Runway Gen-4 characters are old and changeable, which is usually not something that can be solved with a single prompt. Its reference map capability is suitable for improving the consistency of characters, scenes, and objects, but only if the reference image is clean enough, the prompt text does not repeatedly change identity characteristics, and the lens is reasonably disassembled. If you want to stabilize the character, put "Reference Image Quality" before the prompt.
How to choose a reference picture
Prefer pictures with clear faces, uniform lighting, no obstruction, and no exaggerated filters. Don't use multiple photos, low-definition screenshots, strong expressions, half faces, or complex backgrounds. The model needs to grasp the identity characteristics from the picture, and the more chaotic the reference picture, the easier it is to change the face later.
If the character has fixed costumes, hairstyles, and props, it is best to prepare a multi-angle view of the same character, but the style should be the same. Don't mix one realistic, one anime, and one strong beauty, which will make the model oscillate between identity and style.
How to write prompts
Don't redescribe a set of facial features in every shot. For example, if the first paragraph writes "young woman, short hair, round face", and the second paragraph writes "mature woman, long curly hair, pointed chin", the model will think that you are changing roles.
A more stable way to write is to use a uniform title, such as "the same woman from the reference image", and then describe only the action, the environment, and the shot: turn around, walk into the room, look out the window, close-up, side backlight. The identity information is handed over to the reference map, and the lens information is handed over to the prompt.
Don't write too long a shot
Character consistency is most afraid of too many things happening in one clip at the same time: changing scenes, changing clothes, turning around, running, from distant shots to close-ups, and picking up props. Splitting a complex video into multiple short shots, each shot only completes one action, and then using editing to connect it, the stability will be much higher.
If the same character needs to span multiple scenes, Mr. will form a group of character makeup drawings and use these drawings as subsequent references. Don't take a blurry frame of a video every time and continue to expand.
When to lower your expectations
The reference image of AI video is not a traditional 3D character binding, and it cannot guarantee that every frame will resemble the same actor. For scenarios such as commercial advertisements, serial characters, and brand endorsements, the time for manual screening, regeneration, and post-face retouching must be retained. Gen-4 can significantly improve consistency, but don't think of it as a zero-cost continuous role system.
