Question about Project Starlight Results: Mixed Quality and Lack of Adjustability

Topaz Video AI General
1

Hello, I hope that this is an appropriate place to ask the following question.

I’m a fairly long-time owner of several Topaz products, but honestly haven’t put them to much use due to the age and capabilities of the computers on which they been installed. (Previously limited to an aging AMD system with a 2018 AMD Radeon RX570 as my ‘most capable’ device.) Although most of the older versions of Gigapixel, Sharpen, DeNoise, Video Enhance AI and Video AI would run, results were obviously mixed or unsatisfactory for a number of understandable reasons.

Last year, I built a new system with far better, near-high-end, consumer-grade components and specs that far exceed the minimum requirements, albeit not equal to their latest 2025 counterparts. So it was time to dig into and try to learn to use my Topaz tools effectively. (Sadly, I’m still not up-to-speed, but that’s on me due to my being an ‘old guy.’)

With that as prelude, here’s my question: Last night, I submitted my three 10 second samples to the Project Starlight Preview.

The results that were returned were mixed. In some - maybe even most! - areas, the clips looked absolutely amazing, with better-looking video frames than existed in copies of the original 1980s television material sourced from backups of the deteriorating master tapes. (BetaCam of some sort, I think…) I was thrilled to see how good some/most of my 30 seconds of samples looked!

But… There were also some aspects of the video that were less-than-acceptable, ie., unusable. Project Starlight seems to be non-adjustable, in that you load a video and mash the Preview button. This is quite unlike selecting one of the convention models, adjusting parameters, resolutions, grain, and frame rates, etc. In other words, there doesn’t seem to be anything that can be done to achieve better results when you’re not fully satisfied with the video returned. Is this a correct take on its functionality? There’s nothing that can be done to increase the chances that you’ll get a usable result?

I should add that the material I submitted had been through a restoration process using AVI synth scripts and other tools to make it suitable for release on DVD. So it wasn’t exceptionally low-grade material that was being submitted.

For a free preview, this isn’t a problem, but if you’re paying a princely-sum for the service, I’m thinking that there will be at least a small number of unhappy customers.

2

As of right now Project Starlight is a one click process without user input to adjust the export. This is by design and allows the diffusion based AI model to handle the processing as it sees fit.

In the future though the team is looking at adjusting how the new diffusion models are controlled and allow for more user input to their settings.

Project Starlight right now is still in the early research release phase and our team is collecting information and feedback from users on the results for future improvements.

3

Thank you very much for your response. It looked that way, but it’s nice to have confirmation. It’s also nice to know that some tweaking may eventually be facilitated.

My first 3 weekly 10-second clips have yielded mixed results, but the goods have been astonishingly good. The negatives aspects have been few, but strong enough that they could be ‘jobstoppers’.

I’m wondering how this could impact users who paid a significant sum and received something that’s 99% stunning, but leaves Jennifer Aniston looking like the Hunchback of Notre Dame in that remaining 1%, leaving them unable to use the result. I can see some serious complaints flying.

Overall, it’s an amazing technology that you’re birthing here…