New results using AI to cut render time?
jim_doria
Posts: 113
I may upgrade my hardware soon, but for the time being my GPU is a quite anemic. Has anyone found a good way to use the power of AI to reduce just the "waiting for renders to finish" part of the process?
I'm not talking about image generation, or img2img, just using AI to take a 5 minute Iray render and make it look like a 5 hour Iray render. Or a 1 hour render and make it look like a 24 hour render.
The last time I experimented with AI denoisers, a lot of them seemed no better than unsharp mask at reducing render noise. (Maybe because the graininess of Iray renders isn't actually noise but missing data? IDK, that's above my pay grade.) Topaz, which used to be a big player in this space, was also pretty expensive, and I wasn't impressed by the results when I tried the free tiral. But that was one or two years ago, and a lot has changed in the AI world.
This is hard to Google because I get lots of spurious results - Adobe Firefly especially - and because anything older than a year, or even 6 months is out of date already.
Any hot leads?
Daz 3D is part of 
Connect
DAZ Productions, Inc.
224 S 200 W, Suite #250
Salt Lake City, UT 84101
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2024 Daz Productions Inc. All Rights Reserved.
Comments
The only thing like that I have found is using blender cycles to render instead of iray and DS lol. The denoiser in cycles blows the one in iray as implemented by DS outta the water. And if you spend a few bucks to buy turbotools for it, it gets crazy fast.
@TheKD - Thanks for the tip! Requires a pretty massive change to my workflow though. I want awesome results with no effort! Isn't that what AI is all about?
I already use upscaling and interpolation apps on many of my iray animations
are plenty of options out there free and paid
Topaz Gigapixel AI is meant to enlarge images and it works well filling in pixels. Today is the last day of their Black Friday sale and you can download it and try it out with a watermark before purchasing. If you bump up the size quite a bit and download all their "AI models" that are different ways of enhancing the image, then save it, and reduce it to the size you actually want, it might suit your needs.
back in the days of Poser 5 I wanted big pictures with lots in it.
my Pc would just fall over and cry. let alone render anything.
so I had to do multi renders. the image below (posted on May 03, 2005) is made up of aroung 15 (maybe more, it was a long time ago) renders
and put together in photo shop. render time on each render was around 20 mins.
there is more than one way to get fur from a cat
Generative AI is its own Algorithmic process.
Iray rendering is it’s own hardware computational process.
The two are not related.
The AI you are asking for is indeed a denoiser. That is what the denoiser is doing, it is guessing what those missing pixels are. The more known pixels you have, the better the denoiser performs. Denoisers also work better at higher resolutions, again, because there are more pixels to work with. Increasing the size of your renders can thus make the denoiser work much better, and still reduce render times. You have to manually stop the render to save time, as there is no automated method, if that is what you are looking for. You can set manual limits like capping the convergence at say 75% instead of the default.
Do not bother with Topaz for Daz Studio. Topaz is NOT designed for this type of denoising. Topaz is designed for noise in real photos. The noise you get in rendering is not the same kind of noise, and why these denoisers perform so badly.
You need a denoiser specifically designed for rendering. That limits what you can use, because these are not something you can buy from software like Topaz. I use the denoiser made by Intel. It takes a bit to set it up the first time, but once you have it set up you can denoise pics in seconds and set up batches as well. IMO it sure beats trying to work with Blender for anything. Please no, I would rather scratch a chalkboard than use Blender.
Here is a thread.
https://www.daz3d.com/forums/discussion/334881/a-i-based-open-source-de-noiser-for-daz-studio-pc-and-macs/p1
Mcasual wrote a script for DS to use this denoiser. I think it is extremely convenient and works pretty well. Denoisers are not going to be perfect, for the reason I opened up with. But they can do pretty well if you know how to work with them. On its own, it doesn't save you time. But if you look at the pic and decide "hey, this should denoise alright", you can stop it and denoise it. You will get a feel for how well it works after a few tries, and what general percentage of convergence tends to get the results you desire.
I use Waifu2X-Caffe to upscale and Flowframes RIFE to interpolate
@outrider42, @WendyLuvsCatz - Thanks for the tips! I knew NVidia had a denoiser (This looked promising but I don't meet the minimum reqs GPU-wise: https://declanrussell.com/portfolio/nvidia-ai-denoiser/)
I'll take a look at the Intel one, and look into the other tools. Muchos Gracias!
I just researched what you ask about yesterday. I would write a feature request to DAZ 3D to upgrade DAZ Studio to use DLSS 3.5's capability as an option in the iRay render engine to be concise.
