Lesson 2 — Fix the still before you generate the video
Module 1 · ~11 min
This is the mistake that costs the most and gets noticed the least: going straight to video.
You write the prompt, you generate, and you get five seconds where the face drifts, a hand is wrong and the light turns halfway through. So you adjust the text and generate again. And again. Each attempt is a video generation — the most expensive thing on the menu.
Why it happens
Every house model accepts an image as start_image. When you don't give it one, the model invents the first frame from your text — and at that point you're asking for two things at once: that it imagine the scene and that it move it. Two problems in a single generation, and when it comes out wrong you can't tell which one failed.
Handing it a finished image removes the first problem entirely. The model is left with one job: move what's already there.
The method
- Write the scene as an image prompt, not a video prompt. One frame: who, where, what light, what lens feel.
- Generate stills until one is right. Images cost a fraction of what video costs — use that as the cheap part of the process, because it is.
- Look at the still for what video will amplify: the face, the hands, text in frame, the direction of the light. If any of them is wrong in the still, it will be wrong and moving in the video.
- Only then animate, passing the image as start_image, and use the prompt only for motion — the content is already settled.
Why the flaw gets worse in motion
A hand with six fingers in a photo is a mistake you can see. The same hand in motion is a mistake that pulls the eye for five straight seconds. Video doesn't soften a defect — it gives the viewer time to find it.
The same goes for text in frame. A wrong letter in a still, you crop. A wrong letter shimmering through a video is the visual signature of "this was made by AI", which is exactly what you're trying to avoid.
When to skip this lesson
When motion is the idea — an effect where the transformation is the whole point, and there's no single frame the shot is "about". Then there's nothing to settle in advance and you go straight to the generation.
Everything else — a person talking, a product turning, a place being revealed — starts as a photograph. Treat it like one.