds of dollars… i just hope this doesn’t take as long as getting new versons (6.0 and beyond) to even work on that mac mini without crashing after 1 render… took almos 6 months for that… thankfully i was able to find old versions that worked on it in the meantime, but that’s not possible with this upgrade…it’s quite annoying as a paying mac user how late things get fixed/upgraded for mac. Bought a Mac mini back in October 2024 and only as of may was i able to use the new software on my new machine… PLEASE don’t let that happen with starlight mini, ya already hosed us last upgrade, can we jump the line a little bit this time, or are we just always your last priority
1 Like
Even beyond AI, most video frames have key information that can be gained from the frames directly around them. This is how I discovered Topaz in the first place long ago - their focus on improving video through utilizing neighboring frames in an innovative way. I believe they have integrated that same modality deep into their “guessing” generative models and have very likely branched beyond “just” the neighboring frames, which is why other upscaling techniques ain’t cutting it in quite the same way. Out of respect for them, I’m purposely limiting what I’m saying here, and will continue to limit it, but that’s the basics.
We know from previous Topaz postings in these forums that for models other than Starlight, when enhancement is set to Auto the app does an analysis every eight frames and adjusts based on that. They have also cited the use of multiple frame to make “temporal” adjustments as one of the differences between a Video AI enhancement vs running a bunch of still frames through Photo AI. Whether Starlight uses this same scheme and timing is unknown, but the extra VRAM Starlight Mini requires must be doing something differently.
1 Like
Ah that makes sense lol, I appreciate your detailed reply. I’ve learned something new. 
1 Like
Wow, I was sure the ffmpeg inside TVAI is already a customized version of the original, to strictly serve the purpose… And TVAI is forced to use the included ffmpeg. Yet, it seems that it’s not the case.
Thanks for the suggestion!
EDIT: Indeed, with Starlight Mini processing, TVAI is calling the ffmpeg available from the PATH env. variable first, in my case it was the old VEAI installation (1.6.2, I think), that’s why I’ve got the NVENC “p7” error.
Has this model been tested on cartoons?
Any update when Proteus, Iris and Artemis are going to get an update?
I can say, than the destroyed 360p DnD cartoon I’m experimenting the whole time is treated exceptionally well. 
1 Like
Share a sample, please.
