Imogen and Jon approached me a with an unusual brief – to create a looped 30-second visualiser from an image of the two of them. They didn't want to use any off-the-shelf generative AI extensions, and it was important that the source image stayed unadulterated.

I thought the best approach would be 3D Gaussian Splatting (3DGS) – a technique for capturing and recreating three-dimensional space using millions of tiny soft points rather than solid surfaces.

Gaussian splats normally require multiple views of the same object, or a video moving through the scene – neither of which we had.

I first fed the 2D image into a machine-learning model that predicts depth and builds a realistic, if imperfect, 3D representation directly from a single photograph. I imported that scene into 3D software and built a camera move through it, then layered in particle animation, bespoke lighting, glow, noise.

The results looked too similar to visuals Imogen and Jon had done before, so I took an alternative route – generating multiple 360° views from the 2D image, effectively forcing an algorithm normally used on simple isolated objects to ‘hallucinate’.

I then fed those hallucinated views into a second algorithm to build the actual 3D splat. I repeated this process several times, uploaded everything into splat editing software, overlaid different 3D generations by adjusting scale, position and rotation, and created a looped camera movement through the combined cloud.

The final step was bringing more defined silhouettes of Imogen and Jon into the scene. I isolated the silhouettes on the 2D source image first and tweaked the hallucination algorithm to generate interpolated alpha layers, which ended up looking like translucent ghosts, then dropped these in and adjusted the camera movement slightly to keep the silhouettes centred.