Starlight Precise 2.5 slow?

Topaz Video Bugs and Issues
1

In short: After 3 days and 16 hours, “Slowlight” Precise 2.5 turned a 37-minute 1080P 2.55GB absolutely normal video into 20GB of 0.85 fps trash.

To start with, NOBODY better tell me that I don’t have enough RAM or that my gear is not “good enough” for the model:

DxDiag.txt (138.3 KB)

logsForSupport.zip (2.2 MB)

A. "Slow"light 2.5:

3days and 16 hours!!! 2026-05-31 115941
3days and 16 hours!!! 2026-05-31 1159411461×161 18.8 KB

3 days and flipping 16 hours!

Did I actually need to invent a time machine when I purchased Topaz Video for its Starlight Precise 2.5 upgrade? Cuz otherwise, how the heck do I get my RTX 9090：

2026-05-31 011146
2026-05-31 0111461920×936 715 KB

B. Slow-" Trash " 2.5, precisely:

This is “video info” of the rendering result after the 3 days and flipping 16 hours

EBOD-312-1.restored_slp.txt (2.3 KB)

A pile of completely inoperable flipping trash!

When I first opened it, expecting something at least 10% better than the older models like Proteus, a giant piece of sh*t exploded right in my face!! A frozen mess! With individual frames that look nothing like the original 2.55GB video!

The original:

image
image1920×308 271 KB

After being upgraded by Starlight Precise 2.5:

image
image1920×350 357 KB

Conclusion:

Don’t expect me to spend another penny on your Topaz piece of sh**t Video!

*** a preemptive reply to:

 "Well, Starlight Precise 2.5 is a diffusion-based model blah blah blah, you should just use Proteus blah blah"

Then tell me: why is Topaz selling Starlight Precise 2.5?

If Starlight Precise 2.5 renders a lower quality result than Proteus and stuff, why do we bother buying it? What was the point of selling it at all?

Plus, I flipping paid! OK? And I should get what I paid for: not to bother with those parameters, sliding those sliders 10ish by 10ish! If I did not want an AI-smart engine that automatically renders and upgrades my video, that does everything FOR me; if I enjoyed “exploring those different quantifiers and parameters”, I could just use ComfyUI.

It is Topaz, who made Starlight Precise 2.5 so trashy and completely inoperable on the most advanced hardware at this time and age, whom you should turn to! NOT ME!

***another preemptive reply to:

“Well, but RTX 5090 isn’t exactly the best hardware at this time and age. For example, there are RTX 6000 and 6000 PRO with much more VRAM…so not running Starlight Precise 2.5 as expected is completely your fault, cuz Starlight Precise 2.5 was just for the pros”

Bi*ch, even the professionals use the same 5090, ok? It has a VRAM bandwidth much better than those 2 VRAM bloated cards you talked about! Plus, look at it: this Topaz thing does not even CARE how much VRAM you got; VRAM is useless for Starlight Precise 2.5:

2026-05-31 011146
2026-05-31 0111461920×936 715 KB

2

Going over the information you sent in your post, there are a few things to keep in mind, several of which you did already anticipate.

Yes, the model is slow, that has been shared by many users and is a known situation with the model even on high end gear like the RTX 5090. The research team is working on improving processing speeds in the future as they can. Part of that focus has brought about the Neuroserver feature with hopes that it can help with memory resource management, and processing speeds but it will take time.

The original release of the Starlight model in the research phase was too large to run locally, so it needed to process on the cloud servers, and even there it took a long time for processing to complete. Since then they have released 5 new variants for the Starlight family to be used in different situations and with different source material.

Just like all the other models, there is not one model that will work in every situation and on every video loaded in. Looking at the screenshots you attached, and the information, Precise 2.5 could be trying to overly correct this video, and in turn ends up making it appear worse. This can happen at times when using HD footage.

Have you tried the Starlight HQ or Fast 2 model to see if those produce a more desirable result with this content? Looking at the screenshots, the background details are noticeably sharper and more detailed, but the movement of the face and hair is definitely causing issues with the model, and it seems to be trying to force the facial details back into the shot.

From what I can see, the Precise model is not the right model for this scene and this context, but it has been proven to be very capable in many other situations. Just like the other models have done wonders for other videos, but they can still produce bad results in some situations.

I would be curious to hear your thoughts on what the other Starlight models can produce in those same frames, versus Precise 2.5, and also what Rhea or Proteus can deliver. If you are going to run any more tests on it, I would recommend taking that video and cutting out this scene for a shorter processing time, as in version 1.6.0 there is a bug that is causing the trimming to not work correctly with the Starlight models.

3

2 things:

A. On “Starlight HQ”:

It is almost as slow as Startlight Precise 2.5 with 0.9 fps/second. I just quit it and switched to the “more advanced” Startlight model, which is now.

B, on other models:

Speak of the Devil. I am using a preset for another video before, and the speed is obviously much healthier. The overall hardware usage is still capped:

image
image1920×840 536 KB

I will fetch you the log file later.

4

CONT.

After just one grocery shopping, this is what I got:

image
image1478×1150 61.3 KB

logsForSupport.zip (5.2 MB)

It is already hard to say that people should like the Starlight models, now that the older models just keep crushing if one just wanted to combine them into one work flow…

5

You have 0.6 fps while upscaling 1080 to 4K! That’s bloody fast.

I can have only 0.1 fps while upscaling 1242x924 to 2480x1848 on my 5070Ti, and 0,11 fps on 5080.

SLP is slow, we all knew that.

6

I’ve always been of the opinion that it’s a mistake to keep releasing new models at such a rapid pace when you can’t realistically maintain all of them. Fewer models, but better ones, would be the smarter approach. I mean, who is actually using Starlight Fast or Starlight HQ? To be honest, I don’t even use Starlight Sharp anymore.

