Loading...
> >

AI generated textures in products

2»

Comments

  • WendyLuvsCatzWendyLuvsCatz Posts: 38,272

    Stable Diffusion and other image based AI is a indeed mathematically generated procedural function, the thing many have an issue with is the source of data used to train it, not the algorithm itself 

    it is the reason I myself would never use it commercially but I have utterly no issues with noncommercial use which is something many who happily copy other people's intellectual property albeit it skillfully,  seem to find questionable still.

    If it's used to make existing images seamless be it you own photographs or your own artwork which it can do, I honestly don't see that as a problem even for commercial use, setting noise scale very low and enabling tiling works pretty well and by its very parameter value uses the original image, about 2 works well.

    It all comes down to whether you are using other people's work or you own, being trained is not the same as outright plagiarism but one certainly can prompt for others work and depending on the prompts and settings use get things that obviously use others intellectual property just like you can skillfully eye off your favourite comic or game character and sketch it!

    ... or 3D model it in Zbrush

  • TotteTotte Posts: 14,004

    RobotHeadArt said:

    Richard Haseltine said:

    RobotHeadArt said:

    RawArt said:

    Just keep in mind that procedural textures that are generated by programs like substance painter and Filter Forge and the like are not AI textures, they are all math and could be easily confused

    Generative AI is also math so this statement doesn't make sense.  It is a different set of algorithms than those used in generative/procedural generation but at the end of the day a computer can only do math, logic operators, and branching. https://theaisummer.com/diffusion-models/

    With generative AI tools in phones and cameras which capture photos that can be used for textures, Substance Designer, and Photoshop it is going to become impossible over time to discern how much of an image or texture has been generated, modified, or tweaked with AI tools.  Since PAs also use textures from merchant resources, public domain sources, etc. they may not even know the full provenance of all pixels in the image resources they use in a product.

    Generative AI just takes a prompt of some kind, procedural textures require the deliberate construction of a sequence of functions (usually in the form of formual nodes of some kind) to produce the end result.

    This is not exactly accurate.  Generative AI algorithms like Stable Diffusion do not operate on prompts as computers do not have the ability to understand language, they can only perform math, logical operations, and branching based on conditions.  While it may look like Stable Diffusion is generating based on a prompt, what it is happening is there is a separate step before the Stable Diffusion algorithm, where a prompt is converted by a language model like CLIP into embeddings which are just vectors or just a series of numbers.  Stable Diffusion algorithm operates on these embedding vectors, it never sees a prompt.  Also, you don't even need a prompt for generative AI functions.  You can easily use image2image, passing in just an image which is converted into pixels or again, just numbers representing the colors, without a prompt and get a new image.  Just like you can use Substance Designer or Filter Forge to apply an algorithm to an input image, you can use generative AI just on just an input image without any prompt.  The only difference between the two is in the specific algorithms they implement.  At the end of the day, it's just math. Numbers input, numbers output. https://stable-diffusion-art.com/how-stable-diffusion-work/#Conditioning goes into more detail on how text is converted into the embeddings.

    There is one major difference between procedural generated textures and AI generated textures. 
    Procedural textures are created by manually setting parameters, connecting different operators, filters and other processing functions to generate a texture.
    AI generated textures are generated from training datasets of real images where, simply put, the AI learns how a certain thing looks by analyzing thousands of images and then replicate that upon request, mixing different learned images to create different kind of textures.
     

     

  • Richard HaseltineRichard Haseltine Posts: 101,238

    Totte said:

    RobotHeadArt said:

    Richard Haseltine said:

    RobotHeadArt said:

    RawArt said:

    Just keep in mind that procedural textures that are generated by programs like substance painter and Filter Forge and the like are not AI textures, they are all math and could be easily confused

    Generative AI is also math so this statement doesn't make sense.  It is a different set of algorithms than those used in generative/procedural generation but at the end of the day a computer can only do math, logic operators, and branching. https://theaisummer.com/diffusion-models/

    With generative AI tools in phones and cameras which capture photos that can be used for textures, Substance Designer, and Photoshop it is going to become impossible over time to discern how much of an image or texture has been generated, modified, or tweaked with AI tools.  Since PAs also use textures from merchant resources, public domain sources, etc. they may not even know the full provenance of all pixels in the image resources they use in a product.

    Generative AI just takes a prompt of some kind, procedural textures require the deliberate construction of a sequence of functions (usually in the form of formual nodes of some kind) to produce the end result.

