Temporal Specification Guidance Improves Temporal Control in Text-to-Video Generation

Summary

Text-to-video models can depict the concepts named in a prompt while failing to place them in the requested order, timing, or duration. The paper introduces Temporal Specification Guidance (TSG), a training-free inference-time framework that represents these requirements in formal temporal logic and converts them into objectives used during sampling. Its instantiation evaluates traces of visual-concept evidence with differentiable quantitative semantics for Metric Temporal Logic, then uses gradients to create attention biases and a specification-guided denoising direction. The method operates through video-text attention traces and does not require an external concept classifier or backpropagation through the video transformer. It augments classifier-free guidance with an attention-steered forward pass. On a benchmark of 220 prompts covering 11 categories of temporal constraints, including ordered object entry and state transitions, the method improved specification-satisfaction accuracy over classifier-free guidance by 9.5 to 20 percentage points across five video-generation models, with minimal memory overhead. The results indicate that temporal specifications can provide a flexible interface for controlling when visual events occur and how long they persist in generated videos.