ai training data
WendyLuvsCatz
Posts: 41,685
https://www.daz3d.com/ai-training-data
sooo.. what is this???
not really sure which forum to ask in so moving it to the commons as not technically DAZ ai Studio
Post edited by WendyLuvsCatz on

Comments
I'm guessing it's intended for creating 'loras', effectively adding data to an AI image generator so that it can produce repeatable characters or other elements, usually triggered by some keyword that directs it so use data from the selected lora.
I know what a LoRa is, I train them myself on family and pets
but the EULA explicitly forbids this using DAZ content for redistribution
II have used some renders to train one for my own use but mostly go the other way dialing up DAZ characters based on my AI ones and also using the AI characters in Face Transfer
I think they would have more luck selling bikinis to Eskimos, but who knows ( not having a go at you Wendy, just based on how AI has been generaly regarded by folks here ).
https://www.daz3d.com/eula
this
as I said, I have used my own 2D renders, customised to the way I wanted my character to look to train one, and only one LoRa ever, I hardly have touched it though because of the ambiguity of the EULA regarding 2D renders for one's own use and training, preferring to just modify by inpainting and animate the DAZ renders of that character using AI instead, I don't train LoRas on DAZ characters ie Victoria , Michael, PA characters etc
I wonder what might be the source of those "millions" of 3D assets the page says will be available for training AI.
I can understand if they forbade using DAZ images to train an AI engine, because that would basically involve giving away their IP to help another company make a profit, so yes it's confusing. It's really up to DAZ to make very clear what is allowed, since users are not Intellectual Property lawyers.
what I too would like to know
mine is a custom character too, loosely based on G2F Bijou with other features dialed in, who is whom I use for my DAZ renders but at the same time not, it was actually trained on AI altered images of her to look more realistic, human, not raw renders
her renders used in Canny in controlNet Fooocus SDXL1.5 to be specific to create my training images
I still thought that was too derivative according to the EULA so don't use that LoRa
I wouldn't share that LoRa though, I would like clarification for my own use in images and videos though from DAZ
the only video I did using that LoRa
Well that was dammned scary! That image at about 2 seconds in is the spitting image of a lovely lady I used to know many decades ago. It's so close it could be a photo of her, expression, eye colour, hair an all. I definitely felt a chill seeing that unexpectedly. Her name was Jean, haven't thought about for years until today.
I train mine locally with One Trainer
as I said though, mostly on real photos not renders, I want real people
I have uploaded renders and AI enhaced renders to sites to animate them, nothing more
but do that too on my own PC using Pinokio to run Wan, LTX and other AI animation models
she definitely is not a 3D render, I honestly never saw the point of training on one
The person I'm thinking of had two children and was engaged to be married ( her first husband having sadly been killed in a climbing accident ).I really hope she had a happy life, or hopefully is still enjoying one.. We were just friends, but she was the sort of person that made you smile as soon as you saw her.
When I saw that link I assumed that DAZ was aiming at large companies who could train there AI models using DAZ assets (either renders or the models themselves). Note that there is no price quoted, just a "contact us", so these are definitely not intended for the vast majority of the customers here.
The advantage of training on 3D models is you can supply lots of key words with the images, that in theory would make it easier to train, and there would also be no backgrounds to filter out first in order to understand the look of the main subject.
It's certainly only the Daz-owned assets, I'm not sure but I think the 3D content rather than renders. PAs are not included (unless, possibly, they opt in) and certainly end-user art is not included.
I can ask for clarification, but I would think that covers sharing the training results (as the license forbids sharing derivatives of the 3D models) rather than using AI to train for your own image production..
Please ask for clarification, @Richard Haseltine
I have not used Daz Studio renders in that way, but will be good to know for future plans.
It' seems straight forward.....You can't do it with the characters you've purchased due to the license .... but they can sell you their data if you want to use it with AI in a separate license
This comes with a massive dose of "I am not a lawyer", but Daz has the option to sell content under different licences.
We're already familar with this with the Interactive and 3D print licences, which provide permissions that allow you to do things the standard licence disallows.
They can, of course, provide an "AI licence" which does grant this permission.
Again, this has a massive "I am not a lawyer" disclaimer, but the standard licence has granted the permission use use the assets for the purpose of "the creation and presentation of two-dimensional animations and renderings" and with the restriction that the "User may not in any case publish, market, distribute, transfer, sell or sublicense any renderings, animations, software applications, data or any other product from which any Content, or any part thereof, or any substantially similar version of the Content can be separately exported, extracted or de-compiled into any re-distributable form or format".
Even without the AI clause being added later on, I think it would be a stretch to claim that this covered you to distribute a LORA; As AI advocates repeatedly like to tell us, models do not actually contain the original dataset (so it's not reasonably argued that the LORA is a rendered image and subject to the specified freedom of use), and there's also a strong argument that the model is highly imitatitive of the original content and ineligible on those grounds.
There's always been these back-up clauses of a sort in the licence, because there are cases that would, quite reasonably, count as taking the mickey.
For some examples, you probably culdn't just point a camera squarely at a flat plane with a texture on it, render out the image, then claim you have full commercial rights to that texture. Similarly, I imagine that rendering out a model from 40 different angles and feeding that all back into photogrammetry software to get a new 3D mesh out of it would, even though it has been gone purely via your rendered images, not be something you could claim full rights to.
Once again - still not a lawyer, but while for many purposes you can think of yourself as being able to the rendered images mostly as you will, that is not actually the literal text of the licence, and I advise that you read the specific text to get a sense of what limitations might exist.
I don't see that, if a third-party used rendered image to train AI that would not be the responsibility of the user - as long as they didn't upload it to a host that granted others theright to use the hosted images for AI training, of course.
...all this makes my brain hurt. Think I'll just stick to doing the best I can the "old school" way. with the tools I have.
That's basically how I read the situation. None of the images or videos I've been generating are shared with or made availble to other users in the AI engines I use, they're "mine" exclusively. I mean, anyone could grab something I uploaded to the forums here and use *that*, but that's the same with anything anyone posts here, or even the store images for that matter.
For characters, I either use my Webboids models or I run any DAZ characters through my own custom Toon modifiers before using the modified Toons to generate stills and videos. The reason for the extra step of custom tooning my characters first is because I sometimes use the 2D characters to generate 3D models that I print; should I ever decide to share these 3D meshes, I'm comfortable knowing that nothing I convert to 3D retains any semblance of the original DAZ "shapes" and are free from any copyright issues.
You guys do know that you guys are calling a "LoRA" is already trained, for the specific model that it is meant to adjust, right? That using it with your reference images is not "training", but inference?
So:
Yes, the discussion about "LoRA" and generative AI as 2D images is a red herring, I believe. Richard Haseltine clued us in to where the real value lies, and most likely what Daz is trying to market:
The race in tech nowadays is to develop autonomous systems that, from the get-go, woefully lack any concept of the real world, the 3D one we all live our lives. Tech companies are busy building digital twins of the real world, and it's here that Daz's "3D data as training material" comes into use.
Cheers!
Thanks for the confirmation.
I am sometimes editing/improving my renders with AI.
Has not tried to use my renders as training material for AI, but maybe will try it at some point.
yes, thanks for this
I really don't use my renders for training (or whatever the correct term actually is) the software I use is actually called One Trainer BTW, for the simple reason I want realistic LoRas not ones that look like 3D renders
Renders are the target not the source
biut this is still good to know as an option if needed
Not what I expected so thanks for all getting to the bottom of this..... Booting up Ostris as we speak ;)
Thanks for the question. And for the clarification.