The idea for this artpiece came to me as a shower thought. The geometric abstraction allows a much lower resolution image model to drive the 1024x1024 canvas, which means we can run the image model in the browser. What surprised me is that on most devices that can use the WebGPU engine to run the model, images render so quickly that instead of interpolating between images in the output space we can interpolate between noise latents in the input space.
Bears are generated by a rectified flow model reflowed and distilled down to a single step. The denoiser was then exported to ONNX to run in the browser with sampling offloaded to JavaScript. If your device runs the network fast enough, the image transition plays by slerp-ing between random latents; otherwise the image transition is a lerp in pixel-space. Since the latents are random, if you watch long enough you'll see some...imperfectly optimized bears.