Today’s beta release is something many of you have been asking for since the original launch of Starlight: local processing with Starlight mini.
Starlight mini is the second diffusion-based upscaling model Topaz Labs has released, but is the first to be offered as a standard in-app update for all existing and new users.
System requirements for this model are more strict than the overall application. Users will require an NVIDIA GPU with at least 8GB of video memory, but we strongly recommend a GPU with 16GB or more memory for optimal performance and quality. This beta will run on AMD GPU/Intel GPU/macOS systems but will not support Starlight mini.
In addition to these GPU requirements, Starlight mini also requires a minimum of 12GB of system RAM.
More updates:
Live previews of exports are now visible in the Export queue.
Added account-level Generative AI and Cloud Render controls to Pro account panel.
Hurrah! I’ve been checking every few hours waiting for this release haha. Trying an HD clip now, same one that I also tried on the Starlight-Maxi cloud so I can compare results.
RTX4090 TI and Threadripper 3970X - and it’s 0.3fps. Well, fair enough I guess. Is there a another graphics card I can buy that will improve that, for less than £5,000?
Just went on a trip, and finally, there’s a model update to look forward to!
Will there be updates for Proteus, Iris, or some slow-motion models in the future?
Been waiting too long, can’t wait to come back and try them out!
RTX 5090 would be faster, but in all honesty I’d wait for more results from testers before we really understand the speed differences between different CPU/GPU combinations
Using a PAL DVD source, getting 0.6fps. I have to say - the results are absolutely incredible. The additional processing (sharpening, etc.) that the cloud version always gave required extra processing from me to suit my needs. I guess it was geared for a more “bro” crowd wanting unnaturally sharp video? This removes all the noise and artefacts of the source, but still looks completely natural.
The DVD source is set to export as square pixels, 768x576 but produces a 1280x960 file - first bug report?
I’m experiencing a similar issue. A 320x240 input is exported to 1280x960, while I left output resolution set to Original 320x240. Results are interesting, in a good way. Very stable, little artifacting, maybe a little overly smooth, and very very slow. This is a good start. I like it more than Rhea already, except for the speed.
If the file name contains multi-byte characters, starlight_mini will generate an error and will not be processed.
By changing the file name to single-byte characters, starlight_mini will be processed successfully.
Starlight Mini does not download (the package atfer the hw test) behind a proxy server. Is there a way to download it manually? My Web access is ok, in the settings section autodetect proxy is enabled, I get at program start positive license verification, model downloads, except Starlight, works.
Consistent failures on attempts at utilizing starlight mini, with various sources, on a TRX50 / RTX5090 rig. Where should logs for this beta be dropped?
Output aspect ratio is correct. 1280x960 is 4:3, seems to be doing no upscale is not working. It could be Starlight must do an upscale and this is lowest upscale resolution for Starlight
RTX 5070, 64gb ram. All attempts to use starlight mini fail, whether resizing or not, large or small source, 5 seconds or 20 seconds. Of particular interest is this log that shows up before the directory cleanup messages after an export error:
2025-04-29 22-47-32.863 Thread: 16196 Warning qrc:/workflows/TProcessor.qml:442: Error: Cannot assign std::nullptr_t to int
2025-04-29 22-47-33.569 Thread: 16196 Info process exited: 9 1 0
There are quite a few QT error messages. Not sure if that’s normal.
The export starts to ramp up, taking a bunch of RAM, then it consumes about 50% of video ram for a couple seconds, and finally it crashes before processing any frames and without any preview window content.
