Oh I see. I’m doing the previews on the web app instead. Have you tried there?
We need Blackwell support in the app! It’s still much slower to process on RTX 5090 than it was on RTX 4090, weeks after release.
2 Likes
Thank you for the detailed information! I’ll try that for 3090. 
I’m glad that the bug with the real-time preview approaching the end of the video (especially with different fps set with interpolation) is fixed.
Edit: The power diagram still looks like a saw. 121% - yikes!
1 Like
Tony, your marketing team is doing a horrible job positioning this update.
Your company is underscoring things that don’t matter to your customers. They could give two figs about Topaz being a ‘research-focused company’. Shouldn’t that focus be us, the customers? Because, clearly without customers, you have no company
What is coming through is “cloud rendering”. Not - ‘we’ll eventually have all these new models available for high end GPUs’.
I would ask whether anyone there is truly looking at the metrics of how much cloud rendering is being utilized.
And, considering that MANY users are sticking with 5.3.6 - that right there speaks volumes about your customers’ perception of where you are as a company.
Not listening. Not seeming to care. Oh well.
A company that clings to rigid principles without listening to its customers risks irrelevance, lost loyalty, and stagnation.
Before you try and counter my comment … I get it. I’ve read your and your company’s positioning on this. No need to elaborate.
17 Likes
Thanks for the tip - just submitted. we’ll see
1 Like
Don’t click on Cloud Export. Instead, you need to click on Try Starlight on the top right hand corner of Video AI.
1 Like
Better and it better be near future. I am not a fan of cloud rendering at all and will never use it. I’ve paid good money for TVAI for a while now. Now I have a version that is missing a feature simply because it’s a desktop version. If this is the path Topaz is going down then my decision to no longer renew is being cemented by the second. I can only appreciate one thing with this release and that’s a new model. I don’t appreciate it not being included in the desktop version. It can run at 1 frame a second for all I care, but being forced to use cloud rendering and not having it included in the desktop version feels like theft. Not a good way to make your customers feel.
9 Likes
If we could offer 1 frame per second on desktop cards today, we definitely would!
Project Starlight is running on H100 systems right now, and does not run on consumer cards, even at 0.00x FPS speeds.
So I paid for 12 months of updates and now you want me to pay for cloud? Is it possible to get refund if that’s your direction now?
3 Likes
This is a free preview of a model that cannot currently run on desktop GPUs. We have plans for further updates to models that run on your machine.
4 Likes
Thank you for being transparent and I look forward to it being refined to run in the desktop version then. Appreciate your quick response Tony!
4 Likes
Just a quick comment on the preview of Starlight. Biggest issue I can see with generative, which I’m sure is well known to your team, is that it will completely obscure and remake details that even the old (current) models see and deal with. For example a reflection in a pair of glasses. Generative just remakes the glass and maybe overall distortion and misses a reflection, whereas on a current model, even a poor quality reflection is dealt with and represented. The output was impressive but at the moment I don’t think game-changing.
It’s easy to be taken in or starry-eyed excuse the pun by the compute cost and presentation, etc, but if this was a new local model, I think you’d be getting a lot of the same old complaints such as monster faces on very low detail inputs.
That said, I hope you are planning some noticeable evolutions in local models in the next five or so months, perhaps in terms of dealing with old animation, or the elusive lower quality inputs. I have done some absolutely astounding stuff with TVAI, and for context I go back to the old circa 2009 image enhance days. I am loving the faux 4k I can achieve with higher quality inputs and look forward to a time when you guys hit the mark a little more with the older low quality stuff such as TV shows forever locked in video resolution limbo.
Like others I have been frustrated by the bugs, the new less accessible interface, and lack of model evolutions recently, but I do plan to stick with Topaz while it is still doing good things for me. Wish your team best of luck in getting or keeping this barge moving in the right direction.
8 Likes
Yeah I agree 100% that Topaz has definitely improved my medium and higher quality Sony A7S III 1080P slomo footage into 4K with higher fidelity. It still has a way to go in terms of lower quality inputs. Really hope to get some more models to play with. I know everybody on here has a different use case so what might work well for one might not for another. I am actually very happy with the Proteus model. It just lacks the anti-aliasing/moire reduction that Startlight does well on lower resolution lineskipped footage. However, Starlight still has problems adding contrast, chaging color hue and oversharpening/overestoring details a bit.
If the AI team is reading this, it be nice if you can explore into looking at a Starlight evolution of the Artemis Aliasing/Moire preset where it just focuses soley on getting rid of those artifacts and not necessarily restoring detail or sharpening. I don’t know if doing it that way would speed things up a bit and enable a local model. That way I can just use a second pass of Proteus to restore the detail after the aliasing/moire is reduced or removed.
2 Likes
We were unable to test the operation of the DaVinci Resolve Studio OFX plug-in in the beta version.
Will the DaVinci Resolve Studio OFX plug-in work properly with the official version?
The main purpose of buying this software over other solutions, in my view, is privacy. Many of us do not want your cloud logging our videos and as such will never ever use a managed cloud service to render the video. It doesn’t matter how much better it is than the alternatives. I will not use it.
However, it is easy and affordable to rent private H100s nowadays. Why not include the code for the worker and the model weights to run Starlight as it is with the license to the desktop app? That will satisfy everyone seeking privacy and you wouldn’t lose any of the customers who are fine paying for the convenience of using your cloud service.
7 Likes
I remember when I upgraded from a 3080 to the 4080 and yet the performance were almost identical, it took almost a year for the card to be optimised, I think you will be looking anything after August 2025 for the software to use the 50 series properly, that’s why I held off trying to buy one.
1 Like
That’s ridiculous, they just need to compile a TensorRT model with the new version. It’s already supported by Nvidia.
Right now the performance is not identical, the 5090 is far slower than the 4090 was as the TensorRT model is not compatible.
2 Likes
I agree, it’s pretty simple for them to do but somehow they don’t seem to prioritise optimising their software and models well for newer hardware, everyone had the same issue with the newer M4 series came out, I don’t know if that’s been sorted as I didn’t monitor it but judging by how long it took for the 40 series, I wouldn’t hold your breath lol. It may be a simple thing to do to us, just like how simple it would be to make the software bugless and more efficient with hardware, but here we are years later asking the same thing and the issues are still here.
1 Like
You need to add “creativity” option like in Gigapixel.
Some faces are excessively altered.
4 Likes
I don’t think it will be possible soon upscaling a whole movie with different scenes in breathtaking quality without artefacts, but there is still a lot of room for improvement existing models and I expect these deficits like monster faces, artifacts especially RheaXL produces tons of it, to be addressed…but nothing happened here and I suspect it’s because of Starlight which has taken away capacity.
It seems for me that only the B-team was working on the base product in case of Starlight, this annoys more and more customers!
We don’t need hunderts models, we need less but better ones, a model that is more usable run it over a complete video. You can also merge existing models together when the result is a better one.
I don’t hate Cloud rendering, maybe I will use and pay for a video that is very important to me, but due to the high costs, which will never drop to the level of local rendering, the cloud can never be a replacement.
4 Likes
