Despite advances in text-to-image models, prompting does not always lead to desired outputs.
Controlling model behavior by steering intermediate activations is an alternative to reach concepts in latent space.
Experiments show that concept reachability in latent space has a distinct phase transition and can be impacted by where the intervention is performed.
Model providers can leverage steering to reach concepts reliably without costly retraining, dataset curation, and innovate with control mechanisms for users.