As part of a Solarflare project, we needed to visualise an AI, a large language model, “thinking” when users asked it a question.
I came up with a fun point cloud simulation developed using cables.gl and an image to instanced points workflow.
The location of the points could be manipulated using a wide range of image filters, such as noise and transformations to move the particles around. A Feedback loop added trails and some post-processing. The visuals reacted to users entering a question, making it interactive.
As an additional fun challenge, I decided to try to render a mock-up of the visuals running on a large transparent screen in a gallery with some people watching the screen.
This turned out to take almost as long as creating the AI visuals. I spent a while learning how to make the scene look as photo-realistic as possible. Most of the lighting comes from the on-screen visuals. The people were animated meshes imported from Mixamo with standing idle loops. The camera path was fine-tuned with a subtle noise to make it feel handheld. Post-processing added CCD low-light grain, vignetting, and chromatic aberration.
The overall effect wasn’t bad but still didn’t quite cross the uncanny valley and appear completely real. Still more to learn.