Latent Blending

January 2023

We are open-sourcing latent blending, a text-to-video method developed by our team. It allows users to achieve smooth video transitions between text prompts. This process is driven by the advanced Stable Diffusion XL algorithm and works by intricately blending intermediate latent representations and text conditioning. The result is a transition that is not only smooth but also highly customizable in high resolution.

At the heart of latent blending lies a unique approach to the perceptual aspects of video transitions. This method creates transitions so fluid and natural that they seem to challenge the very principles of nature. By avoiding motion during the transition, we reduce the visual system’s ability to detect changes, producing an effect known as change blindness. This results in the perception of a continuous, unaltered image.