Is this only true for SLP or also for SLM?
Exactlyā¦for better quality sources, I recommend Starlight Sharp to 1080p and then use a second pass of Proteus or Iris, depending on what sort of remaining processing you want. According to your tastes, you can add grain, sharpen more, reduce noise, recover additional detail, etc.
I think SLM does not, but Iām not sure
Most of time I use SLP for upscale because reconstruction level is unbeatable, but Iāve had cases where running SLP first didnāt work well because human skin and faces looked much worse.
However, SLM first can be the better option, pends on source. If the upscale looks too soft, you can also post-process it with Deblur. The pixel structure simply needs to be made coarser, with larger and more pronounced blocks. working with soft results is generally not a bad thing because it reproduces very small things good, and there are ways to make it sharper afterward, for example using Iris MQ 1x + Gaia as second model. In this step, you also can add grain if you like.
I want to convert some SD TV shows (480pā576p with a very low bitrate and macroblocks) as a hobby for Plexāshows that no one will ever remaster, so the quality has to be good enough to trick the brain and avoid that cringe moment when someoneās face changes. Iāve seen the workflow: Starlight Precise 2.5 to 1080p ā Nyx 1080p ā Proteus x2 + HDR colors and grain in DaVinci.
@Dianac, I work as a programmer at a large company that sells computer hardware (I can buy a computer at cost price, without the storeās markup), so from what Iāve seen, the difference between a computer with an RTX 50070 Ti and one with an RTX 5080 is about $150ā$200, so I think Iād rather pay the extra than regret it later.
And in the futureāwho knows how these models will evolveāit would be good to be a video editing specialist in my area.
I am doing the same thing. I just finished a months long project to upscale the entire Star Trek Deep Space Nine series. I have also upscaled a number of Film Noir movies that are in the public domain and which are only available as 16mm transfers.
My technique is to use ffmpeg to convert the SD originals to a standard resolution of 540x720 at 29.97 fps. I also make sure that display aspect ratio is the same as stored aspect ratio and, if necessary, convert from VFR to CFR.
I then use Starlight Sharp (for reasonably good quality DVD rips) or Precise (for blurry 16mm transfers) to go from 540p to 1080p. I avoid Nyx, since the Starlight output is already somewhat denoised. I then upscale from 1080p to 2160p with Proteus, using manual settings, to repair any remaining issues (additional noise reduction, anti-aliasing, detail recovery/enhancement, etc.).
I am currently working on upscaling āThe Streets of San Franciscoā, with āThe Invadersā and ā12 Oāclock Highā in the queue. I may need to add some storage to my Plex server to hold them! 
As an example, Taking the 540p original as described above, upscaling a 45 minute TV episode from 540 to 1080 takes about 41 hours. The Proteus step takes about 2 hours.
This is on a PC with an i9 14900 CPU, 64GB system RAM, and an RTX 5080. The 5070Ti takes 50 hours and 3 hours for each step, respectively.
Look at the womanās expression in the original frame and then at each of the upscaled ones. Her expression shifts slightly with each of the upscale models, and it actually shifts the tone of each of the scenes. Probably not an issue when you see it in motion, but interesting to see at the frame level.
@dianac Also a PM here, was head of the Platform team at Marketo, where I designed and created 100+ Splunk dashboards (that was how I relaxed, oddly enough). It was like hunting for buried treasure.
ā
I am currently working on upscaling āThe Streets of San Franciscoā, with āThe Invadersā and ā12 Oāclock Highā in the queue. I may need to add some storage to my Plex server to hold them! ā
And HDDs are, unfortunately, ridiculously expensive right now 
āI am doing the same thing. I just finished a months long project to upscale the entire Star Trek Deep Space Nine series. I have also upscaled a number of File Noir movies that are in the public domain and which are only available as 16mm transfers.ā
Donāt you use DaVinci to first aggressively remove compression blocks and reconstruct details using some model, and then use Starlight to repair the sections cut out in DaVinci and replace them with the ones that had artifacts?
Of course, this wonāt work in severe cases, especially for someone with a keen eye for spotting any artifacts , but itāll do the trick with higher-quality footage.
At least your series is available on DVDāthe first one I want to upscale is a 150 Mb / 25-minute web rip
Iāve had really good results using Topaz and Davinci Resolve (the free version) together. Two examples:
- For dark footage that canāt be easily brightened (because saturation/clipping etc.) I use the SDR to HDR Hyperion 2 model and boost the luminance as needed. Then I overlay that clip on the original in Davinci and use the Composite āluminanceā mode at 50% opacity. That pulls the light from the Topaz clip but keeps the original coloring and minimizes artifacts I previously got from boosting luminance on the original.
- Topaz is already pretty good at de-banding, but I can also take the Topaz clip, duplicate it in Davinci, and apply the debanding effect. That softens the clips details, so I overlay it on top of the base clip with Composite mode of āColorā and opacity of 50%. That blends the softened colors but keeps the underlying detail.
The two work really well together and it means I donāt have to be a color-grading expert to get pretty good results.
The one drawback is that sometimes Topaz rendering slightly offsets the clip frames. Iāve learned how to use sub-frame slips in Davinciās Fairlight audio engine to line those up, but it can be a bit of a pain. Still much fast than trying to do a full color grade on sub-optimal footage though; as it can sidestep the need for key-framing
I actually use DaVinci after completion of the upscale process for content that came from DVDs. However, I have one Film Noir that is a real mess. It is a bad transfer to digital from a 16mm film that was then super compressed down to around 400Kbps. There are scenes that just look like a black screen with some vague movement in the dark. I ran that through Resolve first, adjusting brightness, contrast and gamma, adding gain control (blended back into the original) and fully desaturating the output. I also need to resync the audio track that was slightly behind in the original.
I then ran it through Starlight Precise followed up with a run through Iris to clean up the faces that SLP recovered. Unfortunately, I donāt have the original file on hand or else Iād post a before and after.
SLP was amazingā¦I almost think it was trained on Brian Donlevyās faceā¦it created a very accurate reconstruction in scenes where his face was just a smear.
Well, Iāve ordered a computer with an RTX 5080 
Now Iām thinking about which Topaz subscription to getāIām using it for personal projects, but Iām wondering if Starlight Precise alone will be enough (I know it works best right now, but itās not necessarily the best choice for all types of content), so Iām asking subscribersāwhat do you think? What are the real benefits of the Pro subscription compared to the Personal one?
Are the Fast 2 and Sharp models actually significantly faster for higher-quality content?
For those interested:
Graphics card: RTX 5080, 16 GB VRAM, 32 GB RAM
Video duration: 10 minutes 0 seconds
Input resolution: 720x576, 24 fps
Model: SLP x3
Target resolution: 2160x1728
Rendering time: 9 hours 26 minutes 22 seconds
Average processing speed: 0.424 FPS
Nice performance, thanks for sharing. Wish I could get a 5090 
but would that be worth the extra $?
In my opinion, itās definitely a gimmick.
With the 5070 Ti and the 5080, thereās only a slight price jump, but the performance gain is a real 10% (according to what users have reported), which saves us 100 hours over 1,000 hours of processing.
On the other hand, the 5090 is three times more expensive than the 5080āit makes more sense to buy a second computer with a 5080, get a Topaz subscription, and run the processing on two workstations.
The only potential benefit I see is for new models that will be released in the future and require at least 30 GB of VRAM, but I suspect it will still be more cost-effective to upgrade from the 5080 to the 6080 than to overpay for the 5090.
Curiously, I have 5070 Ti and 4070 Ti Super cards installed in my two Ryzen 9-based rigs, and the latter definitely seems to perform almost as fast as a 5080. Iāve only just started experimenting with Starlight Precise, specifically with Blu-ray rips of Babylon 5 with startling results, but I suspect Iāll get equally impressive output from DVD-quality material. Will need to dig out my Region-2 DVDs of Deep Space Nine and Voyager.
I wouldnāt buy the 5090 today anymore, the price exploded. When I bought mine, the situation was different. I paid around $2,300 for it, and that was when Starlight Mini was still relatively new. In its early days, SLM used 30 GB of VRAM or even more, users had to lower their RAM setting otherwise, the model would use a lot of shared memory, which made SLM unbearably slow.
Iām currently testing Hyperion 2, and Starlight looks like a really fast model compared to it
Deep Space looks amazing after Starlight is done with it! 
Iāve already upscaled both DS9 and Voyager previously, but by using Iris. I had done the first two seasons of Blakeās 7 using Starlight Sharp (with a second pass to convert to HDR), but I suspect re-doing it with Precise may be on the cards!
I was lucky enough to score a 5070 Ti a few weeks before the prices exploded, but even before the AI apocalypse I wouldnāt touch a 5090 with a barge pole! If all of the Starlight models can run on higher-end Intel Arc GPUs, Iād opt for a Arc Pro B70 instead.
