so what you think, do we get 4k starlight? Computing power increases fourfold and the costs… 
This consideration alone makes Starlight a substitute for very few people …until they break down it runs lokaly.
You need to add the user review 
to your page app.topazlabs.com/starlight
I’m impressed with how “Project Starlight” has managed to enhance old digitized VHS video!
1 Like
Thank you for this recipe. I’ll definitely give it a try.
FilmConvert Nitrate is amazing! Thanks for the tips. I actually increase the softness a bit for very sharp footage so it doesn’t look too clinical. For $119 it’s a good buy.
1 Like
Softness is your friend when we’re talking really bad quality upscales. It’s the same with the softness as it is with the grain. No, ideally maybe you wouldn’t want it, and also YouTube butchers grain footage if you pixel peep it (seriously, it’s a mess,) but bottom line is you often absolutely need it if the end result isn’t going to look like some plastic garbage.
By the way, FCN can actually still be useful with the strength set at 0.
Horrible Chinese 720p webstream to 4K.
1 Like
I couldn’t agree with you more Thomas! Softness and film grain is def. your friend especailly when upscaling these low resolution shots. LIke you said, it’s often absolutely necessary so that your footage doesn’t look plasticy or too smooth after upscaling and noise reduction. I know most users on here want super clean, but that only really works if your footage is medium/high quality or high res enough from the start. If it’s super low resolution with a lot of banding, film grain is necessary from it looking like smeared vasaline.
4 Likes
But Tony, that’s just literally not the case even in the home AI market. Basically everyone who is serious about self-hosting AI models, programs, etc. is running dual GPUs (or more) if they can or even dedicated hardware in a rack. You’re clearly still developing the feature, but you’ve chosen to lock it behind an absurd paywall for $1,100 per year.
If the real reason was the lack of SLI/Crossfire in machines, you’d simply stop developing the feature. Why not just keep the model of Topaz Video AI pricing being based on revenue, but include all features in all tiers? Game engines have done this for well over a decade without issue. It works well for all users.
You absolutely should be adding multi-GPU support into the standard version of Topaz Video AI if you’re going to continue to support it as a feature at all, or you’re simply going to lose customers to open source solutions as they become better over time.
I don’t normally edit video but I thought I’d give Starlight a try.
I chose a random video I had saved from the web (Got My Mojo Workin’.m4v, 39MB). First I trimmed it in the Mac Finder QuickLook view because the Starlight site says, “Limited to 3 clips with fewer than 300 frames”. Does that mean the user has to prep the samples first? Hard to tell.*
After some chewing, the process failed for some unknown reason (with no end of the failure in sight):
I thought maybe I borked the clip while trimming, so I uploaded the unedited version. Same error.
Next I tried another small mp4 without editing. This one is progressing:
Finished after approx. 40 mins.:
*It appears the render begins from the start of the video and stops at the demo limit.
Paying for cloud rendering is a non-starter. How slowly a model may run on my 4090 is for me to worry about, not having the option to run locally is BS. If this is going to be the new money grab I’ll find something else.
7 Likes
The future Project Starlight model definitely deserves the name, if not Zeus, then Poseidon.
Starlight cloud times
300 frames to 1080p → 40min
300 frames to 4k would be factor 4x → 160min
150.000 frames 1h30m movie to 4k → 20.000min = 333h = 14 days
upscale cost… (do you really want to know that? )
At this point the only thing we can do is stop suggesting AI model tweaks, how different models combined can restore certain footage and anything else Topaz can, will and are using to improve their models. Simply stop giving feedback for free, when we get nothing in return.
The new models are obviously not for your average consumer (us).
I hear promises, but so far, it’s all empty words.
The promised support for Blackwell, which was “near release” 2 weeks ago according to TVAi themselves. Now? Not a single word from the developers. Last i heard, shortly after the “it’s near release” statement is that it would be released at the latest by end of February.
Hopefully they’ll release the value we provide and will start to give some back.
The v6 GUI that i haven’t heard a single good word from any user anywhere is still the focus. Why? I simply don’t get it.
Edit* The above is especially true since they are only some ~50 total employees(?). I guess they’ve done the math and concluded that cloud-based rendering for enterprise and large corporates will pay for employees that can do what all of us users are doing for free.
3 Likes
Trust me! The best GUI version is 3.0. The following version seems to waste your time!!!
2 Likes
Такое же качества ты получишь при использовании Rhea и theia . Эта модель фикция.
If this turns out to be a subscription based pay-as-you-go for software, then of course they’ll loose thier home users and hobbyists, just like Adobe did.
with the difference that you can use adobe as often as you want with a subscription, whereas with Topaz you pay for every upload. That will be a very small customer base… Provided the base product dies, which they say it won’t, but I’m seeing the opposite at the moment.
4 posts were split to a new topic: Licensing discussion
If video and image upsampling follow the same path as graphics and video editors, then five or ten years down the line I foresee paid versions of TVAI at the “professional” or “bleeding edge” level that perform the most advanced functions and limited-use “hobbyist” versions that are downloadable and perform basic versions at no charge, with an “upgrade to unlock these advanced features” link or popup. Because by then there’s going to be someone posting free open-source alternatives to sites like Github and Filepuma, and if Topaz users migrate to those, the potential base of paid customers starts to dwindle.
So something with the capabilities of Video AI 2.x or 3.x as the freeware low-end version and the latest Pro as the subscription/pay-as-you-go version with stable diffusion and text-to-video generative AI models as the high end, with additional levels in between at varying prices. Hopefully, with both subscription and perpetual license pricing tiers.
1 Like
