Camera
Start frame or text: which gives you a predictable video
If your film has a person, a product or shots that must match, start from an image. Here is why, and when plain text is still enough.
By: Vadász MárkPrompting video models (Seedance, Kling, Veo)
The short answer: if the shot contains a person, a product or anything that has to look the same in the next shot, make a start frame first and generate the video from it. Text-to-video is fine for loose atmospheric shots, where it doesn't matter that every attempt comes out a little different.
The reason is simple. From text, the model has to invent the face, the costume, the light, the framing and the movement all at once. From a start frame, it only decides the movement. Everything else is handed to it.
What the model decides for you
A text-only video prompt carries an enormous load. Whatever you leave out, the model fills with its own average, and that average lands somewhere different every time. In one version the hair is curly, in the next it is straight, in the third the light comes from the other side.
For a single shot you might live with that. In a six-shot ad, though, you end up with six different women in six different kitchens, and no edit will rescue that.
Why the start frame is the safer route
Our planner writes a start-frame prompt for every shot, and only then a video prompt. The reasons are practical, not a matter of principle.
- You see the frame before you spend anything on video: a still is quicker to regenerate than a clip.
- The video inherits its realism from the start frame: skin, light and detail are settled here.
- You can repeat the key traits of the character description word for word in every image, so the face stays the same from shot to shot.
- You check the hands, the composition and the background before anything moves.
A good start-frame prompt runs in this order: shot size, angle, lens, then who is doing what and where, then the scene sheet (time of day, light, palette), the character's fixed traits, the realism block and the composition line. A short ‘avoid:’ list can close it.
Medium close-up, eye level, 85 mm lens: a woman in her early thirties pours coffee from a moka pot at a narrow kitchen window, early morning, low warm sunlight from the left, palette of cream, rust and muted teal. Oval face, straight dark eyebrows, a small mole above the left lip, light olive skin with a warm undertone, faint freckles across the nose, shoulder-length wavy chestnut hair in a loose low bun, oversized grey knit cardigan. Photorealistic, pore-level skin detail, natural matte skin with slight tonal variation, tiny natural imperfections, a few loose flyaway hairs, fabric with real weave and natural creases, diffuse realistic light falling off across the face, no beauty retouch, no airbrushing, no plastic glow. Key subject and action kept within the central square of the frame, comfortable headroom, nothing essential near the top or bottom edges. avoid: waxy or plastic skin, airbrushed look, glossy AI sheen, perfectly symmetrical face, extra or fused fingers.
Two prompts, two separate jobs
The most common mistake is pasting the whole image description into the video prompt as well. There is no need: the image has already said who and where. The video prompt only describes what changes.
One shot, one camera move. Give distance and time rather than adjectives of intensity, and name the opening and closing frame of the move. Camera movement and character movement go in separate sentences.
The camera moves in from a medium close-up to a close-up of her face over the full five seconds, travelling about half a metre. The shot opens on her hands and the moka pot and ends on her eyes and the steam rising past her cheek. She finishes pouring, sets the pot down and exhales, and her shoulders drop. Her gaze leaves the cup, goes to the window and returns. Steam rises continuously from the cup, clearly visible.
This is a Kling-style prompt: natural full sentences, one move, no negative list (Kling takes that in a separate field). For Seedance, 60 to 100 words is the right length; for Veo, the same principles apply, kept concise.
When text is enough
Starting from text still has its place. Clouds, waves, fog over a lake, abstract textures: nobody will notice if the second version differs from the first. It is also handy for a quick check on whether a camera move works at all.
If the shot is going into the film, though, and has to match the palette and light of the other shots, a start frame serves you better here too. The scene sheet covers the whole film, and a text-only generation rarely hits it exactly.
Which model for which step
For start frames with people and faces, Nano Banana Pro is the default, with the fewest hand errors. For locations and materials, use Nano Banana 2; for top-down product shots, Seedream 4.5. For video: Kling 3.0 for silent cutaways and products, Seedance 2.0 for living bodies and emotion, Veo 3.1 for people walking through a space.
How to do it on your next film
- Scene sheet and characterWrite down the time of day, the light, three to five colours and the character's face, hair and costume, once.
- Start framesMake a still for every shot, and check the hands, the face and the composition.
- FixingRegenerate only the faulty image, and leave the character and scene descriptions untouched.
- Video promptOne move, distance and time, opening and closing frame, the character in a separate sentence.
- If it goes wrongIf the movement fails, simplify the move, but keep the good start frame.
Common questions
Should I start an AI video from text or from an image?
If the film has a person, a product or connected shots, start from an image. From text, the model has to invent the face, clothes, light, framing and movement all at once. From a start image it only decides the movement and gets everything else ready-made, so it is the more predictable route.
When is text-to-video good enough?
When it is a loose atmospheric shot and it does not matter that every attempt looks a little different: clouds, waves, mist over a lake, material textures. It is also good for quickly testing whether a camera move works. For anything that goes into the film and has to match the light of the other shots, a start image is better.
What goes in the video prompt if I already have a start image?
Only what changes. The image has already said who and where, so there is no need to paste the full image description. One shot, one camera move: give a distance and a time, name the first and last frame of the move, and put the camera movement and the character's movement in separate sentences.
How do I keep the same character across several shots?
If you do not describe the character, the model fills the gap with its own average, and that average lands somewhere different every time: six shots give you six different women in six different kitchens. With start images you can repeat the key details of the character description word for word, so the face stays the same from shot to shot.