Yes - we don’t do a good job of displaying model unloading/unpacking to users in our apps. We’re working on this for Gigapixel - I’ll take a look at what we can do for Video AI. It is indeed slower the first time but it would be good to visualize this. 
Is this what I get when I try to download the models?
I redid a clean install because I couldn’t download the new Starlight model. This is still the case after the clean install. And the template manager is broken. The only way to download is to load a video and choose a template from the list to download it. But won’t work for Starlight
Making note of this - seeing it requested by a few folks
We aren’t able to share too much information about the model architecture here - Starlight was the first model in the world of its kind, and Mini is the second model of its kind and the first to run locally. But we are definitely working on other versions of Starlight with improved results in some of these scenarios.
4 Likes
Oh okay, now i get it. I was like “why they dont “talk about more of how it works””. But it manage to be of the patent. So if you talk about, lost the “secret”. Its fair in a certain point, but when we ask of how it works, is like, in a comercial way to tell, like samsung does with his ai helper, or google when they share some new information. I know that is from the marketing team, but it can help a lot to be more clear that topaz is the only company to manage a model like this. Its a huge impact in ai bubble and even in video collectors too, but they dosent know any of this huge step
3 Likes
Hiya, Dakota!
It seems like your recent response here did a great job covering many of the concerns from .01b-04b.
You’re absolutely right, beta issues status in a public release are not shared… so, let my clarify. Those concerns were never fully voiced by staff, as for as I know in the last beta .04b, of those issues status before the public release of v7.
It would require me to drink a tall glass of herbal tea… ha-ha… to cover all my major concerns with TVAi; SL Mini destruction to vegetation (I recall staff stating something of a wait-&-see future roadmap to this, though), SL Mini destruction to DOF, all recent models it seems are only heavy focused on low-rez samples/sources but no recent Ai model that is dedicated to HiRez 1080p+, especially 4K & 8K sources, ect.
When will AMD GPUs be supported for Starlight Mini?
2 Likes
So after some tests, my rough and ready reckoning is that even with a kickass system (say RTX 4090, and 14th Gen i9 chip, and 128GB of RAM) you can still count on an hour of processing time for every one minute of footage in Starlight-Mini. After looking at the results, yeah, they’re impressive, but it’s just not worth the time investment for me personally. Starlight has a ways to go development-wise to make it feasible imo.
6 Likes
COMMAND ERROR Setting Option presst to value p6 with starlight mini. Any Idea?
Makes sense - those model changes are not something we can do quickly, they take often months of research. Diffusion models struggle with vegetation - DOF is an interesting one. Our team will make note of this for sure - but any changes addressing things like this may actually appear in a future new model if the model architecture prevents us from fixing some of those deep quality problems.
Very important feedback though to know where Starlight Mini does well and where it doesn’t, this is great for us when considering how to build and update our future models so thank you for sharing.
1 Like
In the future! 
I can’t give you a timeline, I apologize - but we are working on both AMD and Mac support. When we are closer to implementation the Beta users will be the first to know.
3 Likes
Fair enough! It’s not for everyone. It is one of our most powerful and demanding models ever - but we won’t stop here. We’ll continue making Starlight (and our other AI models) better. Stay tuned!
1 Like
I’ve seen lots of p7 issues but first with p6. Team is checking on these, thanks for sharing. If you can send logs that may help us debug.
1 Like
We were sleeping we released at 6pm Texas time
won’t respond to every comment but I read them all! between meetings at least. Many of our team members also read the forums.
3 Likes
Are you working on implementing pause and resume for the Starlight Mini model?
I feel this is an essential feature due to the slow processing time of this model.
4 Likes
Does starlight mini cost credits or is it free like Artemis and all the other models if you buy the lifetime. I don’t want to pay extra credits for a model and the verbiage from topaz suggesting three times cheaper than the regular starlight makes people think that this is a model that cost credits
What amount of Vram is needed to house the entire Starlight Mini model on a local machine?
Starlight mini will not allow me to export, getting the following error:
Unrecognized FFmpeg option: ‘aq-strength’. Check the FFmpeg command.
I shut TVAI down and restarted and also lost all my other previous sources I was working on. 
Nice update and fun to try.
However on my 5090 Astral, i9 285k cpu, 64Gb ddr5, Im achieving 0.1- 0.3fps rendering time on 480x360 —> 1280x960 res (29.97fps).
The result is honestly quite stunning, very impressive, however looking in this forum post 3090/4090’s are exceeding that processing time by a fair bit. Up to 2fps?
Early days, will see if there is anything that can be done to tweak a bit on my side. Will tweak in preferences to not use 100%gpu and lower to 90% for shiz and giggles just to see.
Not sure at what level Topaz VEAI is at yet with full 5090 support? My monitoring software says gpu is running at between 98-100% and memory is maxed on usage.
edit: if 5090’s are not optimised yet and expected to perform worse than the previous gen cards thats fine, let me know, so I can sit back, be patient, and wait for that update in the future. It all looks very exciting though thats for sure.
3 Likes
