Generative models, such as text-to-image models, excel at taking textual prompts and producing semantically faithful outputs. However, the user’s ability to control the level of creativity of the output remains limited. We introduce Concept-Level Temperature (Concept-τ ), a novel framework for controlling creativity in generative models.

Unlike existing methods that operate on low-level features (tokens, noise, etc.), Concept-τ operates at a higher level of abstraction; it systematically identifies decisions left underspecified by the prompt and realizes them according to a creativity parameter, creating a structured exploration space.

In this work, we design general criteria for a creativity parameter, and implement an interpretable, model-agnostic method to satisfy them. We demonstrate our ideas on the text-to-image domain; empirical evaluation, both qualitative and quantitative, shows that Concept- significantly outperforms baseline methods in creative modulation and enhancement.