Same problem. Amazing when an upgrade is a downgrade…
1 Like
I’m nearing the end of a 3x upscale of a 53-minute 1270 x 720p video that I broke into 5-minute segments to guard against losing something in case of a crash. They were all queued up to run consecutively, and have averaged 42hrs+ per segment. So that’s essentially one long render, at 0.1fps. (break-in project on a PNY RTX 5070ti, which has performed flawlessly)
It has about 12 hrs to go on segment 9 of 11.
(Something did go wrong in segment 8 - when I was out of town for 4 days 
)
1 Like
Agree…. when and if it comes to Mac, it is going to be even slower.. I run an M4 pro…..
The closest thing I have come to starlight mini quality is to run the face model in a weirdly named software program called vikpea AI…. Used to be called HitPaw…go figure.
Then run it in topaz AI using Tommy‘s method of Proteus, focus fix normal and sharpness at 30 to 50.
It’s close, but no cigar….. sometimes a face turns out a little weird or you have some black added to the eye socket, but it’s random…..It is a hell of a lot faster.
1 Like
Before I came across Topaz, I tried Hitpaw and wasn’t all that impressed. 
1 Like
I found only their face model to be worth a darn.
But it does give some improvement….albeit at times the results are a little bit disconcerting. Black eyes mainly is what I see.
And they have steadily improved that model.
If you’re on a Mac, it’s kind of the only game in town I have found….that comes anywhere close to starlight mini.
Even then, it’s clearly not as good.
1 Like
Still waiting for Apple Silicon and Windows 11 ARM64 starlight mini support, no matter how slow this might be, the efficiency is unmatched with these ARM chips.
1 Like
The CPU is not crucial for SLM, as the model primarily runs on the GPU (and currently only on NVidia, Tensor cores). The CPU become relevant when it’s so slow you drop into bottlenecks. I run a 5090 together with a Ryzen 7700X and that’s easily enough; the CPU usually has a load of less than 50% when SLm is running.
While it’s purely anecdotal, I’ve found so far the videos I’ve upped to 1080p (or there abouts) via SML hold up really well on my 55’’ LG C3 UHDTV, I think in part because the TV itself has it’s own AI upscaling function. I’d hazard a guess any similar TV with such a function would display them equally as well too.
That is good news…. I had not thought of that aspect at all, but it makes absolute sense.
Thanks!
1 Like
Same issue here.
1 Like
Well, given all the problems I’ve read about this update, I’m going to skip it and wait for the next one. I see that many are experiencing a significant speed loss, and others have models that are continually downloading.
As long as 7.1.0 works properly on my machine, I’ll leave it as is.
1 Like
Hello, when can we expect a Topaz Video AI in German? Thank you
Just posted a sample video to YouTube showing the three stages of an 8mm home movie enhancement using Starlight Mini for pass 1 and Nyx, Rhea and Apollo 8 for passes 2 and 3. Total turn around for a single reel is about 1.5 days, but the results are quite acceptable. This is the link: https://youtu.be/MjsKdoAls_s
2 Likes
Nyx, Rhea Manual or Auto
I’ve been trying to download both the update and models the software needs, but seems like the server is busy or something, the downloadspeed is non existant, over 10h for 270 mb, other things work normal but not the program nor the file-links uploaded here. Nowhere can I download what the software needs to run, and I’m in a production that needs this service to work, deadline is coming up.
I cant use the software since no models can be downloaded and I cant update the software.
How can I access the files or let the program download what it needs?
It’s really not bad. Especially the rear view in 60 fps.
Yes, same issue.
I scan @ 4056 x 3040 dng frames, then open them in After Effects and render to an uncompressed AVI, rescaled to 1520 x 1080 at 15fps. Topaz Starlight Mini locally to ProRes, then back to Topaz 7.1.0 NyxV3 Manual: Noise 100 all else 0, 2nd Enhancement RheaV1 dynamic, detail 100, sharpen 80, frame rate 60fps Apollo V8, replace duplicate frames sensitivity 10, stabilization full-frame strength 16, save as h264.Windows 10 using i7-4790 32gb, GeForce RTX 2080 with 8gb bare minimum system, but works well in its own time.
YES! Same issue…It’s about what 6.1GB re-download everytime! However, once I got one in the oven and say that it was going to be 48 hours or so I folded up my tent and went back to normal upscaling routine. I’m figuring it needs some time yet. Could be my RTX 4060 but I have lots of beef around that so who knows. The usual way I do it continues to work great and relatively fast.
Typical fanboy reaction.
Even if there isn’t any other software. One can stop paying for upgrades.
I did.
I do not have 7.1.1 and am stuck at 7.0.2
It seems I’m not missing out on anything interesting and I feel more at ease now.
It would even be worth some money to not get excited for a new version and then finding out nothing really has improved.
I’ve seen the introduction of Rhea and Rhea XL.
I try them from time to time, but each time they give me worse results than Iris.
Iris itself is not being improved.
The improvement I would like to see there is speed.
I moved from a 4070 super to a 5080 and expected a real boost.
It’s a tad faster, but in all one big disappointment.
But go on!
Tell them there is no competition and let them relax and sit back.
2 Likes
