ELROND: Exploring and decomposing intrinsic capabilities of diffusion models

Researchers introduced ELROND, a framework that disentangles semantic directions within diffusion models' input embedding space. By analyzing gradients from stochastic prompt variations, it isolates interpretable steering directions for fine-grained control over single concepts. The method also mitigates mode collapse in distilled models and establishes a novel estimator for concept complexity. This approach moves beyond analyzing output features to directly manipulate the underlying generative process for more predictable AI image creation.