3 ms·

And yet watermarks are a thing. I'm skeptical that no one will use this, especially if you can put it into a pipeline of multiple image transformations that "pr

And yet watermarks are a thing. I'm skeptical that no one will use this, especially if you can put it into a pipeline of multiple image transformations that "protect art". It could be bundled into tools or on platforms themselves as part of their value add. YouTube already transcodes your videos, why not DeviantArt encode your art except to verified human accounts or something? Their efficacy is the more interesting question to me.

dragonwriter 4y ago
> And yet watermarks are a thing. The full-resolution examples I’ve seen have shown this as much more destructive to the underlying art than usual watermarks, to the point where I’m not even convinced people would use it on samples when the underlying art wasn’t displayed online but sold through other channels. > YouTube already transcodes your videos, why not DeviantArt encode your art except to verified human accounts or something? So the idea is to be “protected” against automated scrapers building datasets for training, say, the core StabilityDiffusion models, but not against the huge number of individuals training (and remixing) checkpoints/loras/TIs/Aesthetic Gradients and exchanging them that are driving the ecosystem?
- cobertos 4y ago
  Maybe you have a different viewpoint of the ecosystem than me? Everyone I know that's used this is not an AI expert but still tech savvy. None of them have trained their own models. The path of least resistance for them would be to work around the "protected" art if they were doing something more custom For someone with more effort and time to dedicate to cloning a specific artist, maybe not as useful. But I know less of this type of use case/the tools in that part of the ecosystem
  - dragonwriter 4y ago
    > Maybe you have a different viewpoint of the ecosystem than me? Everyone I know that’s used this is not an AI expert but still tech savvy. None of them have trained their own models. Look at something like Civitai [0] and the number of users submitting things there: pretty much everything there (there’s a few things like poses and tools that aren’t) is one kind or another of trained model (though some are merges of existing models and extractions of one kind of model from another.) [0] https://civitai.com/ https://civitai.com/

dragonwriter 4y ago

> And yet watermarks are a thing. The full-resolution examples I’ve seen have shown this as much more destructive to the underlying art than usual watermarks, to the point where I’m not even convinced people would use it on samples when the underlying art wasn’t displayed online but sold through other channels. > YouTube already transcodes your videos, why not DeviantArt encode your art except to verified human accounts or something? So the idea is to be “protected” against automated scrapers building datasets for training, say, the core StabilityDiffusion models, but not against the huge number of individuals training (and remixing) checkpoints/loras/TIs/Aesthetic Gradients and exchanging them that are driving the ecosystem?

cobertos 4y ago
Maybe you have a different viewpoint of the ecosystem than me? Everyone I know that's used this is not an AI expert but still tech savvy. None of them have trained their own models. The path of least resistance for them would be to work around the "protected" art if they were doing something more custom For someone with more effort and time to dedicate to cloning a specific artist, maybe not as useful. But I know less of this type of use case/the tools in that part of the ecosystem
- dragonwriter 4y ago
  > Maybe you have a different viewpoint of the ecosystem than me? Everyone I know that’s used this is not an AI expert but still tech savvy. None of them have trained their own models. Look at something like Civitai [0] and the number of users submitting things there: pretty much everything there (there’s a few things like poses and tools that aren’t) is one kind or another of trained model (though some are merges of existing models and extractions of one kind of model from another.) [0] https://civitai.com/ https://civitai.com/