Does anyone own an RTX 5070 Ti / RTX 5080 Ti and know what the current rendering times are for different Starlight models ( Mini, HQ, Fast 2, Sharp ) for 24 fsp SD → 1080p and SD → 4K?
I currently have an RTX 3070, so I’m not even considering a Pro subscription, but I’m wondering how the 5070 Ti and 5080 Ti perform. Or maybe I should wait for the 6000 series? 
I can’t find this information anywhere, so I’m starting a thread ;p
I have both a 5070TI and a 5080 (in different computers). On my 5080, Starlight Sharp renders at approximately .6 fps or 1.9 seconds per frame. Precise is slightly faster - about .7 fps. I haven’t used Mini in a long time, but it runs about the same as Sharp. I haven’t done enough work with Fast or HQ to comment. Pretty consistently, the 5070Ti is about 8% to 10% slower than the 5080.
I should note that I have my 5080 overclocked to +325 Core clock boost and +2200 VRAM clock (MSI Afterburner settings). The 5070Ti I own is VERY overclockable and runs +375 GPU clock and +3000 VRAM.
Overclocking is only marginally effective in Topaz, and you have to be careful about VRAM - push it too far and you run the risk of black boxes or dropped frames. Ultimately, actual performance is tied to how many compute cores you have available on your GPU. A full-power 5090 should get around 15% better render speed than a 5080 at OOB settings. How much overclock you can achieve, if any, will of course vary.
Did you forget to mention whether you’re talking about SD to 1080p or SD to 4K? This is crucial information.
Are you actually a woman? 
Diffusion upscaling (Starlight) is a relatively new type of upscaling. There’s still huge potential for improving both the quality of the results and processing speed, so I wouldn’t recommend changing your graphics card just for upscaling.
But if you do decide to go for it, keep in mind that the Starlight !mini! models are VERY slow (6 seconds of video at 25 fps took 28 minutes on my GPU). The result is decent but quite specific — it has strong smoothing.
However, there’s a newer model — Starlight Precise 2.5 — which I personally like much more in terms of quality, and it works significantly faster.
I have the regular RTX 5070 12 GB (not the Ti version). Today I processed a 740×460 video that was 3 minutes 7 seconds long at 25 fps. Upscaling it 3× to 2262×1386 took 7 hours 6 minutes.
Based on my observations, full 4K upscaling of the same file would probably take about twice as long — around 14 hours on my system.
However, when watching the video normally (“by eye”), I don’t see any noticeable difference between 3× upscaling (1080p / 2K) and 4× upscaling (4K). The difference is only visible when comparing static frames.
That’s a pretty impressive result! It means you can render a 25-minute episode in 2.5 days. By comparison, it would take several weeks on my card.
And since you brought it up, I’ll ask: at what speed does the Starlight Precise 2.5 render SD to 4K or x3?
If we take popular SD quality 720×576, you can’t upscale it directly to 4K from that resolution. x4 maximum gives 2880×2304. That’s why I can’t say exactly how long it will take to reach 4K — you’ll need to do a second pass or pre-upscale the source resolution beforehand. I have a personal subscription, so I can’t say anything about the others.
Sure, let me know if you can how long it takes to scale 3 minute 7 secound ( 25 fps ) clip scale x4 on Starlight Precise 2.5?
I also encourage other owners of the 5070 Ti or 5080 to share their thoughts on the main topic (scaling in Starlight Precise 2.5, Mini, Sharp, Fast 2, and HQ) 
I don’t know English, so I might be misunderstanding the translation. If you need the X4 metrics specifically, then I ran a test on a 3-second source video at 720x576, upscaling it to the maximum resolution supported by the program (X4) — 2880x2304. The result was that 3 seconds of video (25 fps) took 12 minutes and 5 seconds to process.
Speaking of which.. If the results are the same as in the cloud version, Starlight Precise 2.5 looks much more realistic than Mini.
For comparison:
Original ( 576p ):
After Starlight Precise 2.5 ( 1080p ):
After Starlight Mini ( 4K ):
After Starlight Precise 2.5 ( 4K ):
I assume that if Precise 2.5 works locally, it probably makes the most sense to scale it by a factor of 2 and then run it through Proteus or another moodel.
Best is IMO SLP 2x and then a second SLM pass as finisher. That gives really nice results - it just takes forever.
See the TopazVideo 1.6.1 patch thread for examples.
Yes, I use Starlight Precise 2.5 as a base, then I eliminate some negative artifacts - during interpolation and AV1 codec compression with a -cq 30 setting.
I own both 5070Ti and 5080.
5080 is 10% faster than 5070Ti in Starlight family.
That’s valuable information, but even though it’s 10%, I think I’ll go with the 5080—I believe it will be more open to future game optimizations. The 5090 is out of the question because I have to build a PC from scratch; I’ve been using a laptop until now, and a full build with an RTX 5090 is already nearly twice as expensive.
Yes, that is my picture, and I am a woman, and I am a nerd. 
I just recently retired from product management at a division of Insight Software, where I managed data warehousing and master data management products, so I’m a bit more computer literate than most women my age.
Keep in mind that the 5080 these days is selling for almost 50% more than the 5070Ti ($1500 and $1000 respectively).
I did that. What’s interesting is that SLP shows you 2304 in the rendering output section but it is not, when you overshoot, you always end up with a maximum output of 2160p. The question is: does the model internally generate higher than 2160 and then downscale to 2160p or not.
Right, I hadn’t noticed it before because I don’t use x4. It’s probably still a rendering limitation after all — otherwise, why impose the limit? The ffmpeg build would have handled any resolution without any problems, so it’s unlikely to be an issue with the assembly of the final file.
I tested in Astra app your theory, so I upscaled the 1080p footage generated by Precise SD to 4K using SLM. The result is disappointing—in my opinion, the 1080p footage still looks much better. The Mini smooths it out too much.
Oryginal 1080p ( after Starlight Precise 2.5 ):
Starlight Mini 1080p → 4K:
This is because you have a quite well source to start with. The SLP/SLM procedure is mainly for really LQ content.
