In the realm of AIGC, the art of text prompts is crucial. The release of Runway's Gen-3 Alpha model has sparked a revolution in how to precisely control generative AI. This is not just a triumph of technology, but a leap in creativity and expression.
When using generative AI products like Gen-3, you might find that the results sometimes fall short of expectations. This is usually due to insufficient or inaccurate prompts. For example, a simple "desert sunset" is too vague to guide the model to understand your true intentions.
The correct approach is to provide detailed visual descriptions, such as "a desert at sunset, with golden sunlight cascading over undulating sand dunes, a line of camels returning from afar, kicking up a lot of sand, in high-definition realistic slow motion," which allows the model to more accurately capture your creative vision.
Gen-3's Text Prompt Structure
Runway recommends breaking down the details of the scene, subject, and camera movement into different modules, as this structure is clear and yields the best results. The standard prompt structure for Gen-3 includes text prompts to establish the scene and additional details, for example:
Close-up shot: High-definition fingertip with a curious mini dog standing on it.
Scene transition: Flying over an Irish castle to discover skyscrapers of a futuristic cyberpunk city.
The effect of Runway Gen-3 is excellent, but the high membership fee deters many users. If you are hesitant to subscribe to Gen-3, that's okay, as the same prompts can be applied to Kelin.
However, different products may have some logical differences. For instance, the prompt "high-definition fingertip with a curious mini dog standing on it" might not easily generate the desired result on Kelin. This is because Kelin's logic adheres to real-world logic, so the generated dog might not be as mini as the one produced by Runway Gen-3, which offers a blend of realism and fantasy.
To achieve such an effect with Kelin, you would need to slightly modify the prompt, specifying how small you want the dog to be.
Below, AIbase uses this logic to recreate a miniature zebra on a fingertip. The prompt is as follows: A high-definition fingertip with a cute mini zebra walking on it, as small as an ant, in a miniaturized style.
Then, you can obtain such a cute mini zebra:
Using the same approach, you can simply replace the animal name to create various miniatures like pandas, tigers, and more.
Here are detailed prompt examples provided by the official:
Seamless Transition
Continuous ultra-high-speed FPV shot: The camera seamlessly flies over a glacier canyon to a dreamlike cloudscape.
Camera Movement
A glittering ocean at night, with bioluminescent creatures underwater. The shot starts with a macro close-up of a glowing jellyfish, then widens to reveal the entire ocean lit up with various glowing colors under the starlit sky. Camera movement: Begins with a macro shot of the jellyfish, then gently pulls back and upwards to show the glittering ocean.
Text Title Card
A title screen with dynamic motion. The scene starts with a wall covered in colorful paint. Suddenly, black paint flows down the wall, forming the word "Runway." The dripping paint is detailed, textured, centered, and has excellent cinematic lighting effects.
Prompt Keywords
Keywords help achieve a specific style in the output. Ensuring that keywords align with the overall prompt makes them more prominent in the output.
For example, in a wide-angle shot where the camera cannot tightly focus on the face, adding keywords about skin texture is not very useful. Instead, adding more details about the environment could greatly benefit the wide-angle shot.
Keep this cohesiveness in mind, and here are different keywords to try when drafting a prompt:
