Launch unblocked,
£500k funding raised
The client's product turns a story into a sequence of generated images. Initial versions of the images were inconsistent with user outfits, scenes and even gender changing between shots. This made the product unshippable. To fix the problem we built an evaluation rubric that scored images according to consistency and story alignment. We then developed a novel strategy that combined reference images with clear story plans.
The character wouldn't stay the same person
Generating one good image is easy. Generating a sequence that holds together is an open problem in generative story telling. Across a single story certain aspects can (and should) change. These include what the user is wearing and what room they are in. Other attributes though, such as who the user is, should remain consistent.
As images were created in parallel context was not between shots. Individually an image may have been good, but when incorporated into the story it presented something very different.
The usual fix, feeding earlier images back in as a reference, was not tenable due to the latency of generating images in serial. We therefore needed to develop a novel strategy that presented both the changes in the story and the consistency of the character.