AI is going to be our biggest game changer
This discussion has been closed.
Daz 3D is part of 
Connect
DAZ Productions, Inc.
7533 S Center View Ct #4664
West Jordan, UT 84084
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2025 Daz Productions Inc. All Rights Reserved.
Comments
I'm getting less concerned about its ability to steal art from other artists. No matter what I do, I can't get anything remotely similar to my art or likeness by putting in my name. Maybe it would for more famous people. But even inputting Elvis Presley or Jim Morrison didn't really look like them but just an essence them. But things like Wonder Woman could be pulling in from movie posters and comic books, not sure. My artist name Wonderland just brings up Alice in Wonderland stuff but I'm wondering what other artists names that are more specific would produce.
Ok. Thanks. Sounds pretty interesting but doesn't sound like it would work for narrative work.
Sorry. But AI's still have a way to go if they can't even generate a man-eating chocolate chip cookie *rolls eyes*
We are not nearly famous enough for that to happen. Even when your prompt asks the AI to copy an artist's style, like Artgerm or Sakimichan, it rarely looks THAT similar. It can copy things like anime series and celebrity likenesses very well, but it wouldn't know you or me from Adam.
Ask Midjourney to make that, it probably will do a very good job.
that would be a interesting challenge... but even conventional intelligence may misunderstand your instruction.
Not unless you train it with your provided images
on a local SD install on your own
GPU .
Will you settle for a cop and his giant cookie partner going undercover? (I know they're low quality, I made them in about 30 seconds)
Wolf, have you trained any pure 3D characters? I'm very curious to know if it's possible to get non-3D results using a LoRA or checkpoint that was trained exclusively on 3D renders.
oops, forgot one. "Cookie Crooks" (1983) starring Alec Baldwin and Joe Pesci
Theoretically Daz should provide an ideal environment to set up optimal images for LoRA training... might be a worthwhile endeavour.
combine that with the new controlNet extension and already you're putting together a formidable suite of programs that can tackle a wide range of subjects.
i know its hideous but it's just a bit of fun... i aint got much time to spend on other peoples whims.
That is amazing, is it Midjourney? People have come up with some methods for getting fairly consistent characters in Midjourney (there are videos on Youtube), but you don't have as much direct control as you do in Stable Diffusion now.
Yes it's MJ... you can't beat it for quick n easy but i much prefer SD (or a combo)
One of my favorite combo results
Nice, thanks. I do think Midjourney gives better results out of the box and with less fiddling than SD, but there's more creative potential (and potential for using 3D renders) with Stable Diffusion.
i think the fiddling will become the most rewarding part of the process... the results i may or may not care for.
But i'l always get satisfaction from the process.
deleted
So you guys had 100 tries at getting a decent image out of these things, and one of them looked pretty good. I have a 1st Class Hons in AI but from the 90's (not pulling rank, I just thought about these things maybe more than most people have) i.e. before the "Deep Learning" breakthrough. I thought back then and still think today that the hype this stuff generates with our already primed imaginations (primed by science fiction) is far, far greater than the practical reality. If you could show me where these things are useful or likely to be? You type in some words and it produces something with the same statistical properties as the 2 million tagged images you trained it on. You can't actually ask it what those properties are (that would at least be interesting).
There are some very focused cases I've seen where it seems pretty good and useful, like Cascadeur physics, protein folding and Flowframes video upscaling. But in the general sense I don't think anyone's going to be out of a job in the creative industries any time soon. The lack of control is the main problem here. You give it a word soup and it returns an image soup. That's about it!
I'd rather say "where we would go", which means both that humans design those systems, as well as humans decide if it's accepted at all. In addition, the purely generative systems won't get more magic and more precise at the same time, technically. It doesn't progress linearly, exponentially neither. It's not certain that we will be that much better within 10 years. Not saying such can't happen, but it's clearly not certain.
Concerning precision... a way more complex system, consisting of different types of components, including different types of ais, may make more precise posing possible for the generator. That's light years off the current implementations, i would say. Can't fully foretell the speed they're moving at either, though.
Actual advancements i could imagine, would be using generative tools (or just algorithms, like daz3d scripting) to create images, that you can be pretty sure of, that they're tagged correctly, and an improved learning+detection system then is capable to run detection much better than what you get by "ordinary" supervision (so far). That's not yet the generator. Maybe it'll somehow be possible to use the inpaininting abilities and design systems, one of which poses something like the shapes or say wireframes rather, and the other doing texturing and effects, which when combined with a generative system with the actually artwork-wise precious training data, will then interpret stuff in a feedback loop (somewhat) to actually create stuff with precision and taste and sugar (and so on...). We're back at what i wrote above, then... "currently sci-fi", "maybe in range within decades".
(Concerning accountability: With the current lawsuits no one who trained on images against consent of rights owners wants accountability. Competition might, though.)
(With very good detection, there could be some shortcuts to some quality, e.g. if you can make different ais for "objects" as well as background and postwork/rather - that would allow for a relatively dumb composition. But those kind of systems already would be more complicated to build, and likely clearly more expensive to train. Plus, of course, the yet-unknown-to-me/what-i-forgot-to-mention...)
