I created a new project, placed a file, select SL sharp, do Ctrl+E and nothing happens.
1 Like
So I think this is probably where I draw the line and look for something open source as has been coming out here recently.
When they start locking everything behind subscriptions or other editions that is where I quit updating the software and I no longer subscribed to updates or anything else.
Sorry guys I donât do greed very good Peace out.
15 Likes
left-click on preview frame first, and then Ctrl+E. Model is limited to 1080p, but you can select 2x, 3x and 4x upscale. What they do is a real upscale with selected factor, but then they do Lanczos downscale to 1080p.
One might wonder why they do downscale. I think itâs because they wanted to limit it due to their cloudâŠbut thatâs just speculation. Anyway my impression is doing 3x or 4x gives better results even the downscale happens, it is structured a bit more finely.
There are bugs. One is knew; when close TVAI Beta the application freeze and I have to kill the task manually.
Shown output resolution can be wrong, yes this bug is still there, I had 2x output when selected 4x and in the output windows was 4x shown, but results was 960p. If this happens close and reimport source chose other model and go back, play around solves it, until next time.
Sure you canât see if itâs 3x or 4x on 1080p limited output. Experience comes with time and you can see in the speed what the upscale resolution is, with my 5090 doing 4x, 480p to 1920p (and Lanczos down to 1080p) I get about 0.7 - 0.9 fps. sometimes fps jumps wild around and you have to wait about half an hour to get the fps but Iâm not sure whats final speed I will see it when itâs done for now I would say itâs about 2-3 faster than SLm. Interesting is VRam Usage does not exceed 16GB, itâs because the model needs not more or have they limited the model qualityâŠ
1 Like
So, SL Sharp shows up on macOS, but clicking on Export (Quick Export or Export asâŠ) is not allowed⊠(resolution <1080p). Also, model not available for download in Model Manager⊠Was this supposed to work? There is nothing in the release notes that says so, so i guess itâs just a placeholder? Thanks.
Iâm sorry, what? I subscribe, paying a very high price per year, to receive model and program updates, but now you tell me Iâm not subscribing hard enough so I donât get the new starlight model? No wonder they arenât optimizing starlight mini for speed. Theyâre cooking up new models to gate behind a higher tiered paywalls.
Maybe you should introduce a new subscription tier called Pro Plus thatâs $24,999 a year. Youâd make more money that way, right? Why stop there? How about $49,999 a year? Disgusting company. I hope there is competition in this space so that they are brought down to reality. This will be my last time subscribing to this product.
If you want to price gouge someone, do it to those who can afford it. Gouge your commercial clients and fortune 500 companies, not the normal people using this to upscale family videos.
26 Likes
I find the situation complicated right now.
On the one hand TL does a good job, on the other hand there are open source models provided by large Chinese companies that also work well.
I am in two minds right now, my main topic is enlarging and I would be interested in what kind of models I could train with my knowledge.
I already have a programmer I could contact who I think could help me.
I will wait and see what the current beta of PhotoAI brings, if it gets an update.
Itâs a pity that TL charges so much money for the Pro version, half the price would be justified, or the option to buy models.
You buy the software with basic models and you buy the models that specifically fit your work separately.
Even cooler would be that an algorithm is provided and that it would then be possible to train commuity models which can then be shared.
3 Likes
Itâs incredible how every time the user interface changes, the new version is always worse than the previous one: the new layout of the models is much less intuitive than the previous one, at least for me!
11 Likes
I actually wish we had the in-app option to downscale to a user selected resolution as I suspect using original->starlight (mini/sharp)->downscale back to original resolution (reimport as source)->proteus 2x, will be good for some content.
4 Likes
yes and and I wish no fix built in sharp it must be customizable
1 Like
No itâs an entirely different model architecture. SLM speed improvements are still very much in progress! 
I assume SLM will be faster than Starlight Sharp when those optimizations are done.
2 Likes
You do get a sneak peak, but this is good feedback - talking with the team here to see if thereâs a way we can grant Pro licenses (for production, not just Beta) to our wonderful Beta users for the next post here, as long as you have an active Video AI license. Weâll see what we can do.
There will be some features behind a Pro tier, but the Pro tiers will look much different than they do now, and will be more accessible. We will also have to boost the system requirements for the Pro version as a result of models like this one.
8 Likes
That would be appreciated. Iâve been a loyal paying customer for many years and I have taken the time to give detailed feedback on various versions and betas to help you improve your product. All I want is to be treated fairly and with respect.
5 Likes
Ok itâs adifferent model but why you call it Starlight? Hopefully we get the faster SLm, thatâs what everyone is waiting for 
After long experimentation on how to combine the models, Iâve come to the conclusion that the SL-Sharp upscale + SLm afterwards upscale doesnât work, but the opposite it does very well. Unfortunately upscale already upscaled SLm footage is no option because SL-Sharp is limited to 1080p. Whatâs produced with SL-Sharp can hardly be further âprocessedâ, it has to be the final encode.
1 Like
Doing this shouldnât be a huge loss for the company. Itâs unlikely that any of us beta testers here will be pushing Proâs biggest business difference (the $1M commercial use limit). Itâll just be models and multi-GPU functionality.
2 Likes
This is long overdue, and not just at TVAI.
Itâs really annoying that 8gb GPUs are still being sold.
4 Likes
Quick question⊠Is the âSharpâ model just for beta testers or can anyone use it??
I am also unable to test the new âStarlight Sharpâ model locally: when exporting, I can only select cloud encoding⊠is this a bug or something intentional since I am not a PRO user? Were you not supposed to unlock local testing for this beta?
4 Likes
Iâve now tried out Starlight Sharp - itâs the SeedVR2 model. I get absolute identical results between the two. So for everyone concerned about having to pay more for the âproâ tier to use it, just install ComfyUI and SeedVR2, or the installer/web interface that SECourses has written on their patreon.
Itâs nice theyâve got the block size at 285, but for me, and the material Iâm using this software on, the quality is much better with Starlight Mini.
Topaz canât just make commercial software that is a copy of open source software, there are people out there willing and able to do âuser friendlyâ versions and not charge (or charge a lot less). Text and small faces - Topaz need to implement additional models into the software to correct this problem. Iâm having to use additional text recognition and face swap software (which are also free) as additional steps and then use Vegas and masks to layer them all together. The only thing USP that Topaz can provide going forward is to have all this in one application.
14 Likes
PRO user only, but there is a backdoor you can use it select Sharp model, then in the middle of the right panel click on the Sharp text description area, then a new windows appears, the Sharp downloader opens, simillar to SLm first use. Download is about 3GB. When download is finished, close and reopen TVAI import source choose the model, do your settings. Left click into big video View window and then press CTRL+E (thatâs the the trick) to start quick export.
4 Likes