So given what you said about your current card and what I just told you what you need in an RTX 4000 series video card really. Any of them will do but I researched today the sweet spot for cost vs performance is in the RTX 4070 12GB card.
Using that information, I just ordered the parts to build me a new desktop in the next couple of weeks and after realizing 2025 and the next generation of nVidia GPU/AIPUs will be here before we know it I decided not to get an RTX 4090 as they are going up suddenly in price being in low stock and instead get the best value for today's current GPUs, which I read on multiple sites as being the RTX 4070 12GB (Nvidia finally has a Goldilocks graphics card | Digital Trends). I ordered one and am picking up tomorrow at Walmart. I have to wait on the new SSDs, SSD enclousures, motherboard and so on to be delivered before I can test that card but returns are allowed until 31 Jan 2024 and I'll have the parts I ordered within 2 weeks.
What you are talking about is something the folk that work for nVidia on the iRay engine could integrate as an option in the iRay render engine as used in DAZ Studio. Read this article for why:
Nvidia's DLSS 3.5 is what ray tracing always wanted to be | Digital Trends
Basically to summarize that article, the iRay render engine in DAZ Studio needs the options SR (Spacial Reconstruction), FG (Frame Generation), and RR (Ray Reconstruction) so that you can turn on and off, so when you turn it on you can render at FHD or 2K and use DLSS 3.5 to upscale the results to 4K. It probably will solve a lot of out of video card memory problems that some DAZ 3D products like some of OOT's hair products can cause too. (well, they would have to fudge it because it's not like you really have 2 or more frrames like you would in a game or animation that had ray tracing).
Now this new DLSS 3.5 was trained on a much larger data set than prior versions and so it's a prebaked AI algorthm that uses equations gleened from solving all those data sets.
@nonesuch00 - Cool, thanks for the great info! I've been looking at my next hardware purchase and this info about GPU capabilities is right on time!
Does DAZ Studio automatically use the latest IRay engine, or do they have their own version? I'd imagine I'd have to upgrade DAZ Studio, which I've been putting off...
DAZ Studio is close to the latest iRay engine but not always the latest. The DAZ Studio public beta will have closer to the newest iRay engine but not always and usually it's not long before DAZ Studio release follows with that same iRay version..
In my tests of my RTX 4070 12GB today with DAZ Studio public beta 4.22.0.7 it's been running about twice as fast as the RTX 3060 12GB I used prior, eg the test scene in this forum section used to test render speeds in DAZ Studio ran at about 299 seconds with the RTX 3060 and 139 seconds with the RTX 4070.
It just depends. Sometimes Daz can be pretty up to date with Iray, and sometimes not. We are currently running Iray 2022. Iray 2023 released in June. The Daz beta tends to get stuff faster, there have been times when the beta has received the latest Iray just days after it was made available by Nvidia. But not this time.
You can scan the Iray benchmark thread for some user scores with a little scene we made. The 3060 is still being produced by Nvidia, and the 12gb model is still one of the best bang for buck cards you can buy. Obviously a 4070 would be better, but it costs more, too. So you just have to weigh out your options and decide what works best for your budget.
VRAM is the other big stat. You need memory to keep the scene on GPU, or it drops to CPU only mode. So having enough VRAM is vital. Because the fastest GPU does you no good if you run out of VRAM. It basically becomes a paperweight at that point.
You have a challenge though. You need to consider how large your scenes will be, and that might be tough since you do not have a viable GPU now. So it can be tougher to gauge how much VRAM you may use. Iray can be a massive memory hog, but you can also work to optimize a scene to fit it into memory.
You also cannot forget RAM. You need that, too. You will need more system RAM to be able to use that extra VRAM. If you don't use all 12gb of a 3060, then maybe it isn't a big deal. But if you do use closer to 12gb VRAM, you might also need 32-48gb of RAM as well. This is hard to determine because every scene is different. I personally use a lot of RAM because I frequently hide things in the scene tab instead of outright deleting them. I also started using larger and larger textures and assets. I used to be extremely careful about memory because I had small GPUs. I started with 2gb. Then I had 4gb, later 11gb, and now 24gb VRAM. I also upgraded to 64gb of RAM. So perhaps I have developed a few bad habits with this extra memory, but I still know what it is like to have a smaller GPU and to be memory constrained. It is not fun!
On the plus side, though, I can tell you this, whatever you do get will be a lot faster than what you have now, assuming you have the memory. Like comically faster, if that makes sense. You can try out the benchmark for yourself, and use that as a ballpark guide to how much faster a particular GPU might be.