Eh, I've got about a hundred very good tries saved in a folder, so I'm not complaining.
I don't think you've paid any attention to AI since the 90s either then. The addition of ControlNet to Stable Diffusion basically solved character posing and composition issues overnight.
Lol, I can't get jealous of my own art, can I?
If you are anywhere on the internet it should find you. If you put your name in the title of your art before you post it to major sites, it should find you. Somehow it found my watermark and put a mixed version of that on generated art and I could tell it was originally my watermark but thankfully, not my art. Somehow it can read print imbedded in images. Out of curiosity, I tried putting in Mousso, Daz3d, but I just got a weird mix of ad copy with illegible print on it with some interesting characters that didn't look anything like Mousso's.
Another time, I tried to put in Shag the artist's style in there and it almost banned me lol, so I put in his actual name Josh Agle and the style was fairly similar but no where near as good and very convoluted. I think it may be trained to alter everything so much that it's only reminiscent of a certain famous artist or celebrity. I wasn't using the prompt photoreal, but the celebrities looked more like an expressive artistic depiction of them and some were barely recognizable. Elvis Presley's hair and leather jacket had more likeness to Elvis Presley than his face. But I guess if you kept trying you could get closer. It might be programmed to avoid lawsuits lol. Out of curiosity, I'm going to try the photoreal prompt with celebrities and see how close I can get. If it's too real, that could definitely be a problem for them!
Well, I don't think celebrities have to worry. The hair may be right but not much else.
Maybe. The most sensible thing would be to not train with images against the consent of the rights owners in the firs place.
More complicated, slower, more expensive - yes, but societally more sane. Could be from the cookbook of modernist tech people: Simple, fast, for free.
Other's call it the devil's path, but ok ~ i'm just passing through...
Technically, it's encyclopedic by nature, as it trains to reproduce source images as close as possible. Theoretically resulting in "technically no style", because it learns all styles (in a way). Of course it's not magic, but it comes close to both the statistical part due to tagging and model size, and the encyclopedic part, due to being trained to reproduce as closely as possible (with the model size, source image resolution, and so on). So to me it's no wonder that you can actually get such and similar results. A sufficiently large model with perfect tagging might perfectly reproduce, if you train it on high resolution images. From there on it would be a fundamental decision about at which numbers it's safe to assume "fair use" - we've had distantly similar questions about thumbnails in some context.
TBH face recognition is the hard part since the beginning of the AI mostly because even with humans modern science is not capable of giving a realy complete answer what makes us so good at recognizing faces, so it'd be unfair to expect better. Lots of human artists fail at making portraits too.
This might be of interest here:
I'm no spring chicken so I have to say I've heard this before with other technologies, more than once. Nothing came of them. Worse perhaps the more you increase the complexity, the less control you have over them. They're pretty opaque already. You wouldn't put a black box in charge of something important would you, or make your studio's £50m investment dependent upon it.
I think the limitations will be centred more around legal restrictions than creative ones... If the Stable Diffusion database is determined to be transformitive instead of derivative, then that will open the door for all sorts of applications to use it - we could even see a future where you could create a 3D scene as normal and then select an art style for your renders - and the same thing is happening with music and writing as we speak.
I quite like the ControlNet OpenPose
shame batch image won't work for me with it for making videos
There are creative restrictions too, also with transformer technology. The point is YOUR creativity with using such tools, though. So if the tool enhances that, it'll be another game.
Moving "the database" to transformative... not sure what you're pointing at there. I guess creativity fed from statistics and past works is more a miss than a hit ~ as i said... up to YOUR creativity. Of course i can imagine a combination of different kinds of tools, streamlined for creating swift (mock-up) scenes, using precise and generative tools alongside, up to whatever quality. However i still assume, that you can't have arbitrary amounts of magic and precision at the same time - which doesn't contradict it possibly becoming a killer tool in the hands of a creative person.
Concerning creativity... we must be very careful, i've seen few different kinds of takes on that (noteworthy for me):
1. Human creativity isn't so huge either. It essentially does the same. (I call this romanticism, as humans are not actually static, even if their evolving isn't as great as in some of the fairy tales. On the other hand, there are some concepts for manipulation of people, like "framing" or "setting the stage", e.g. verbally, and humans will in average act slightly differently, similar to a prepped prompt. But that's still not touching the fundamental difference i intend to point at.)
2. Imagine supervised learning, which is done for all those systems for a part, imagine you're the training part of the ai, and you're asking the human supervisor all the time "Is this creative?"
3. The tools are not really creative by themselves. It all depends on whomever is using the tool. However they have the potential to help you break barriers and take repetitive routine work off your shoulders. That would be a game changer, already. (Note that different people may have different kinds of games in mind.)
4. (Independently: risks with tools that minify the amount of work ~ floods of still irrelevant and not really creative content, which looks as if it could be something, spam, manipulation. Question is if there will be balancing effects, like people not wanting to pay for something they get tired of, or detection of generated content working very well, for instance.)
We're closer (to some things) and very far (from some/other things) at the same time. Like with secret services, you never know what they have in the basement by now, but societally, we obviously can't rely on such untold secrets (neither scientifically).