Impact of Text Prompts on Video Generation Results #50

In the paper, the block diagram suggests that different videos are assigned corresponding text prompts. However, in the inference examples, only very short prompts are used.
​​1.Prompt Importance Comparison:​​

• Has there been any comparative analysis on how prompt length/structure affects the final results?
• Is the model relatively insensitive to prompts because ​​Wan 2.1​​ inherently supports short prompts for image-to-video generation?
• Conversely, would models like ​​DiT​​—which rely on longer, structured prompts—be more sensitive to prompt variations?
  ​
  ​2.Follow-up Questions:​​
• Were there ablation studies on prompt design (e.g., concise vs. detailed descriptions)?
• If prompt sensitivity varies across architectures, could this explain performance differences in controlled experiments?

This would help clarify the role of text conditioning in video generation. Thanks!