Infinite Wonderland
Today I’ve been looking at a new project from Google Labs called Infinite Wonderland. The project invited artists to train a model that can create images in the artist’s distinctive style, and then to connect that image generation to the text of Lewis Carroll’s Alice’s Adventures in Wonderland. This essentially creates a generative illustrated version of the novel in the artist’s own style.
This operates on Google’s Imagen 2 image generation model and uses a new tool called StyleDrop. The biggest difference between StyleDrop and other image generation tools is that you can adjust the style of generated images directly without needing to retrain the underlying model. Somewhere in the video they made, StyleDrop is described as an interface that allows the artists to make aesthetic modifications in real time.
So it’s a fabulous and powerful new interface that allows artists to more easily train a model on their own styles/datasets, which is amazing because even though it’s become common to approach artistic production through the creation of artist datasets, it hasn’t yet been super easy.
But one of the things I’m really psyched about with this project is the connection to the text part of it - the way they’ve entwined these aesthetic experimentation with the novel itself. You can highlight each of Alice in Wonderland’s 1600 sentences (does that seem like a lot? Intuitively it strikes me as a small number but that might be just me) and the project will create an image (in whatever style) that illustrates the text.
For those of you who know me, nothing could ever be more pertinant than a generative novel, or any kind of project in general that creates narrative through recombinant text and image systems. Like my projects Space Video (2012) (that addresses ideas of inner and outer space exploration through a generative system that mixes an original non-linear narrative with YouTube videos uploaded in real time), Why Some Dolls Are Bad (2007) (a dynamically generated graphic novel built on the now-defunct public Facebook API that engages themes of ethics, fashion, and artifice in order to re-examine contagion and environmental decay) and Grafik Dynamo (2005) (that loads images from LiveJournal and Flickr into a live action comic format).
Infinite Wonderland of course is predicated on demonstrating the creative potential of StyleDrop by using it in connection with a familiar book, but I’m psyched to think about what this might open up when working with other kinds of texts. Fragments, poetry, other original novels and stories, but also technical manuals, legal documents, almanacs, or any other bananas document you can think of broken down into segments and matched with an aesthetic approach. For that matter scripts - into a storyboard or stop-motion animation, film or whatever we don’t have words for yet. This connects to the idea of Lean Narrative I wrote about awhile ago but builds on it… and so many other layers of potential in terms of how Gemini changes the reference text into an image description that becomes a prompt.
