we need a “dislike” button on the forum
5 Likes
Local rendering is just a better experience - no argument there! I wish we could have released this locally, and hope we can release a version of it soon.
For some context, our top priority is to release the best quality models as quickly as possible. Topaz is meaningless if we can’t do that.
With that in mind, for us the question was “should we release Starlight cloud only first, or should we wait months until we can optimize it?” We chose to release cloud-only now. (Unfortunately we can’t commercially release a model that takes 20 hours to process 1 minute of footage on an H100 , even if a few people would use it 
)
While we can certainly be faulted for not wanting to wait, it’s because of a desire to move quickly. We’re not trying to hold out on you. We’re also losing money on Starlight until we can optimize it, so you can bet we’re pretty incentivized to do so.
7 Likes
Appreciate it. Diffusion tech is pretty obviously the future for image/video enhancement, but there’s a lot of progress yet to be made
3 Likes
It’s not, really. Our purpose is always to get the best quality models out as quickly as possible.
In that mission, we’ve been pre-empted before by others who were willing to use cloud to get larger models to market, while we were using local as a constraint and falling behind. We’re trying to not repeat that mistake.
2 Likes
I hear you on this. We’re working on it
provide quality grades between existing models and Starlight
You’ve hit the nail on the head and are exactly accurate as to what we want to do. It’s just that we need to build Starlight before the existence of those additional quality grades is possible.
Pushing quality forward is a research problem with indefinite outcomes, rather than an engineering problem with predictable progress. We needed to see how good something could be without constraints before adding them in.
the moment the DeepSeek equivalent of video upscaling hits
We’re looking forward to delivering this to you
4 Likes
Sure. But as with the diffusion based image enhancements we definitely need text recovery mechanisms also in video upscaling as otherwise the videos will be vastly ugly with about every sign, clothing with text or logos on it (even on cars - company logos, number plates, stickers,…) being that terribly distorted.
Not to speak about burned in subtitles.
So I think this should be priority number 1 - followed by number 2: avoiding of monster faces and face alterations.
1 Like
Really appreciate the constructive feedback. We wanted to release first and then optimize, but it’s a reasonable opinion that we should have waited to optimize first. We currently can’t share metrics about the core Starlight model, but might be able to share some details about the derivative models that come from it. Hope you understand.
3 Likes
I can’t find a specific example right now that’s publicly available, but Starlight is actually fantastic at preserving text - way better than the older models. It’s one of its main strengths.
Monster faces will still happen to really small faces, but way less frequently 
I my short tests it was quite terrible - similar to redefine in Gigapixel without preserve text option.
And that also in about any other uploaded sample I looked at.
See here:
look at the "5"s on the vest
or also e.g. at that example done by another user where I’ just took a screenshot to show some flaws:
The occurrence of monster faces seems to have been drastically reduced, indeed. And Starlight seems to finally have at least some “knowledge” of foreground/background, not trying to make every background face or other out of focus objects razor-sharp 
Great my slowmo video montages tend to only be 30 seconds - one minute long so I don’t mind waiting 10-20 hours for it to finish. lol jk
1 Like
Ah I see, you’re right. I think the model didn’t recognize the emblem as text.
Here’s one of the examples I was thinking about - it’s a lot better than what our previous models can achieve:
Of course, plenty of room for improvement still.
3 Likes
Eric that example looks great! Thanks for sharing!
It still alters the font so it doesn’t really “fit” anymore - and it looks kinda like “stamped in”. The original is much more natural and in those cases I firmly do believe it would be better to leave those items with text and logo alone / not trying to enhance them (as you do in the PhotoAI and GP apps with preserve text option).
Another recent example from this thread:
With that distortion everyone that only has had loose contact to AI upscales will realize in less than a second that this has been done with AI.
Which I try to avoid in my upscalings - more isn’t always better.
Ohh - so this new model is only for cloud processing? lol
Many of us have high end computers. You can split up a clip and run it on multiple seats.
COUNT ME OUT – I am not paying for cloud processing. What a shame.
I knew this was too good to be true.
3 Likes
How much is in-house development, and how much is using GitHub - NJU-PCALab/STAR: STAR: Spatial-Temporal Augmentation with Text-to-Video Models for Real-World Video Super-Resolution? Because it looks VERY much the same. I’ll probably be banned for mentioning this, but the similarities are questionable.
1 Like
Thanks for bringing it to our attention. I just looked into it. It seems they can only support 420p. Not sure whether it’s useful for you. We are looking at billions of pixels when restoring one pixel, it’s a totally different approach.
What I’m hearing feels more like a justification for internal business decisions rather than an acknowledgment of customer concerns.
A few key points stand out:
- No one is questioning that local rendering provides a better experience. The frustration isn’t about whether it’s possible now—it’s about the lack of focus on what customers actually want.
- The fact that many users are sticking with 5.3.6 contradicts the narrative of progress. That alone should be a significant signal about customer sentiment.
- While there is mention of optimizing for local use, there’s no clear commitment on when that will actually happen.
- The overall message still comes across as “this is the direction we’re taking, whether users prefer it or not.”
Cloud-based rendering isn’t just a temporary solution while optimizations are in progress—it’s also a revenue stream.
Framing it as “we’re losing money on Starlight until we can optimize it” gives the impression of shared struggle, but in reality, your cutting edge model pushes customers toward paid cloud processing, aside from a few free token renders.
Your own Video AI landing page says: “Local. Secure. Processing built for pros. Directly process video using your existing hardware. Work on confidential and protected images and video without ever uploading to a server. Local processing in all Topaz products guarantees your work remains secure, while making for lightning-fast processing, too.”
Meanwhile, I continue to receive marketing emails promoting cloud rendering —something I neither need nor want. I quote from one email your marketing team sent within the last 10 days: " *Take advantage of faster processing. Use Cloud Credits to speed up your renders on Photo AI 3, Gigapixel 8, Video AI 6, and Gigapixel iOS. They also allow you to batch process images and videos to make your workflow more efficient even on low-power machines."
Do you honestly expect power users to spend $2,400 a year for the convenience and lightning speed of cloud rendering video, regardless of whether they use Starlight? Really???
And before you mention other less expensive tiers of credit purchases - let’s be clear, that’s just splitting hairs.
If this shift wasn’t about encouraging users toward a more profitable model, then why does every update seem to make cloud services more integral, more important, rather than less?
And if this really is about channeling users toward a cloud-based solution, BE HONEST and UPFRONT about your ultimate goal. That, I can respect.
But everything I’ve seen from you and your team fails to address the true dissatisfaction that long-standing customers have clearly stated in these forums.
I understand that optimization takes time, but let’s not ignore the financial aspect. If cloud rendering weren’t profitable, would it really be the first and only option available?
Your response doesn’t alleviate my concerns—it reinforces them.
17 Likes
Our internal test shows it’s quite critical to maintain a large model capacity for better quality. We invented many different models last year. The Starlight model was the only model that can handle aliasing/moire, and has a broad capability of improving video quality. It took us three month to optimize it and make it available on cloud. Otherwise it’s prohibitive even running in the cloud. We are working on smaller models to make it more accessible, leveraging the knowledge we learned from developing Starlight. But unlikely it will have the quality close to Starlight according to our internal benchmark.
4 Likes
Okay, thank you for the detailed explanation. 
