Discussion about this post

User's avatar
John Giacobello's avatar

I'm not sure why we have to spend so much time explaining and justifying using AI tools in art creation. Let's just experiment, play, and enjoy it. Whether the gatekeepers deem it "art" or "slop" is entirely up to them and not our concern.

David Yates's avatar

I like this conception, but I also think there's a spectrum between curation and creation that is not appreciated in the discourse about this subject, perhaps partly because of the UX of generators most people have encountered. The UX of ChatGPT Images is that you ask the AI to make you an image, maybe based on one or more images you supply, giving as much or as little detail as needed, and then in an ideal world you get out the image you wanted, or even something better than what you had in mind. You can ask the AI to make changes, either in the chat or by making comments on parts of the image, but it's nonetheless a pretty hands-off process.

There is a whole other world of stuff you can do if you have more control over the diffusion process with local models or a more expansive API than the one OpenAI provides. Setting the denoise level and targeting specific parts of the image you're working on lets you do everything from smoothing out lines to completely altering the style and/or content, and there are tonnes of other techniques like ControlNet (famously responsible for Spiral Town), regional prompting, edit models and so on. There's a plugin for the Krita painting program that I think represents the best implementation of this idea, and plenty of YouTube videos of people using it to visually shape the AI's outputs in real time.

I think even sufficiently detailed prompting probably reaches a point where it's more than curation. Ideogram recently released an open weight model that's trained on a specific JSON format for prompting which allows you to specify bounding boxes for elements of the generation, and you can get very granular with it and really control the composition. https://ideogram.ai/blog/ideogram-4.0/

I'm hesitant about advocating some kind of Labour Theory of Value for art where the amount of human effort required to make it determines the work's value, but I'm really interested how works produced with the tech can blur the line between curation and creation like this. Personally, I lean towards wanting to create and get the output as close to my initial vision as possible, but an undeniable part of the fun of AI image generation is when the model surprises you with an unexpected interpretation of your prompt that may be better than what you first had in mind.

13 more comments...

No posts

Ready for more?