ShiftDDPMs: Exploring Conditional Diffusion Models by Shifting Diffusion Trajectories

Diffusion models have recently exhibited remarkable abilities to synthesize striking image samples since the introduction of denoising diffusion probabilistic (DDPMs). Their key idea is disrupt images into noise through a fixed forward process and learn its reverse generate from in way. For conditional DDPMs, most existing practices relate conditions only fit it reversal unconditional process. We find this will limit condition modeling generation small time window. In paper, we propose novel flexible model by introducing utilize extra latent space allocate an exclusive trajectory for each based on some shifting rules, which disperse all timesteps improve learning capacity model. formulate our method, call ShiftDDPMs, provide unified point view related methods. Extensive qualitative quantitative experiments synthesis demonstrate feasibility effectiveness ShiftDDPMs.