Is that a bug? I hope that I can render just part of a clip. That’s how I can easily test models and see which is best. I don’t want to have to cut it up first.
1 Like
I’m currently rendering a segment, so it works on my end.
P.S. shown rendering fps is completly wrong, it shows me 32fps 
3 Likes
Yes, SLP resolves block-like structures less well than SLM; a combination of both would probably be optimal.
By less “smooth” do you mean less blurry? SL mini is incredible, but it’s slightly blurry. I liked starlight sharp in terms of sharpness, but the resonstruction wasn’t as good as mini.
So, I sent the log file.
I’m sure they will fix it.
Are you picking this up? Good. This model is crazy. It works well at 1x (1080p in this case). I already had this 1080p video, but it fixed the details pretty well. All screenshots below are 1080p source with 1x output resolution.
I’m testing the Command & Conquer intro.
Subtle details, but eyes look a lot better.
Look at the womans face in the top left. Insane. This is awesome stuff!
Look at this. JUST LOOK AT IT!
The detail is fantastic. Look at the crystals.
Good improvement even with animation.
Now all I’d like is some speed optimization. It looks like the model isn’t using all of my 5090 VRAM. It’s using about 16gb as in the notes above. Topaz team, if you keep pumping out local models with this type of quality, I will stay subscribed for a LONG time.
4 Likes
Oh, Command & Conquer cutscenes videos! 
1 Like
Wow. Precise locally!! Yay. I see there are teething problems but boy oh boy I love this comapny. Great job Topaz. This looks incredible. THANK YOU!!!
3 Likes
This is the best money I have ever spent on software, cloud or otherwise, and I think this is just the beginning. We may be seeing exponential growth in these models over the next two years.
4 Likes
oh my, my first test on an old blurry 2D 640x480 animation gave me a way better result from Starlight Mini, guess I need to redo all my previous SLM output again with SL Precise, good and pain XD.
Ok, adding my exporting rate below, for this 640x480, 24 FPS clip, using SLP x3 under RTX 4090, the actual exporting rate is around 0.465fps; RTX 5090 is 0.529fps. Hmm
1 Like
logsForSupport.zip (631.9 KB)
This is happening to me too. I have 64gm of memory. All of it gets filled up.
Is it possible that the Neuroserver triggers a Bitdefender cryptomining warning?
None of the Starlight models other than Starlight Mini are working for me:
- Starlight mini: Works.
- Starlight Sharp: Hangs in an endless loop doing nothing.
- Starlight Fast 2: Just says “Error”.
- Starlight HQ: Just says “Error”
- Starlight Precise: Just says “Error”
Logs attached.
logsForSupport.zip (155.5 KB)
SLM: Finely drawn objects—the lines are rendered clearly and thinly, and compression artifacts are well resolved down to the pixel level. You can really see the computational power behind making this possible.
That makes the frame look “soft,” blurry—which in a way it is. But the advantage of SLM is that there’s a lot of information contained in a single frame, which provides a very good starting point. I’d rather have an image that’s a bit too soft than one that’s overly coarse and oversharpened. In Topaz, almost every model adds sharpening anyway, so you can use a finisher like the “Deblur” function (Iris MQ, Proteus Natural). So I never hat problems with SLM is “to soft”.
What I see with SLP is a higher degree of reconstruction compared to SLM so this is crazy how they did that—that’s probably where the model’s strength lies. However, it also appears somewhat Coarser in structure, compression artifacts aren’t resolved as finely as with SLM, and overall it looks a bit rougher. The ideal solution would be to combine the best of both 
1 Like
So the logs are huge—can you delete them, then reproduce the error and copy that log out for posting?
Looks great, but I don’t like the “gray” grain that gets integrated, especially on faces, makes a old “dirty” look. So maybe doing SLM first then SLP helps, need to experiment with that a bit.
