> If I’ve spent 8 hours getting the perfect prompt, fine tuning a few LORAs, mixing them, choosing between 8 different checkpointed stable diffusion models, and have done a bunch of in painting, does this constitute a copyrightable work?
The 8 hours doesn’t matter. By the USCO’s ruling, iterative prompt refinement probably wouldn’t, but that’s not super clear (partially, because the USCO’s description of what the model is doing justifying its determination is a fundamentally strained metaphor, so while its clear how it applies to the exact case it describes, its not clear how it generalizes.)
If you are fine-tuning a model (LORA, Checkpoint, whatever), then your input isn’t just a prompt to the model (once or in an iterative process with review of the output), so, your pretty far outside of where the ruling provides clear guidance.
> and have done a bunch of in painting,
inpainting is probably the thing most (even though it again involves prompting, it involves specific selection of where within the image to apply that based on aesthetic concerns) similar to the traditional creative parts of visual art, and the strongest argument given the shape of the USCO description of its rule.
But, I have a feeling that that a rule that appeals to tradition and vague analogy to lower tech visual techniques probably won’t be anywhere close to the final word on copyrightability in this space.
IANAL but I'd assume "inpainting" is legally considered equivalent to arranging. You're not pushing pixels in Photoshop, you're just running the AI again on a particular part of the output, thus combining that output with previous output. This seems equivalent to bricolage at best.
The 8 hours doesn’t matter. By the USCO’s ruling, iterative prompt refinement probably wouldn’t, but that’s not super clear (partially, because the USCO’s description of what the model is doing justifying its determination is a fundamentally strained metaphor, so while its clear how it applies to the exact case it describes, its not clear how it generalizes.)
If you are fine-tuning a model (LORA, Checkpoint, whatever), then your input isn’t just a prompt to the model (once or in an iterative process with review of the output), so, your pretty far outside of where the ruling provides clear guidance.
> and have done a bunch of in painting,
inpainting is probably the thing most (even though it again involves prompting, it involves specific selection of where within the image to apply that based on aesthetic concerns) similar to the traditional creative parts of visual art, and the strongest argument given the shape of the USCO description of its rule.
But, I have a feeling that that a rule that appeals to tradition and vague analogy to lower tech visual techniques probably won’t be anywhere close to the final word on copyrightability in this space.