Dancing Diffusion assigns a dedicated continuous condition block to each span of timesteps, turning conditioning into a latent program. A compiler maps text features into this program, and a fixed pretrained generator executes the blocks over successive sampling intervals. The trajectory therefore retains its executed history while its continuation can be redirected. A shared prefix is run once and reused; later operations can be rewritten while content fixed by earlier steps is preserved. The compiler is trained through the images it produces, with visual feedback on both the requested change and the retained content. Experiments on the SANA benchmark show that reusing an executed prefix preserves far more of the scene than regenerating from noise at comparable edit strength, learned conditions retain more background than native prompt switching from the same saved state, and the learned program improves generation quality over native SANA. Subsequent branches weaken the edit, limiting semantic control over individual operations. By converting a single‑instruction trajectory into an addressable sequence of operations, Dancing Diffusion lays a foundation for structured intermediate computation in diffusion models.
Review