High-level semantic understanding of long-form videos critically depends on accurate temporal action segmentation. Although generative diffusion models have recently introduced a new modeling paradigm for this task, most existing methods initialize the diffusion process with random gaussian noise, which induces substantial uncertainty in the generated outputs and makes it difficult to ensure stable and reproducible predictions. To address this limitation, we propose a video feature-guided temporal action segmentation method (VFG), which uses semantically rich video features as a deterministic prior for the generation process to guide the model toward learning task-relevant discriminative representations. Specifically, instead of relying on purely random gaussian initialization as in prior diffusion-based methods, we construct a feature-guided latent representation during training by combining encoded video features with a scheduled amount of ground-truth guidance. This provides stronger task-aligned supervision and improves the accuracy and consistency of action boundary prediction. In addition, we design a dedicated decoding architecture that enhances sensitivity to local details while also modeling long-range contextual dependencies. Extensive experiments on three benchmark datasets show that VFG is competitive in accuracy and computational efficiency.
Li et al. (2026) studied this question.