- This topic has 28 replies, 14 voices, and was last updated 3 years, 2 months ago by
.
-
Posts
-
So would the actual model and render take place on their server or would your creation file be downloadable?
If you can DL them then could you remove the data by altering it in another program?
Could they make a format only be understood by Daz?
May be the beginning of the end
The rendering is still done on your own hardware. What you get back from the Tafi server is just a scene file (DUF), which is essentially just a list of the objects and parameter settings--the coordinates to position each object, which texture file to apply to which surface with which shader parameters, and so on. Any necessary objects that you don't already own would likely be downloaded at the same time.
The scene file is in the same Daz format (DUF) that they use in DS4, and has no information about your creation prompt in it; that prompt text is only needed by the AI to generate the scene file. The scene file is just an ordinary old Daz file like you'd get if you created a new scene yourself and saved it. There's no AI-related data to "remove" from it.
The text-to-3D AI coming in Daz Studio 5 is very different from the generative AI that you see in things like Stable Diffusion, etc. It does not generate images at all, and it does not render anything.
The sole purpose of the AI in DS5 is to generate scene files for you, which you would render in Daz Studio on your own computer. The "intelligence" that it adds is the ability to find suitable models for your scene, to save you the trouble of having to search through 100,000 products in the Daz Store to find one that meets your needs.
It is basically an improvement on the Smart Content concept that was introduced with DS4, and instead of being limited to the products you already own, it will consider products from the entire store. It selects the objects for you, and creates the scene file that Daz Studio can use to edit and render that scene. The parameters for the objects and morphs will be preset for you in this scene, to fit the prompt you provided, but of course you can edit things as usual to refine the result.
It could be a real time-saver when it comes to creating rapid prototypes, providing you with a scene that is "mostly" what you want right away, so you can build upon it from there and finish it by hand.
Remember, this could (theoretically) be applied to any parameter setting, not just object morphs. It could eventually be used, for example, to specify volumetric lighting effects ("the headlights of an oncoming car can be seen through a blanket of dense fog"), so the parameters for camera positions and angles, light sources and intensities, etc. would be preset for you in the scene file that you receive.
@duckface wrote:
Than copy over the content from that server like they did in the beginning with the activation server from Windows when they introduced online activation. Not saying it’s easy to copy the content from that DS5 AI server but with some hacking tricks it’s possible.
No, the AI code and its database are never exposed to the user. You provide your text prompt to DS5, DS5 sends that text prompt to the Tafi server, and then it's a black box that just sends DS5 back a Daz scene file (DUF) in response. It may also trigger the download of any store assets your scene needs, if you don't already own them.
So no, there's no way to get access to (or download) the AI code or its database, and it's not immediately clear to me how that would be of any use to someone anyway. The whole point of having the AI done at the server end is for speed and efficiency--that AI code and its database are running on clusters of servers stacked with RTX A5000 and A6000 cards, running training processes at least once a day as the store adds new content to its database. That would take weeks to train on a typical computer, and by the time it finished its database would be weeks out of date.
There has to be a way to select from different assets or this won't work. If Tafi sends a scene with a cowboy hat that the user doesn't want, then they're not going to pay for it. And there are dozens of cowboy hat assets to chose from. So sending a completed .duf would be kinda stupid if the buyer/subscriber hasn't already approved the assets of the scene. Must be some way to interact with the interface so that Tafi presents all the options for the hat and the user selects the one that they want. Then it would make sense to build a scene and send it.
I'd assume that you'll get some sort of preview render rather than a direct download so you can alter your prompt until you get the desired results.
Assuming it isn't all pie-in-the-sky vapourware of course.
@ulysses, the user has paid a flat fee to subscribe to the AI features, so there's no additional charge for the assets. He's not paying for the cowboy hat, he's paying for access to every asset in the store. If he doesn't like the result he gets, he can try a slightly different prompt, and it doesn't cost him anything more.
Bear in mind that the prompt itself can specify assets, if you have a specific one in mind. Like, "male cowboy with Genesis 8 Male body, Dean 8 face, and hat from sku 12345".
As with all the current generative AI engines out there, getting the result you want comes down to crafting your text prompt properly and refining it one iteration at a time. Fortunately because this particular AI engine is not rendering images it should be able to give you a response in seconds rather than minutes. (The scene file it sends you is just a relatively small text file in DUF format.)
I personally would never use it because its easier to create with more control via sliders. BUT, one way this new thing could be handy is if you have a general idea in mind, it could save time to see roughly what it would look like quickly. Right now just to load a character or base can take minutes, then more to load each other asset, etc just to see an idea. With this you could quickly get an idea what it would look like, and THEN if its what you like take the time to load it all in locally and do it up
Yeah, the whole point of the AI is just to get you 80% there quickly, so you can finish things manually. No matter how carefully you word your text prompt it's not going to give you a result that perfectly matches your imagination, but it might get you 80% there, and that becomes your starting point. It's not about giving you a finished product, it's about saving you a lot of scaffolding work so you can focus on refinement.
