Conditional diffusion models are known for generating high-quality images that closely match a text prompt. Yet the authors note a practical limitation: when sampling with different random seeds, the outputs tend to look alike, which restricts creative use cases where variety is desired.
The paper proposes a technique based on "pullback directions" to make sampling more controllable and diverse. Rather than relying on random noise alone, the method steers the generation process along directions that are designed to produce distinct but still prompt-relevant variations.
This work addresses a gap between quality and diversity in generative models. By offering a way to explicitly control how much samples differ, it could give artists and designers more flexibility when exploring visual ideas from a single prompt.