I’ve found the results going from 1080p and higher to be disappointing as well. I don’t think the model is trained for that. There are better and faster models available. Starlight’s fastball is upscaling from a LOW resolution.
5 Likes
So that’s the trend.
The enhancement from SD to FHD was so amazing that I couldn’t help but hope for even more…
But now I’m even more convinced that I should wait to upgrade.
Thank you!
Yup. I’ve mentioned this concern to staff not so long ago, but TL just basically shunned it by stating; “it could take months to train such a model” and my inner-thoughts were like… DUH! you only had a few years now to raise the bar to add this potential hi-rez/HQ model for 1080p+ sources. So close but still so far away it seems, at least from TL.
1 Like
Could be the model is not good trained doing 4k otherwise Topaz Labs offers 4k Cloud Starlight so the model has to be 4k trained.
As higher the resolution is, as more Vram is needed to keep the quality level. I think to bring 4k local to work, the model must be massively reduced…pherhaps when you have a 96GB Vram GPU, then 1080p to 4k looks great, this is just a speculation
Today, this is no longer true. 16K is being easily crushed with today’s hardware with the PROPER memory management & PROPER compression algorithm(s). Racing to the bottom of silly high VRAM claims by the industry is ONLY for the uninformed and… credulous.
2 Likes
Does disabling live export improve export speeds? I mainly use Prores 422 HQ for output.
I can understand the staff’s reaction to some extent. In other words, the release of Starlight Mini is a way for TL to attract more users to Starlight.
I don’t intend to pay a lot of money to use cloud rendering, so I’ll just use a model that can get high performance locally.
I am also having this issue on 7.0.1. With Starlight Mini as the selected model, if I did my last export with output resolution set to 2x, changing it to 1x and triggering a new export through “Export as…” starts exporting as 2x again.
The workaround of switching to Proteus to change the resolution, then back to Starlight Mini does work though.
2 Likes
that’s new for me, is doing 1x Starlight mini possible with this trick then?
1x Starlight Mini was always possible, but there’s a bug with the UI where updating the resolution doesn’t necessarily affect the export. Switching to Proteus to change the resolution there and then switch back seems to resolve it as a workaround.
1 Like
ok thanks I try this out
Quick question: After updating to 7.0.1, I noticed that it downloaded an additional 7MB for StarLight Mini. My question is: What changes does the additional 7MB bring?
changes are in the release notes
Applying rotation no longer highlights the export queue
For Mac, fixed navigation with arrow keys, such as using the keys to select a file
Fixed disconnected drives from being picked up on app relaunch
Improved starlight-mini downloader reliability
Updated Starlight-mini!
Resolves loading model hanging
Improved performance on higher end hardware
Fixed certain videos producing black bars
I was mainly asking about StarLight Mini. For the rest, I had already read the release notes hahaha. I just wanted to know what changes were made to StarLight Mini.
Not better with Iris MQ
The ‘runner.exe’ and its ‘comrades’ got an update, which fixes a bunch of annoyances, the model files are unchanged. 
2 Likes
In my experience no. I tested two things. 1. Lowering the memory usage to 70% and 2. Disabling live preview. Neither gave me any speed increase with starlight so I’m back up to 100%.
I’ve also noticed that teeth look a little better in the new starlight model. I was doing my usual beavis and butthead clip and the teeth are slightly better with this version. Less jagged and weird. Anyone else notice this?
I noticed that in the latest version my rtx 3090 began to use not 16 but 18 GB of video memory. I also noticed that now the video memory settings in the program affect the amount of video memory used by the video card. For example, at 100% there will be 16 GB, and at 70% there will be 11 GB of use, but strangely enough, now the amount of video memory does not affect the quality of the output image. If on 7.0 I had 2 options 100% and <100% which gave output videos of different quality, now the quality is always the same. P/S/ this all happens when using Starlight
Iris LQ works better for me in this situation. It sharpens the softness of Starlight Mini output.
1 Like
I compare 3 options after a starlight export. Proteus, iris lq and iris mq. Then I pick which one looks the best.
