Depth as Time Reveals Denoising Trajectories in One-Step Generative Models

Analysis shows that the multi-step denoising trajectory of diffusion models unfolds across the depth of a single forward pass in one-step models. By decoding intermediate layers with the model's output head, researchers can recover the sampling steps and observe how computation depends on the transport task.