👉 Temperature, in the context of mathematical models and machine learning, particularly in generative models like text-to-image or text-to-video, refers to a parameter that controls the creativity or randomness of the output. It essentially scales the logits (the raw, unnormalized scores predicted by a model) before applying a softmax function. A lower temperature results in more conservative, deterministic outputs, favoring outputs that are closer to the most probable class or topic, leading to more focused and less varied results. Conversely, a higher temperature increases randomness, making the model more likely to explore less probable but potentially creative or diverse outputs, resulting in more varied and imaginative results. This parameter is crucial for balancing between fidelity to the input data and novelty in the generated outputs.
