This is great! I had been meaning to ask you what you meant by Gaussian noise loop. I work a lot with noise after image generation, or in some cases by prompting for lower image resolution - I wonder what other strategies people are using.
In my case it’s just a matter of asking, “what kinds of prompts would induce the system into making something it really shouldn’t make?” And so after exploring how these tools worked, and trying to figure out how to do with Diffusion what Brakhage did with “mothlight,” it occurred to me that I should try to type “Gaussian Noise” and hit enter. I did, and my theory was correct: it created something absolutely weird.
In addition to post-hoc manipulation and forcing low resolutions, we've been exploring prompts that would require examples unlikely to exist in the dataset (eg. embroidered monograms) with models that had not been optimized for type (stable diffusion). And 'conceptual prompts' (or that's the best i can describe it for now), like prompting for images of a generative model's dreams. More recently we've had some interesting results for prompts with the structure "X in the shape of Y" but are still exploring that!
This is great! I had been meaning to ask you what you meant by Gaussian noise loop. I work a lot with noise after image generation, or in some cases by prompting for lower image resolution - I wonder what other strategies people are using.
In my case it’s just a matter of asking, “what kinds of prompts would induce the system into making something it really shouldn’t make?” And so after exploring how these tools worked, and trying to figure out how to do with Diffusion what Brakhage did with “mothlight,” it occurred to me that I should try to type “Gaussian Noise” and hit enter. I did, and my theory was correct: it created something absolutely weird.
In addition to post-hoc manipulation and forcing low resolutions, we've been exploring prompts that would require examples unlikely to exist in the dataset (eg. embroidered monograms) with models that had not been optimized for type (stable diffusion). And 'conceptual prompts' (or that's the best i can describe it for now), like prompting for images of a generative model's dreams. More recently we've had some interesting results for prompts with the structure "X in the shape of Y" but are still exploring that!
Love it! Are examples online anywhere?