Convert a Reference Video to a Kling Prompt Without Copying It

A reference video is a brief you did not have to write. Someone already solved the hard parts — where the camera sits, when the movement happens, how the light falls — and your job is to keep those decisions and replace the content with your own. That is the whole craft of reference-driven generation, and Kling is a good model to learn it on because it follows motion instructions so literally.

What a reference is for

People usually bring a reference for one of three reasons: a look (a grade and lighting pattern), a move (a camera behaviour they cannot describe from memory), or a rhythm (where the beats of the shot land). Know which of the three you want before you write anything, because that determines what must survive the swap. If the move is the point, the move is sacred and the subject is interchangeable; if the look is the point, the grade has to be described precisely enough to reproduce.

Extraction is what turns that instinct into language: SocialToPrompt reads a clip — from 27 named platforms or an uploaded file — and returns the fields a reference is made of.

The three-keep, three-swap rule

Keep from the reference	Swap for your own
Camera verb, height and stability	Subject
Light direction and quality	Setting
Pacing and where the change lands	Palette and grade

The keeps are mechanism; the swaps are content. This is also the ethical line made operational: a person’s likeness, script, brand marks and distinctive set design live entirely in the swap column, and if you find yourself unable to describe your output without naming the original, you have copied expression rather than technique. The longer version is in the reverse-engineering method.

Converting to Kling clauses

Reference (a 6s perfume ad): black set, bottle rotates on axis, single hard rim from behind-left,
background falls to pure black, slow, ends with label facing camera.

Kling prompt, your product (a hot sauce bottle):
A glass hot-sauce bottle rotates slowly on its axis against a deep charcoal set; camera locked-off,
eye level with the bottle, no drift; single hard rim light from behind-left tracing the glass edge,
label edge-lit; condensation beads sliding down the shoulder; saturated reds against neutral grey,
high contrast, blacks crushed; 6 seconds; ends with the label square to camera.

Every structural decision is inherited — the rotation, the rim position, the locked-off frame, the ending — and nothing about the original subject survives. Two Kling-specific touches: naming the surface behaviour (condensation sliding) and stating that the background falls to near-black, which is what makes the rim light read.

Locking constants across shots

Reference work is usually a sequence, so declare the constants once and repeat them verbatim in every shot prompt: the subject noun, the light source and position, the palette words, the format claim, the aspect ratio. Kling has no memory between generations, so a reworded noun becomes a different bottle by shot three.

Keep the constants in a short header block above your prompts and paste it into each one — the same discipline as the multi-shot structure in Seedance prompting. When you have several references in play, that is a library problem: shot-list extraction covers the multi-shot case.

Testing the reference held

Put the render next to the reference at the same size and ask three narrow questions: did the camera do the same thing; is the light coming from the same side; did the change land at the same moment? If all three are yes, the reference converted correctly and any remaining dissatisfaction is about your subject, not the mechanism.

If the camera failed, fix the verb first — that is where extraction and imagination most often disagree, and the corner test in the camera movement guide settles it in seconds. Then re-run with the one-variable discipline from the eight-slot template. Related conversions: Kling field-by-field extraction, Seedance conversion, and Sora prompts from existing video.

Frequently asked questions

What does it mean to use a reference video for a Kling prompt?

It means taking the structural decisions someone already made — camera behaviour, light direction, pacing, where the change lands — and refilling them with your own subject, setting and palette. The reference supplies the mechanism; you supply the content.

How much of the reference should I change?

Exactly three things: subject, setting and palette. Keep camera, light and timing intact. That split is also the rights boundary, because everything you are allowed to reuse lives in the structural column and everything protected lives in the content column.

My sequence shots don't match each other. Why?

Because the constants drifted between prompts. Repeat the subject noun verbatim and restate light source, palette words, format and aspect ratio in every shot; models have no memory of the previous generation.

How do I know the conversion worked?

Compare at the same size and answer three questions only: same camera behaviour, same light side, same moment of change. If all three hold, the mechanism transferred and anything you dislike is about the subject rather than the shot.