Starlight cloud precise two, limited to 18 seconds? On ASTRA

tryed the new starlight precise 2 but it seems limited to 18 secomds and the creative model is aswell? I like precise 1 just wondering why the limit per render? Also im tslking on the astra web interface

Also i noticed on precise mode, it seems to dumb the colours down?

The new Starlight precise 2 in astra.app is great for old cine films (better than Starlight mini ) and even deals with most tram lines on the film. BUT it is ludicrous to limit the time to 9 seconds (i use 50P files). Sort your servers out Topaz, an absolute joke.

Starlight Precise 2 has a 450 frame limit.

Regarding the color issue, can you share a screenshot of the input vs output?

this is PRECISE 1 but precise 1 and 2 both dumb down the colours, it is quite a difference in colour, if you look at how bright the fireworks were in the original VS the 4K output on the right side, this was upscale from 1080p left side to 4K output RIGHT SIDE, I wish there was an option to enable “vivid colour enhancement” to keep the original colours or make them better

Can you email in and send a request ID number for a file ran within the last 7 days so we can take a look?

If you don’t have one available within the last 7 days, please share the input and output files with us.

I wrote a python script that can take a bunch of 18 second clips from Astra and stitch them together into a single video file. Would anyone be interested in me posting that to github so you can use it?

Yes please, if you can post that GitHub link that would be very helpful for stitching them back together. How have you been splitting the short video clips (from same source) before uploading to Astra? Do you use LosslessCut? Also, curious if your Starlight Precision 2 rendered clips from the same video source are coming out visually looking the same once you join them back together? Since the model only works with the 450 frames for each clip independently, the same subject/props end up looking different/colored different when combing the clips together (That’s my experience).

@margaux.topazlabs Would you happen to know an estimate of time for when the Starlight Precision 2 model will eventually accept longer clips, past the 450 frame limit? It’s an amazing model and I hope we can start uploading longer clips soon like the older models.

I haven’t uploaded it to GitHub yet, but I’ll work on it. I don’t split the clips before uploading to Astra, I just upload it as one clip and then render 19 agonizing seconds at a time. Yes, there is always a little jump in the image from one clip to the next because the AI creates them slightly differently. For my script, it’s best if the clips overlap by at least a few frames, up to a second or so, so it can figure out the sequence. also I earnestly recommend that you put it in a directory, and hand it to Claude Code to be the pilot; it helped me write it and it will have a good grasp of the parameters, it can run it automatically in a powershell, even alter the code for you according to the results you’re getting. I’m happy to help you if you need it.

At the moment, there isn’t an estimated timeline for extending Starlight Precision 2 beyond the 450-frame limit but it’s something the team is actively looking into.

And for Starlight mini?

I don’t really understand this limitation to such extremely short clips anyway. What is the difference if someone uploads a 2 hour movie compared to 24x 5 min clips that then are processed one after the other?

But this max 5 min. issue renders the whole cloud encoding quite useless.

If someone wanted to take that hassle (plus the obvious temporal inconsistencies at the split/merge points) he could as well go for SeedVR2…

For Starlight Mini in Cloud, there is currently a 9000 frame limit per job.

As of today, Scene Detection in Astra is now available for clips up to 10 minutes long (18000 frames). So if your video contains multiple scenes, you can upload up to 10 minutes and the system will automatically split and process them per scene.

Single scene limits still apply.

2 Likes