SDPO: Importance-Sampled Direct Preference Optimization for Stable Diffusion Training
This addresses training challenges for diffusion models in generative AI, offering incremental improvements over existing methods like Diffusion-DPO.
The paper tackled instability and off-policy bias in preference learning for diffusion models by proposing SDPO, a framework that uses importance sampling to correct bias and emphasize informative timesteps, achieving superior VBench scores and human preference alignment on models like CogVideoX-2B and Wan2.1-1.3B.
Preference learning has become a central technique for aligning generative models with human expectations. Recently, it has been extended to diffusion models through methods like Direct Preference Optimization (DPO). However, existing approaches such as Diffusion-DPO suffer from two key challenges: timestep-dependent instability, caused by a mismatch between the reverse and forward diffusion processes and by high gradient variance in early noisy timesteps, and off-policy bias arising from the mismatch between optimization and data collection policies. We begin by analyzing the reverse diffusion trajectory and observe that instability primarily occurs at early timesteps with low importance weights. To address these issues, we first propose DPO-C\&M, a practical strategy that improves stability by clipping and masking uninformative timesteps while partially mitigating off-policy bias. Building on this, we introduce SDPO (Importance-Sampled Direct Preference Optimization), a principled framework that incorporates importance sampling into the objective to fully correct for off-policy bias and emphasize informative updates during the diffusion process. Experiments on CogVideoX-2B, CogVideoX-5B, and Wan2.1-1.3B demonstrate that both methods outperform standard Diffusion-DPO, with SDPO achieving superior VBench scores, human preference alignment, and training robustness. These results highlight the importance of timestep-aware, distribution-corrected optimization in diffusion-based preference learning.