Video data consists of sequences of visual frames and can be analyzed by language models for useful insights. A practical task is generating a suitable title from video content. section 1.2 section 6.0
A video-analysis request typically combines:
- Selected video frames, submitted as images. section 6.4.1
- A text instruction such as “Generate a concise title for the video.” section 6.4.1
- A generated output, such as a concise title describing cars or a visible location. section 6.4.1
Generating a video title
Send a selection of frames together with an explicit instruction to GPT-4o. The model can return a title based on visible subjects and text in the frames, but each submitted frame increases processing cost. section 6.4.1
Figure 6.3 Multimodal prompt for video processing: based on a selection of video frames, the language model is instructed to generate a suitable title.
Use the prompt illustration to inspect how selected frames and a title-generation instruction form one multimodal request. section 6.4.1
