I know - because SeedVR2 yields much worse results. Starlight works with 8Gb of VRAM, while SeedVR2 requires as much VRAM as possible.
Pretty cool! Do you keep the āMax memory %ā preferences in Slm at 100%?
Kind of puzzling that the TL subreddit gets more responses from the company than their own forum 
thanks that explains, āadvancedā sounds always good, says marketing guys 
yes I set VRAM to 100% into TVAI, for all my tests
1 Like
In my case all the āplay aroundsā on settings, discussions about VRAM didnāt lead to make SLm any faster.
Then I speak about āfasterā I think about at least 2,0 - 5,0 fps render speed and not about an increase from 0,9-1,0 to 1,2 
So, I render further my SD footage with 0,9-1,0 fps in the background and stopped with experimenting on settings.
Perhaps on one day in the future Topaz Labs decides to give us the (already devoloped) faster local model.
Perhaps on this day Iām nearlly finished rendering my old SD footage 
Just a little first hand experience:
I started my Starlight Mini adventure with an EVGA RTX-3080Ti FTW3 Hybrid (i.e. water cooled). It had a max power consumption of 450 watts, and it pulled every available watt when running SLm at 1.02 seconds per frame and heated the card to about 80Ā°C.
I upgraded to an Asus RTX-5070 Ti Tuf OC which pulls 300 watts (Asus upps that to 330 max), and averages 0.83 second per frame on the same content. I am pulling a full 320 watts but the GPU runs cooler (never goes over 60Ā°C with air cooling only).
So, Iām saving ~220 watts and getting ~20% better performance and generating less heat. It seems the overall speed of Starlight Mini is determined solely by the number of pixels per second in the input. Larger memory doesnāt speed it up, it lets it recover more details. More powerful Tensor cores helps some, but not that much (which I think indicates SLm uses a dense model). After all, even Cloud processing of SL on server grade AI systems is very slow relative to the other models.
interesting welcome to the Club I also have the Asus 5070 Ti TUF in my main PC and I love everything is made of solid metal! I get about 0.5fps with it and the same content on my at 875mV adjusted 5090 gives me 0.92fps. I didnāt undervolted my 5070Ti yet, comes next! 
Having spent considerable time SeedVR2 now, it isnāt better or worse than starlight once you configure it optimally. If you feed it low resolution tiles of your video that are overlapping at very high frame batches (Iām at 197) and blend them it performs better than starlight VR on many things (detail recreation, fine textures, text etc.) but it lacks the top notch smoothing and natural look that comes from the finishing methods of Topazā model.
Itās a heap of work but I ran a video through seedvr2 for itās superior detail and texture recreation and then starlight for itās ānatural lookā smoothing and edge handling. This approach is the best outcome Iāve seen.
3 Likes
I have a Gigabyte Aero OC 5070ti 16 GBā¦.
Am closing in on a decent undervolt that will drop power usage and tempsā¦.and improve performance.
I doubt I will see a FPS gain in SLMā¦.but in Benchmarks the performance is slightly higher and much more āefficientā as temps, noise and power draws will definitely decrease.
Will post when I feel they are totally stableā¦.at least for me.
Rememberā¦what I get will probably work with almost any 5070tiā¦.but also may not be your āoptimumā.
The 5070ti models appear to be manufactured pretty close to a ānormāā¦.regardless of company like Gigabyte or Asusā¦.they all seem to be about the same unless they have been binned for an Overclock model which will be in its title āOCā.
Not for everyoneā¦. but if you PC is in a hot room or you werenāt satisfied with the cooling of your PCā¦.. or you just like to tinker.
It may be something you want to explore.
Iād even say itās mandatory Itās simply a way to keep the card running healthier
Thereās also the option to undervolt by simply lowering the power regulator, which also causes the card to draw less power, but that doesnāt allow me to achieve the same high clock speed at the selected low voltage.
So i can recommend as desribed above, choosing a voltage that best matches your profile (try out) and drag the curve up to āvery highā MHz, then by flaten the courve Afterburner goes auto down and gives you the highest possible Mhz at your selected voltage pointā¦and this is stable.
