kleio

An intuitive interface for exploring generative imagery

Team : Fergus Laidlaw, Jon Evans

Current image generation controls are unintuitive
Current image generation methods rely on parameter adjustment and text prompts as inputs. These prompts and parameters can be tweaked and iterated upon, but this workflow is punctuated by generation times and doesn't create a fluid workflow.
Kleio makes content creation for design, art and culture intuitive and truly accessible to a new population of creators by giving real-time control over certain inputs. Similar to a mixing desk in the music industry, Kleio allows you to combine different inspirations and fine-tune the importance of each to create your own unique imagery in a natural way.
Image generation is a powerful tool for combinative creativity
Current image generation controls are unintuitive
Current image generation methods rely on parameter adjustment and text prompts as inputs. These prompts and parameters can be tweaked and iterated upon, but this workflow is punctuated by generation times and doesn't create a fluid workflow.
Image generation is a powerful tool for combinative creativity
Kleio makes content creation for design, art and culture intuitive and truly accessible to a new population of creators by giving real-time control over certain inputs. Similar to a mixing desk in the music industry, Kleio allows you to combine different inspirations and fine-tune the importance of each to create your own unique imagery in a natural way.

This project aims to maximise the creative potential of diffusion models used for image generation. We've designed an intuitive platform for exploring the latent space within these models - the space between discrete concepts, that humans struggle to pin down, but artificial intelligence treats just like any other. By mapping the paths between concepts, it helps give us an understanding of the logic that is embedded in the model, which can both inspire our own creation and educate us about the function of a tool that is becoming more and more ubiquitous in our lives.

You can explore kleio yourself at https://www.kleioapp.com/

How it works

*This pipeline was created in January 2023 when image generation was slower than now and pre-caching the image set was necessary for a responsive interaction, with the expectation that as image generation got faster, the generations could be done in real-time without the need for pre-caching.

Design Process

Prototype 1

Below is one of my first manual attempts to map the space between seemingly unrelated or adjacent concepts. By incrementally adjust prompt weightings that are input into a stable diffusion model I created a canvas that can be wandered across with cursor and naturally explored. This particular canvas juxtaposes flowers and architecture on the x-axis and plastic and sunset on the y-axis, allowing the user to explore these visual concepts as they correlate to spatial dimensions. This prototype was made in processing, manually pre-generating the images and displaying them according to the mouse position within the 5x5 grid of the canvas.

Prototype 2

This second experiment adds a third dimension to the latent space. Affording each characteristic a physical slider on a midi keyboard allows the user to explore those characteristics easily in realtime, understanding how each has an effect on the other. Music is a natural analogy for this kind of interaction where the user needs control over a multi-dimensional space in real time (as with volume, bass, treble etc.) so a midi controller is a good first step, but I'm looking forward to developing the most natural physical interface for this specific interaction.

To save manually generating sets of images, we created a pipeline with gradio that iteratively weighted the embeddings of each prompt according to the parameters we gave. It correspondingly labelled the images with their weightings, so we could recall those images in touchdesigner according to the position of sliders (each physical or within the touchdesigner user interface). This pipeline also enabled us to explore what the optimal range of weightings was for compelling imagery, alongside the optimal resolution for a seamless user experience.

prototype 2 - backend
prototype 2 - front end

User Feedback

At each stage of the design process we took feedback from creative professionals, including a focus group of 10 architects that helped inform the user experience of Kleio. The video below shows one of our first user feedback sessions with the focus group, identifying kleio as a replacement for softwares like pinterest during the design discovery phase of an architectural project.

Other projects: