The Silent Prompt: Initial Noise as Implicit Guidance for Goal-Driven Image Generation

Text-to-image synthesis (T2I) has advanced remarkably with the emergence of large-scale diffusion models. In the conventional setup, the text prompt provides explicit, user-defined guidance, directing the generation process by denoising a randomly sampled Gaussian noise. In this work, we reveal that the often-overlooked noise itself encodes inherent generative tendencies, acting as a "silent prompt" that implicitly guides the output. This implicit guidance, embedded in the noise scheduler design of diffusion model formulations and their training stages, generalizes across a wide range of T2I models and backbones. Building on this insight, we introduce NoiseQuery, a novel strategy that selects optimal initial noise from a pre-built noise library to meet diverse user needs. Our approach not only enhances high-level semantic alignment with text prompts, but also allows for nuanced adjustments of low-level visual attributes, such as texture, sharpness, shape, and color, which are typically challenging to control through text alone. Extensive experiments across various models and target attributes demonstrate the strong performance and zero-shot transferability of our approach, requiring no additional optimization.