Abstract
Multimodal diffusion models have become a central approach for high-quality content generation. Some recent studies explore cross-timestep gradient relations to guide improvements in diffusion training. However, commonly used full-gradient cosine similarity can be distorted by parameter heterogeneity and layer-wise scale differences, leading to misleading estimates of the coupling among timestep-specific objectives. This paper revisits cross-timestep gradient measurement and proposes a corrected protocol for multimodal diffusion training. The corrected measurement shows that gradient similarity decays rapidly with timestep distance, and gradients from distant timesteps become nearly orthogonal as training progresses. These results suggest that diffusion training consists of weakly coupled timestep-wise objectives. They further imply that directly applying history-accumulating mechanisms across timesteps may introduce interference among weakly related timestep objectives. Our study refines the empirical basis for timestep-wise optimization dynamics analysis and provides guidance for optimization design in multimodal content generation.
Funding Information
  • National Natural Science Foundation of China (U2541212)
  • National Natural Science Foundation of China (62302023)