Forum Replies Created
-
Posts
-
I do AI for a living. Agree with a ton of points made here, but a few missing.
1. This stuff started being able to make small 32x32 images five years ago. Imagine how good we'll be five years from now (the latest models are beginning to make small videos now). Then when it's completely photorealistic and commoditized it'll be everywhere.
2. Copyright is dead. There's no way to prove what's what in terms of training images or weights of a neural network. And much like policemen didn't go to people's houses to check they weren't using Napster, they won't be checking you don't train on some large corporation's cartoons either.
3. I think in the long run consistency will creep in those AI tools (it's hard to say create the same character twice, or tell them to do another camera view of a given scene now). This will come via developments in neural rendering. When that happens and you can easily do a series of images, comic-style, it might give even medium-level 3d tools a run for their money, so I doubt Daz will have much to offer then.
Bonus observation: if I can stick 10 new portraits on DeviantArt with Stable Diffusion and Lightroom in a (busy) evening... does this mean soon there will be much more creation than actual eyeballs to appreciate the art ?
Yeah, I mean, we all have slightly different use cases, and that - photorealism - is what they were trying to sell, because 8k textures, right ? But even that they couldn't execute correctly.
Improvements to the figure are very incremental and some parts go backwards - you want a better portrait, but it's clear that say, ears are much worse and distractingly bad now. So now we'll need Natural Ears instead of Natural Eyes... not a clear sell at all.
I think one thing that helps save Daz is realtime, instant dForce : would fix polygonal hair & geometric issues with clothes, also would bake in soft body physics. It's possible with AI methods in 2022 but they don't know/can't afford the coders to make that happen. And that's a cost to their business, not an opportunity to immediately bring cash in.
Thanks for your answer ! I like hi-res woods and marbles 🙂 More seriously, indeed in your case it served no purpose, but I really think this is only an issue because Daz doesn't allow for user-defined prerender texture resizing, something one could implement in 100 lines of (Python) code or less.
I also agree that in many instances the obsession on megapixels is sad opportunistic marketing hype (like a lot of Daz things...) - one would not put 64k textures clothing on a G2F - there is a clear point of diminishing returns, but due to the nature of raytracing, that point sits at a multiple of your screen resolution like @Littlepleasures mentioned, so we could easily see quality gains till 8k or in the limit 16k textures become widespread.
This obsession with texture resolution is bizarre. Obviously more information is always better. Downsampling is trivial, and upsampling (superresolution) is now realistic and easy up to 4x with AI-based techniques (GANs, diffusion). That stuff is pretty user friendly now once one has located the relevant .jpgs or .png, not sure it's worth making a fuss about.
(Of course there is a world in which DAZ would include those tools and enable them on-the-fly before Iray passes textures into GPU VRAM, but one can dream).
I'm disappointed in the wardrobe essentials. It's 2022 and we still don't have a photorealistic pair of jeans. (We have an androgynous figure with bikini, swimwear, and 'diva desire' though...)
I wonder how long they can keep doing this. In a couple years, G11 will have 32k textures and crash on startup; no matter which GPU you have.
Its hair will still be polygonal, unless they give outoftouch an explicit monopoly on the store.
dForce will recognize a single of your two 4090Ti's and take 15mins to run, but you'll have realtime physics on the AR app on your phone.
@Littlepleasures thanks for a detailed and well-reasoned post. To clarify when I said 'fix it ourselves' I meant the obvious loading times elephant in the room (and possibly include basic runtime versioning for recovery, git-style, as @Amberkat once mentioned).
As for the coding, you make my CS101 point for me - I hinted in a post above at the MySpace-esque level of software engineering in play. Daz is never gonna be Maya/Arnold or even Blender/K-Cycles for that matter. But assuming a beefy GPU and minus maybe 3 4 glitches less, it could be a functional toy.
I do AI research for a day job. The stuff I see is 20 years ahead of the QT nonsense Daz offers. It is progressively percolating in the 3D community and already offers tantalizing possibilities like machine-learning powered shader conversion (to get out of the Daz ecosystem conundrum) and the like. Of course, nobody's willing to pay the upfront development costs. Razor and blades, like you said.
indeed. at that point we might as well ask them to open-source the program and fix it ourselves... not like they treat it as valuable anyway, they seem all-in on their digital asset catalog being their source of economic value added.
True true. On the other hand, G8F loading time seems to roughly scale linearly with number of chars (I'm at ~2min with around 400 G8Fs), which points to just about the most naive code writing you could imagine.
Said otherwise, it's a bit weird that we users have to expect DAZ coders not to know about fast sublinear algorithms or hashtables, and yet here we are. Like, imagine if login time to your favourite social media had grown linearly with number of users, all the way to a billion.
Note that the above forms the basis for all modern CS101 courses and all software engineering interviews ever, hinting at DAZ being treated like abandonware by management. Fixing this would certainly be easier than trying to integrate blockchain smart contracts and NFTs into the program...
Has anyone tried the script at https://github.com/TheNoobDucky/Daz-Package-Manager ? It seems to be doing more or less what @Legolas18 suggests (I haven't tried it or Turbo Loader yet).
