Joint Generation

A generation strategy in which multiple editing states are conditioned jointly on the same initial visual context, so no error is propagated between the states of different turns, analogous to denoising multiple related samples simultaneously in diffusion models.