Diffusion models generate images by iteratively denoising a sample over many steps. As the abstract notes, current models condition every step on the same instruction, so the trajectory cannot be addressed at the level of individual operations.
Dancing Diffusion is introduced as a method that learns continuous programs for image generation and editing. Instead of a single fixed condition, the model can vary the operation at each step, allowing more precise control over how the generation process unfolds.
The paper's contribution is the shift from a uniform instruction to a per-step program. The abstract gives limited detail on implementation, but the stated goal is to enable more flexible and targeted manipulation of images.