William Raphael Battle, IV

General: I discovered the Version History feature, which allows users to recover previously generated clips that were inserted into the timeline. This proved valuable when experimenting with multiple generations and wanting to revisit earlier outputs.

​

Veo Model: 

Prompts: Through experimentation, I found there is a “sweet spot” when writing prompts. Overly detailed prompts often attempted to describe actions that were unrealistic for an 8-second generation window, while prompts that were too minimal lacked sufficient context for the model to produce strong results. To mitigate this, I kept a Google Doc open to structure prompts into smaller segments and generate clips iteratively. I also created a second Google Vids project file to store clips that were useful but not immediately needed in the timeline, preventing potentially valuable assets from being lost.

 

Uploading Images: Uploading up to three reference images generally worked well. However, it was unclear whether the model consistently recognized the intended order of the images (e.g., first, second, third) when referencing specific details. Because of this uncertainty—and the need to regenerate clips multiple times, sometimes waiting until the next day for the generation limit to reset—I experimented with preparing images beforehand using the Nano Banana model. This allowed me to add details and generate stronger point-of-view compositions before feeding them into Veo.

​

Audio: The auto-generated audio sounded strong when evaluated individually. However, when multiple clips were combined, the audio transitions became disorienting and disrupted the overall emotional continuity. To maintain narrative cohesion, I removed the generated audio and instead composed an original score in Logic Pro, which allowed me to control pacing, emotional tone, and rhythm across the full sequence.

​

Nano Banana Model: The Nano Banana model proved particularly effective for refining images before using them as references in Veo. It was useful for enhancing visual detail and establishing clearer contextual cues for the video generation process.

​

Voiceover Model: I expected a broader range of voice models similar to those available in AI Studio with ElevenLabs models. Ultimately, I chose not to use voiceover in this concept piece because it did not align with the visual storytelling approach or the tone of the narrative.

​

Upload Feature: I appreciated the integration with Google Drive and Google Photos, which made accessing assets convenient. However, a direct drag-and-drop upload option within the interface would improve workflow efficiency.

​

Captions: The preloaded caption styles were useful and visually polished. A valuable addition would be the ability to create and save custom caption styles as reusable templates.

​

​