It depends what video. Starlight has done the best so far in lineskipped aliased/moire video (if you take a look at my example above) in reducing those artifacts while Artemis LQ is a close second. Granted both aren’t perfect and still have their problems but really keen to see how much they can improve on it. Also, I know lineskipped slowmo video is probably niche footage that not everybody is doing so I understand if some aren’t seeing that much improvement.
edit: oh my mistake, I saw that you were referring to another video.
I don’t understand what you mean with this statement. How are clock speeds lower at lower voltages than default when you maintain the default curve? Not that it really matters that much, since 99% of the time it’ll either use 3-4 of the highest voltage points under load, or at idle, the lowest. Maintaining the default curve will be more stable. Not by much of course, but it will.
I’m not saying your approach is wrong. But it is not a good approach when trying to help someone who’s never fiddled with undervolting, overclocking or the likes before.
An undervolt to <=800mV is pretty extreme as a starting point. It’s more prone to crash under transient loads, which is exactly what @mikmod1 was referring to regarding TVAi.
The 3090 and 3080, while both are from the same generation, there is a significant difference in silicon, memory config and power requirements. Just compare the amount of cuda cores and tensor cores between the two and you’ve already got yourself two GPU’s which will process videos with generative AI very different from one another. It’s also the likely reason mikmod1 is seeing higher transient spikes compared to what you’d do on a xx80 card. At 800mV you don’t have any margin, and transient loads will be more prone to cause a crash. GPUs also work different at different temperatures. Another factor that matters with a significant undervolt.
What software are you using to monitor gpu power usage and at what interval?
Next time you check, could you perhaps use the one in Msi Afterburner so we can get a view of voltage, power draw in Watt, power usage, GPU usage, clock speeds and memory controller load?
Pretty crazy that it is hitting north of 120% with a slight undervolt and a power limit set to 85%. It’s almost 50% more than what you are allowing it.
Have to say I really have doubts about the future of this product (and company)… you are trying to push customers down a route that 99% don’t want to go down.
I will never pay for cloud rendering, not least because some of the videos I’m converting are highly confidential, and we have no way of knowing if any data is retained by yourselves… or if you have some caveats buried in your T&Cs that allows you to “own” or claim ownership of the footage. Plus the cost is currently ridiculous… even if you divided it by 50 it still would be expensive for me.
Think I got quoted something like $120 for a 1 hour video when I tried it a few months ago. Totally ridiculous when I might have 30+ hours of footage to upscale every month. I could buy 2x 5090s for the cost of 1 month’s worth of conversions!
You cannot force people to use and pay for something they do not want by making it the only option which the cynic in me says you’re doing.
I cancelled by subscription late last year when it was clear there were no worthwhile updates forthcoming for local rendering. Unless you move back to focusing on this then I won’t be renewing ever, unless something comes along that makes it worth it as a one-off.
It’s System Informer, with kernel driver turned ON in settings, at 100ms intervals, all as administrator.
Exactly, it shows crazy values. It’s like Nvidia hardcoded that behavior in the driver.
But while rendering SDXL image, the reins are in place, somehow.
Could you provide more examples with multiple faces (in color) and landscapes?
Also, how much slower is the render speed on a Mac M4? If I can get results like what you show (and definitely more examples of very blurry faces and landscapes) – then I would definitely purchase a copy.
If you watch the Starlight NASA launch example on YouTube, you’ll see an eyeball on the Astronauts forehead in the before and after. The results are amazing, but not perfect.
Interesting how you won’t give specs on “large model” and “high-end machines”…
I’m sure every user of this software has terabytes of storage and RTX cards.
Personally I keep 50 terabytes available at all times and have 4 machines running dual RTX 4090s. I’m pretty sure I can handle anything you throw at me, and I’m positive I am not alone.
Just be honest:
Topaz feels left out of the data gathering of copyrighted material, and this is the way to get it legally and charge your users for the priviledge of training your models, including Starlight.
Some of us even enjoyed using the GPU on our CPU to use additional processing for renders. Now only available for more cash. I understand you have to pay the bills.
Open source models are gaining ground every day so I think you’d better re-examine your priorities for long-time customers such as myself.
Yeah I get that – but amateur video footage (in color) from the 1980s (not film) examples with multiple faces moving around would be a common use for this “cloud” feature.
I added some VHS footage to the online browser and it cleans up really well…text excluded though. LOL.
And much better result from a different source.
Although the text is soft, the diffusion AI is treating the text as though it’s very narrow which I wish it didn’t. A heavier more fuller version of the font would have been preferred. That said it did a pretty amazing job.
Actually, the online examples posted for this were so small it was impossible to see any difference at all. And even though it takes my 4090 system a few hours to crunch through the enhancement of a 90-minute video, it is really no problem, and at no extra charge.
I remember when the server credit system was introduced someone stated that this meant the downfall of this program, and that soon we will have a A.I model that only will be used with credits, and here we are. I got a $2000 dollar pc, and a A.I model i can’t use. 100% sure that within 4 years this program will be 100% server credit only because it will gets them more money.
Within 4 years, someone will probably have posted an AI diffusion-based video enhancer open source app to Github and a commercial app like TVAI will have to be downloadable in free and pro versions like DaVinci Resolve to attract users.
The window to make a profit from bleeding edge tech shrinks rapidly from the first day someone thinks of it.
