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.
As machine learning advances, machine learning as a service (MLaaS) in the cloud brings convenience to human lives but also privacy risks, as powerful neural networks used for generation, classification or other tasks can also become privacy snoopers. This motivates privacy preservation in the inference phase. Many approaches for preserving privacy in the inference phase introduce multi-objective functions, training models to remove specific private information from users' uploaded data. Although effective, these adversarial learning-based approaches suffer not only from convergence difficulties, but also from limited generalization beyond the specific privacy for which they are trained. To address these issues, we propose a method for privacy preservation in the inference phase by removing task-irrelevant information, which requires no knowledge of the privacy attacks nor introduction of adversarial learning. Specifically, we introduce a metric to distinguish task-irrelevant information from task-relevant information, and achieve more efficient metric estimation to remove task-irrelevant features. The experiments demonstrate the potential of our method in several tasks. Our code will be available at: https://github.com/iwhoyoung/PriFU.