GEN-3: The Ultimate Prompting Guide

Runway ML’s Gen 3 has launched as the successor to the highly popular Gen 2 model, signaling a major step forward in AI video technology. This video script provides an in-depth guide to using Gen 3, following the author’s research and experimentation. The speaker highlights Gen 3's ability to create advanced AI-generated video with less emphasis on keyword spamming and more descriptive, detailed prompts, showcasing its potential and growth since Gen 2. An example demonstrates the difference between an early Gen 2 text-to-video clip and a more refined Gen 3 output, showing remarkable improvement.

This section dives into the mechanics of prompting in Gen 3, which supports more descriptive prompts and encourages creative experimentation. It discusses the importance of including key elements in prompts—such as subject, action, setting, shot type, and style. A specific example shows how adding details to the prompt can significantly improve the quality of the AI-generated video. The speaker also provides a list of shot types and styles, encouraging users to experiment and test various approaches to optimize results. A PDF guide on Gumroad is available for users who want to explore further.

Here, the speaker explores Gen 3’s behavior when closely adhering to prompts, often using dissolves or cuts to bridge gaps when it can't fully achieve a requested action. One technique discussed is the use of seeds to reroll a prompt and maintain stylistic consistency while making iterative changes. Examples include a cyberpunk woman prompt, where the AI generated a scarier version on a second roll, and a workaround method to control the results. The speaker also talks about a music video project inspired by these techniques and mentions copyright issues with sharing it on certain platforms.

This part explores how using specific keywords like 'suddenly' can lead to dynamic effects in Gen 3 videos. For example, the speaker demonstrates a prompt where rain falls over a city, resulting in a fast zoom into a coffee shop. Though the results weren't perfect, the AI managed to deliver most of the desired effect. The section also highlights an impressive Gen 3 feature: its ability to incorporate text, showcased through a community member’s prompt mimicking the MCU Marvel opening sequence. The speaker notes that some names and keywords trigger content restrictions in Gen 3.

In this segment, the speaker discusses how imaginative prompts can generate unique visual concepts, such as a miniature civilization building castles and cities from a map. The prompt draws inspiration from a community member’s idea, and the speaker’s own attempt to recreate a Game of Thrones-style intro using a medieval fantasy map. Despite challenges in achieving time-lapse effects, the speaker appreciates the originality and potential of the output. The section ends on a lighter note with an amusing example of a puppet talking to a man, illustrating how even simple prompts can yield entertaining results.

This part focuses on Gen 3’s limitations, particularly its inability to accurately recreate specific scenes from actual movie scripts. The speaker shares an attempt to recreate a scene from *The Dark Knight* screenplay, resulting in humorous, DIY-style output reminiscent of the movie *Be Kind Rewind*. Additionally, Gen 3 performs well with time-lapse prompts, as demonstrated by a simple prompt where days rapidly turn into nights. The speaker advises users to rate their outputs to help improve the model, which is still in its alpha phase.

In the final paragraph, the speaker discusses exciting future features for Gen 3, such as image-to-video capabilities and the possibility of motion brushes. Though the speaker speculates about how these features might work, no official details are provided. The section encourages users to continue exploring Gen 3’s potential through prompting experimentation and to share findings with the community. The speaker closes with an invitation for viewers to leave comments and prompts to further advance the collective understanding of this evolving AI video technology.