After 500 experiments, we discovered the simplest formula works best. Six components in the right order is all you need.
Let's talk about what everyone does wrong. People write a prompt like "a cat sitting on a table." They get a generic image of a cat on a table. Then they say "AI doesn't work well." No. The AI works great. The prompt is the problem.
After 500 experiments (yes, we counted), we arrived at a simple formula that always works. 6 components, in this order: Subject + Action + Style + Camera + Lighting + Mood.
Worth knowing: Subject: Not "a woman" but "a 30-year-old woman with red hair and a white shirt." Action: Not "walking" but "striding confidently through heavy rain." Style: "Cinematic, film grain, Kodak Portra 400." Camera: "Close-up, slow dolly-in." Lighting: "Golden hour, Rembrandt lighting." Mood: "Nostalgic, dreamy."
Before: "cute cat." After: "A fluffy orange tabby cat curled up on a velvet armchair by a window. Rain streaks on the glass. Warm candlelight. Shot on Hasselblad X2D, 85mm f/1.4. Cozy, intimate atmosphere." The difference in results? Like the difference between a phone snapshot and a magazine photo.
This is the first question in every workshop, and the answer surprises people: write in Hebrew. The engines understand Hebrew well at the level of intent, and you can describe a whole scene in Hebrew and get exactly that. What is true is that professional cinematography terms come out more precise in English, simply because the engine was trained on millions of English descriptions, and shallow depth of field is one clear concept to it while the Hebrew equivalent is three words it has to guess at.
Worth knowing: The formula that works: a Hebrew sentence with the professional terms left in English inside it. "A thirty-year-old woman stands at a large window, rain outside, shallow depth of field, golden hour." That is not a compromise and not laziness, it is simply what produces the best result.
After people learn the formula they make the opposite mistake. They write an eighty-word prompt with four actions, three camera moves and two lighting setups, and they get a mess. A video engine does not execute a task list, it draws one moment. A five-second shot contains one action and one camera move. That is it. If you wanted her to turn, lift the cup and have the camera fly backwards, that is three shots and not one, and if you ask for all three at once the engine will do a fourth thing.
Worth knowing: A simple test before you click: read the prompt out loud and ask "how many things happen here?" If the answer is more than one, split it.
Some things people try to solve with phrasing simply cannot be, and knowing them up front saves credits. Hebrew text inside an image: most engines draw letters as shapes and learned a left-to-right direction, so Hebrew comes out reversed. Two engines in the system do solve it, and for all the rest you add the wording on a separate layer. Exact counts: "five people" will sometimes return four and sometimes seven, because the model estimates quantity rather than counting. And Hebrew speech inside a video shot does not exist at all: the voice is built separately and synced afterwards.
This is the most profitable rule in this whole article, and it is not about phrasing at all. Low resolution while you are still searching for the scene, full resolution only once the shot works. Hunting for an idea at four credits a shot feels completely different from hunting at twenty-one, and the final result is identical, because all that changed is the tier you closed on. Same with stills: generate five cheap options, pick one, and upscale only that one.
And one more money saver: when something comes out almost right, do not write a new prompt from scratch. Change one component. If the face is good and the lighting is flat, touch only the lighting. A brand new prompt sends you back to the start and loses what already worked, and people do this over and over without noticing.
The formula is the same, but two components change weight. In a still, the camera is an angle: close-up, from above, profile. In video, the camera is a movement, and a movement is a decision in itself. "The camera pulls back slowly" and "the camera tracks from behind" produce two completely different videos from the exact same scene. It is also the easiest place to ruin one: a camera move that is too fast in a five-second shot turns everything into noise.
The second component that changes weight is the action. In a still you can write "sitting" and get a good image, because a still is a frozen moment. In video "sitting" is nothing, because nothing happens across five seconds. Video needs an action with a beginning and an end inside that time: "puts the cup down and lifts her gaze". An action that does not finish within the shot length will be cut off mid-way and look broken.
Our most important tip: Think like photographers, not programmers. Describe the scene as if you're explaining to a director what you want to see.
עמוד הבית · בלוג · וידאו AI · תמונות AI · קול AI · כל המנועים · קרדיטים ולא מנוי · AI בעברית · קורס מתחילים · מדריכים