For a client's brand timeline, The style needed to match the moodboard given, as well as some objects that were too obscure for the AI to know them. Using bleeding-edge techniques, I was able to create a workflow that allowed me to match the given style to overdraw images of obscure objects or generate images from scratch.

Back in early 2023, it was still a challenge to get AI to follow the schematics of an already given image. ControlNet was just released and nobody really knew or understood it all too well, so when I was asked to generate abstracted artwork reminiscent of the Brandenburg Gate, I took that chance to test out the various different ControlNet approaches.

When a new image generation model was released by a German startup in August 2024, I immediately jumped on the experimental support for it in ComfyUI to see whether it is useful for current or even future projects. Realizing that the model was a huge step up in prompt comprehension as well as generating text, I tried out some diverse range of possible applications to see in which use cases it would be of help - and it fared incredibly well.

Even in the Disco Diffusion days (where you needed an A100 graphics card and about 30 minutes to generate a single picture, those were the times), people were already trying out to generate some form of videos with AI. Come Stable Diffusion which made generating a lot more resource friendly, I tried to do something similarly to how video generation worked then: A video in which the camera moves slightly while the prompt changes every few seconds to make the sense of travelling across the world.

Of course the result isn't - contrary to the name of the model - stable, as it creates a new image for every single frame, so keep in mind this was before any of the fancy video capabilities seen in the last years.