More cards has to be tested, for the 5090 itās something like that:
850mV: highest efficiency, save 30% power (or more), but speed can drop up to 10%
875mV roughly nearby same performance as the original (when you run Vram at 2000Mhz it is slightly faster than orig card when you do nothing.) Power consume drops about 15-20%, max watts on 600V connector is about 470W instead 575W
900mV higher performance than the original, but just littel bit lower or same power consumption as factory card
925mV and higher, Overclocking, fps nerd gamers
1 Like
I will be interested in your results with the 5070Ti. Iām running a pretty standard overclockā¦power limit to 116%, voltage to 100%, core clock +400MHz, mem clock +200MHz. While running Starlight mini upscales core reaches 3285MHz and memory is running at 14200MHz. Rock solid - Furmark ran for 3 hours, Kombustor for 3 hours without errors or artifacts. Temperature hovers between 55Ā° and 65Ā°C. pulling a max of 330 watts.
I have had great results upscaling and improving the detail of my DV videos captured with Scenalyzer using SLmini. I am wondering if the new project at GitHub - NVIDIA-RTX/RTXNTC: NVIDIA Neural Texture Compression SDK will improve the speed of renders as it effectively doubles or quadruples your VRAM The gurus at Topaz should take a look at this, it may work better than quantizing Starlight for local use
1 Like
Iām not sure if this enhancement suggestion has already been mentioned, but I wanted to say that running TWO passes of starlight mini is even more miraculous for difficult source material.
I didnāt think a second pass would make that much difference, but it did.
After the first starlight pass and Davinci Resolve there were still some issues:
- The source was dark and I could not get Davinci to solve everything and the color in the dark regions was still not correct. Some middle shadows in the brighter areas still color shifted as well.
- Far away faces were still problematic: still some artifacting (eyes & mouth in particular), and still quite blurry
- I also had some residual pincushion effect (larger dots regularly spaced)
- And there remained a few, difficult to correct, tone/hue shifts upon panning
Nevertheless, I wanted to run a final enhancement pass on what I thought was the best I could doā¦ but none of the models worked very well. Each displayed their typical weakness.
On a whim, I tried a second pass of starlight on a problematic section and ALL of the issues were corrected.
3 Likes
That sounds so extreme, but maybe Iāll try it in a few years when I can afford better hardware.
Gotta try this
1 Like
Iām so far not at the point to color grade the improved Starlight Mini footage.
Iām render further the original SD footage with SLm.
Thank you for sharing your experience āwhat comes after the first SLm renderingā.
Sounds interesting!
Do you resize the rendered SLm footage again down to 640x480 or 720x576 before you render it a second time with SLm? Also, in order to ensure higher render speed?
Or do you render it with the bigger resolution, which SLm demands for the first rendering?
1 Like
I had wondered about thisā¦.thanks!
Sounds like a very solid card.
That is a fairly aggressive overclock but those temps are still very reasonable.
Now here is what Iād like you to testā¦.if possible.
While you are seeing good frame rate improvement in gaming benchmarksā¦.
How does that extra performance you are getting equate in Starlight Mini?
Are you seeing improvements in render speeds?
Does a normal Render in Starlight Mini at say .5 FPS jump to .6 or .7 FPS??
In the endā¦a gain of render speed is what we all are searching for unless you are simply undervolting to save energy, lower heat and noise and extend card life.
The original was 640x480 and I have come to conclude that I get better results (most of time) feeding SL the untouched video.
When I first started using SL, I would sometimes (not always) run a light Artemis and/or some basic lighting/color correction beforehand. But, turns out, SL does a better job than what those pre-SL enhancements were doing. Now, I do all of the touch-up at the end.
For SL pass #2. I just fed the output from pass #1 into TVAI and ran SL again.
Pass #1 produced 1280x960 and thatās what I used.
I thought it best to give SL #2 all of the video information to see what would happen.
The speed was the same as usual: 1fps.
Fortunately it was a relatively short clip.
After SL #2, I did Davinci corrections and the upscaled with modified Nyx (sometimes will use Proteus or Rhea instead).
