Как писать промпты для видео в нейросети, как режиссёр

Расшифровка английского закадрового текста.

0:04A video prompt is a director's brief to a crew that takes every word literally and fills every silence with a default. So a good prompt isn't poetic, it's precise, ordered and complete. In this lecture we'll take apart the real prompt that generated the photo, line by line, and you'll leave with a structure you can reuse on every shot.

0:27Here's the structure, in order. The camera, shot size and movement. The subjects, who is in frame and exactly where. The action, what happens as a sequence of physical events. The performance, emotion shown through visible behaviour. The environment and the light. And finally, the sound and the constraints. Same order, every shot. The model reads it as a plan.

0:57Start with the camera, because it frames everything else. Medium two shot, dolly in. Close up, static camera. One movement push shot. And if it moves, say how fast. A slow dolly in is a different emotion than a fast push. If you don't name a movement, the model will pick one for you. And that's how you get the floating camera.

1:22Then place your subjects, and anchor them. Here's a line from our real prompt. Thomas sits behind the steering wheel in the driver's seat, while Emily sits beside him in the front passenger seat, leaving the rear passenger seats completely empty. That last part looks redundant. It isn't. It's there because in an earlier take, Emily ended up in the back. Write the positions in every shot, even when they don't change.

1:51Now the action. Describe physical events, in order, with a beginning and an end. Thomas turns his torso across the centre console toward Emily, and extends an old photograph toward her, while Emily reaches forward and receives the paper between both hands. Every movement has a direction and a destination. And for objects, protect them. Our prompt says, keeping the photograph as a single, unchanged physical object. Objects that appear, multiply or melt are the most common physical failures.

2:26Write the action at the speed of the shot. Five seconds is two or three small actions, not a whole scene. If your action would take a real actor fifteen seconds to perform, it won't fit in five.

2:41Then the performance. Never write a feeling the camera can't see. Don't write, she is overwhelmed. Write what overwhelmed looks like. Tears well in her eyes, and a faint smile appears on her face. Give the actor a small, specific action. A thumb brushing the edge of a print. A head tilting down. Small actions read as truthful. Big ones read as acting.

3:07If your model can generate several shots in one clip, write the prompt as a timeline. Zero to five seconds, the first shot. Five to nine seconds, hard cut, the second shot. And so on. Our prompt has four time -coded blocks, one per shot. Each starting with its summaries and its movement. The words, hard cut, matter. They tell the model this is a new shot, not a camera move.

3:33When you give the model reference images, name them. And say what each one governs. Our prompt says, the vehicle interior and ambient lighting of image one. And that the character reference governs her identity and wardrobe. A reference without a job is a suggestion. A reference with a job is an instruction. In Screenweaver, the prompt director writes these bindings for you. From your shot actions and your app mentions. And the eye icon on the video node shows you exactly what the model receives.

4:03Then describe the world and the light. Once. The same way in every shot. Hours ends with, the entire scene takes place inside the continuous car interior. At night, during steady rain, illuminated exclusively by soft street lights and city bokeh through rain -glazed windows. That's the light logic from our visual bible, word for word.

4:26And close with sound and constraints. Our prompt says, no subtitles, no captions and no on -screen text. And the soundtrack consists solely of natural interior car ambience. Rain pelting against the glass, the muffled city hum, Thomas' spoken line and the friction of fingers on paper. With no added music score. Without those lines, you'd get burned in captions and a generic music bed.

4:54Here are the five mistakes I see most often. Adjective soup. Epic, cinematic, stunning, which describe nothing. Two camera moves in one shot. Emotions without actions. Text that contradicts the reference images. Like a red coat when the sheet shows beige. And prompts so long that the important instructions get lost. If a line doesn't change the picture, cut it.

5:24Your exercise. Take one shot from your shot list and write its prompt in the six -part order. Camera, subjects, action, performance, world and light, sound and constraints. Then generate it and check every line of the prompt against the result. In the next lecture, we'll go deeper into references. How to keep your characters and your sets on model, shot after shot.