Others might use it for creative inspiration, rather than limiting themselves to the small set of assets they use all the time. You might discover that a different cowboy hat out there works really well for the scene you had in mind, but you'd never have thought of trying it (or didn't know it existed). Loosening up your text prompts gives the AI more leeway to select assets, so if want to see a broader range of what's available just be more general in your prompt (e.g. "cowboy hat" rather than "cowboy hat with curled brim and gold hat band").
@Queensknight
I am surprised by your knowledge of AI DS5, are you a Taffi insider ? (just a question, not an accusation, no problem with that). If I understand the process you give instructions (prompts as in SD) to the AI to prepare your scene, you receive a duf file and than you render on your computer. I am afraid it will practically not work : you will have to download on the fly all the missing stuff the IA has selected for you and DAZ servers are notably slow and unreliable... GigaB of data will have to be downloaded in real time by hundreds of customers who will relaunch a new prompt because they are not satisfied with the current render... I see horrible path problems and version issues, and I remember DS is not multiprocessor compliant... What happens when you are out of RAM, when you have the same object referenced in two different paths (on your own computer and on the servers) further how can you use non DAZ assets (notably NSFW thingies) ? More trouble than solutions I guess but I am quite willing to be educated.No, I have no official connection to Tafi (or Daz), apart from using the DAZ Studio SDK for a few projects. I'm just basing my thoughts here on the limited press releases, and making some reasonable inferences about how the stated features are likely to be implemented, based on those clues.
As for the server bottleneck problem you raise, I would expect Daz to have to change to a CDN anyway, likely contracting with an established cloud provider like AWS, Google, etc. With or without the AI aspect, offering subscribers access to their entire catalog requires a more distributed and robust fulfillment network to avoid the very issues you raise.
Local caching of assets would also speed things up, particularly with minor prompt refinements, where most of the assets don't need to change between iterations.
Assets that you already have installed locally would not need to be downloaded either, so if you have storage space for a large cache of assets you could presumably speed things up even further.
As for non-Daz assets, those would not be available to the AI, presumably. You could use them in your scenes as usual (manually), but the AI would know nothing about them as they would not be part of its database. Adding metadata to third-party assets would not help in this case, unless Daz themselves choose to let those assets get included in the AI's training set. It would not be something the user has any control over, since the AI's database lives on Daz's servers.
Basically, PAs who want their products to be available to the AI engine would have to join the Daz marketplace and sign a consent contract for those items.
Yeah, I'm not seeing it. The results look too generic, too pre-baked. Like 400 other people will get the same looking results. I also doubt that you can save a specific look for later, at least not without paying $$$.
The biggest fail is that it smells like an attempt to make a subscription service out of the whole shebang.
A minor nitpick is that, once I create specific characters in DS (and before that Poser), they're mine, unless I choose to share them. They belong to no one else, and I own full rights. I build stories around them. I give them personalities. I know I'm not most users, but I like to tweak my playthings' meshes in Blender or ZBrush (depends on my mood and how much time I have). I doubt this thing will let me do that.
Finally, because it most likely uses LLM for it's AI base, I'm not going to get anything unique to me, no matter how hard I tweak the prompts. Instead, I'm going to get what it was 'trained' to base its results on, which means four things that immediately come to mind: shitty porn, waifu/manga bullshit, done-to-death tropes (scifi, medieval fantasy bullshit, and bodice-rippers) and I guarantee she'll be dressed like a slut.
No fucking thanks.
I'm pretty sure the AI part of this will never be on Daz Studio 5 itself, and will be a separate frontend/backend with a subscription. The rest of what @queensknight said has 95% chance of being exactly what they are doing.
USD will open up new revenue streams, more than the AI frontend will, that's just a gimmick.
If they release DS5 with USD export real players will actually look at Daz, which has not really been the case past 20 years. However, they are competing with Metahuman and Unity, so they better make some extremely fast upgrades to soft body physics, texture map options, and animation rigging, else it will be just a pretty pinup generator like it is now, and people will still export and re-rig everything.Also, the Omniverse Bridge is for the Enterprise version, not the free version like what Reallusion uses.
One issue about that is, it’s a $9k per year subscription for the Enterprise version! There’s no way DAZ would able to have their users use it for free, unless they go subscription (which is a huge concern), or NVIDIA waves the fee for DAZ Studio 5 users.
- You must be logged in to reply to this topic.