All of that could be integrated into SLP. I still remember when a staff member on the forum wrote, around the release of SLM, that Topaz wanted to learn how different aspects of the models could be influenced so that users would eventually get adjustable parameters. Unfortunately, nothing has come of that to this day.

Where I disagree with the criticism, however, is the claim that SLP produces poor results. In fact, the opposite is true. I haven’t seen another model that achieves such a high level of reconstruction. At times, the results are genuinely astonishing. That said, it does come with its share of problems that need to be addressed—most importantly, the jittering.

8

Maybe you don’t, but Sharp is absolutely critical. Using Precise on high quality input rarely works. Precise is too aggressive with facial features and often “enhances” shades into textures (like 5 o’clock shadow turning into a bad case of acne). Precise has done wonders for blurry source material - I just used it on an old Film Noir that was a poorly done transfer from a 16mm film original…

image
image1920×932 336 KB

That is what Precise is precisely designed to do. But I find using it on good quality original content causes more harm than good. Starlight Precise isn’t the “BEST” model anymore than Proteus is the “BEST” model. There are different models for different needs. It is impossible to create a model for EVERY image problem, since fixing ‘X’ in one case will hurt ‘Y’ in another.

9

Okay, in cases like that, it can be great. The issue with SLS is that it’s a 1080p model, while NYX is used for every other target resolution. NYX is a denoiser, so it works well in your case, but in other situations it just oversharpens. What’s crazy is that although SLS is an FHD model, it doesn’t even offer FHD as an output option.

What I see is that the program keeps adding more and more models, but they can’t keep up with maintaining all of them. Maybe we’ll get “Starlight Accurate” next month, and SLP will be forgotten.

10

I’m not sure what you mean by “1080 model”. It does 2x,3x and 4x like any other Starlight model. To prepare a 480x640 original I just batch the files through FFMpeg to enlarge to 540x720 (i.e. just a bit less than PAL resolution). That aligns it the HD resolutions of 1080p and 2160p. My standard process for DVD rips is just that: FFMpeg upscale to 540, Starlight Sharp at 2x to reach 1080, and then 2x again through Proteus, pushing high noise reduction and anti-aliasing settings. I never use Nyx, nor is it automatically invoked.

Perhaps running it through Precise and then Mini may accomplish the same output, but would take over 4 days per episode, versus 2 days, since Starlight Mini runs at less than 1/10 the speed of Proteus.

11

For that: get a Mac…

12

The error you are seeing there is the system overloading and not able to process and run the Stabilization model correctly. Applying 5 models on a file is going to overload even a 5090 card and cause issues. Best to break it up into multiple passes to reduce the memory load.

13

so I broke it uh?

Then how do YOU explain what I got is 0.854fps? Did I operate the model any differently than u?

14

This is another example of “you got not what you paid for”: you HAVE the muti-layered double or trible enhancements available on the menu, yet now you are telling me “you can’t do that, it will be slow or not to your hardware’s best capibility”.

What didn’t you have that warning on the payment page or right on the main menu???

15

If, for example, your output resolution is 1440p or 4K, SLS always performs a model-enhanced upscale only to 1080p. The target upscale from 1080p to 1440p or if you have set 4K is then handled by NYX. The model is trained only on FHD (1080p), works differently from SLM or SLP.

However, if you catch the process at the right moment just before it completes, you can copy the temporary output file before the NYX stage finishes. This allows you to obtain the 1080p version without the NYX upscale. I know this because I’ve done exactly that multiple times myself.

16

Obviously, you did something wrong — your file is upscaling quite successfully.

upscale.zip (3.8 MB)

There was probably an error in the video file (source was wmv?) which is why you ended up with a file that has broken fps. You needed to fix the file before processing it.

Also, such long videos should be processed in chunks. Split the video into several parts, process them separately, and then combine them into one file. Additionally, during processing you can preview the result without waiting for the entire process to finish. To do this, copy the output file to another folder — then you’ll be able to open it (but note that this is not an official method).

You’ve already been told this, but HD videos usually upscale worse than SD ones. It makes sense to downscale first, for example to 720p. There’s no need to chase 4K numbers — it can actually worsen the result. Personally, I think 2K is quite sufficient. You should first test a few seconds of the video at different resolutions and only after that make the final decision for processing the full video.

20

but is it spposed to be that slow? I don’t think so

23

If you just actually target 1080p, as I do, it doesn’t try to run Nyx at all. I have compared running SL Precise and then SL Mini (to soften the overly enhanced SL Precise output) and I get the same, or better, results from SL Sharp followed by Proteus. The hours that I save are a bonus.

My position is that if you want the diffusion models to run locally, you MUST have different versions for different use cases. A single Starlight version to cover every circumstance would require a server level machine (with multiple 16 lane slots) and multiple server level GPUs.

Nobody wants that!

24

Excuse me, but 0.8 fps is actually pretty fast for a Starlight model variant. Just saying.

25

This will probably annoy some people, but I absolutely love starlight precise. I’ve been disappointed with most of Topaz’s releases over the last year, but Precise 2.5 has exceeded my expectations.

Did anyone really expect it to run at the same speed as Proteus or any other non-starlight model? All of the starlight models are slow, but I can confidently say that Precise 2.5 has been a game-changer at upscaling old, low quality clips. In fact I’m so happy with the results that I deleted all my old clips that were upscaled with SLM and am re-processing them with Precise.

I really don’t think you have anyone to blame but yourself if you thought a model like this would process faster than what you’re experiencing.

27

Starlight Precise 2.5 in newer version ist slower than the last version. It needs more than 4 days for rendering a 10 minutes video without upscaling.