    This is not exactly accurate.  Generative AI algorithms like Stable Diffusion do not operate on prompts as computers do not have the ability to understand language, they can only perform math, logical operations, and branching based on conditions.  While it may look like Stable Diffusion is generating based on a prompt, what it is happening is there is a separate step before the Stable Diffusion algorithm, where a prompt is converted by a language model like CLIP into embeddings which are just vectors or just a series of numbers.  Stable Diffusion algorithm operates on these embedding vectors, it never sees a prompt.  Also, you don't even need a prompt for generative AI functions.  You can easily use image2image, passing in just an image which is converted into pixels or again, just numbers representing the colors, without a prompt and get a new image.  Just like you can use Substance Designer or Filter Forge to apply an algorithm to an input image, you can use generative AI just on just an input image without any prompt.  The only difference between the two is in the specific algorithms they implement.  At the end of the day, it's just math. Numbers input, numbers output. https://stable-diffusion-art.com/how-stable-diffusion-work/#Conditioning goes into more detail on how text is converted into the embeddings.

    There is one major difference between procedural generated textures and AI generated textures. 
    Procedural textures are created by manually setting parameters, connecting different operators, filters and other processing functions to generate a texture.
    AI generated textures are generated from training datasets of real images where, simply put, the AI learns how a certain thing looks by analyzing thousands of images and then replicate that upon request, mixing different learned images to create different kind of textures.
     

    Both of which are what I was trying to convey with the shorthand "prompt"

  • JasmineSkunkJasmineSkunk Posts: 1,906
    edited September 1
    It's a really good question. I would also ask, what constitutes as "AI"? What about filter created textures that use noise or other algorithms to generate an image? What about fractals? Are these things that have been used almost forever now considered AI? Or does this refer only to text/prompt generated imagery? I'm asking what is Daz's stance rather than the community's personal interpretation of these things. Have they made any distinctions vendors follow?
    Post edited by JasmineSkunk on
  • TotteTotte Posts: 14,004

    JasmineSkunk said:

    It's a really good question. I would also ask, what constitutes as "AI"? What about filter created textures that use noise or other algorithms to generate an image? What about fractals? Are these things that have been used almost forever now considered AI? Or does this refer only to text/prompt generated imagery? I'm asking what is Daz's stance rather than the community's personal interpretation of these things. Have they made any distinctions vendors follow?

    My personal view is that to be considered AI, it should be created using a dataset and a Neural Engine, otherwise it is generated, 
     

  • FantastArtFantastArt Posts: 313
    edited September 1

    What I was talking about is not a pattern like texture, but maybe an image of a dragon, animal, character, used for an image in the product or maybe a print on G9 Shirt. As I said, I don't have a big problem with it, but it would be good to know, because there are now people who are obsessively searching for hidden AI using questionable programs. Then, for example, it would be shown in one of my DAZ images and my image would be shown as AI, even though it isn't. That would annoy me.

    Post edited by FantastArt on
  • TorquinoxTorquinox Posts: 3,359

    That's why it's all becoming a game of whack-a-mole. People will think or suspect, blah-blah, but there's often no way to know what the deal is in something like a texture. The ai detectors are looking for ai-like noise patterns in the image, but they'll pop positive for jpeg compression noise. So say one did use something ai-generated in a texture. The person making the texture will process the image for use. That may mean applying the texture to a 3D model for baking into a UV-specific texture mask. Then the customer will apply the texture to the model in DS (true for many people here) with posing, lighting, etc. then the whole thing is rendered and maybe post processed to create the final image. Ok. How does anyone actually prove the texture came from or included output from an ai image generator? And it's not going to get easier going forward.

  • TorquinoxTorquinox Posts: 3,359

    To be clear, I am still strongly opposed to ai-generated imagery due to the way in which the training data was harvested without consent or compensation to the creators of the original works used to train the ai generators, myself included. That hasn't changed.


    Even so, it seems perfectly obvious to me that saying "don't do that" Is not going to stop some people from doing it. And it's already difficult to prove if someone is doing it. So, for the most part, rules against using ai generated images in textures are honor system and rules intended to make people feel better. Sorry. That's the way I see it.

  • RawArtRawArt Posts: 5,907

    ..and there is the rub.

    The time may come when there will be no choice but to accept ai, not because we want to, but because it will be so deep into society that it will just be there. If we like it or not, it is not going to go away

     

  • TorquinoxTorquinox Posts: 3,359

    @RawArt : Seems that way

Sign In or Register to comment.