Skip to content
Log inSign up free

Prompting guide

How to prompt Magejutsu

How to prompt Magejutsu: mapping each reference photo to a person or object in your clip, protecting what must not change, and a fill-in template.

Try MagejutsuMagejutsu model page

Recast a dancer
Made with MagejutsuRecast a dancer

Quick start

Your first great Magejutsu prompt

Everything a beginner needs in one minute: what to know first, the rules that matter most, a template to fill in, and prompts to copy.

What to know first

Magejutsu is an edit, not a generation: the clip you supply is the master, and the result keeps its length, shape, motion and sound. A director model watches your clip and writes the edit for you, so one clear sentence is usually enough — but you must say which reference is which, and what has to stay exactly as filmed.

The rules that matter most

  1. Name every reference by its position

    Your photos are numbered in the order you added them. Write "image 1 is the person on the left, image 2 is the person on the right" rather than describing them — the model matches by number, not by description.

  2. Say what must not change

    The clip is the master, so protect it: "keep the choreography, the camera move and the timing exactly as filmed". Without that, the model may re-stage or re-light what you wanted untouched.

  3. One clear photo per person beats a collage

    A single well-lit, front-facing photo matched as well as a multi-view card in testing, and is far easier to supply. Use a sharp photo where the face is large and unobstructed.

  4. Change one thing at a time

    A new cast, a new location, and a new outfit in one request pull against each other. If the result drifts, split it: recast first, then change the setting.

  5. Use it for objects and places too

    A product, an outfit, or a location is the same mechanism: give a clear photo of it and say where it goes. You do not need a person in the shot.

Fill in this template

Replace each part in square brackets with your own idea, and leave out a part you do not need.

Template
Replace [who or what] in the clip with image [N], taking [face, hair, build, and every garment] exactly as shown. Keep [the camera move, the background, and the timing] exactly as filmed. [Any second change, named plainly.]
The template, filled in
Replace the dancer in the white shirt with image 1, taking her face, hair, build and every garment exactly as shown. Keep the choreography, the camera move and the timing exactly as filmed. Move the scene to a neon night market.
  • [who or what] Point at the subject by what is visible: the dancer in the white shirt, the person on the left, the bottle on the table.
  • [N] The position of that photo in your reference list, counting from 1 in the order you added them.
  • [face, hair, build, and every garment] What to carry over from the photo. Being specific stops parts of the original subject leaking through.
  • [the camera move, the background, and the timing] What must stay as filmed. This is the line that keeps the edit faithful to your source.
  • [Any second change, named plainly.] Optional. One more change, stated as plainly as the first — or leave it out.

Prompts to start from

  • Put a character into a clip

    Put a character into a clip
    Replace the dancer with image 1, keeping the choreography, the camera move and the timing exactly as filmed.

    Why it works: The reference is named by number and the source is protected in one line, which is the whole structure of a good recast.

  • Two people, mapped

    Replace the person on the left with image 1 and the person on the right with image 2. Both keep their positions and timing; nothing else in the shot changes.

    Why it works: Mapping each photo to a named position is what makes a multi-person edit land on the right people.

  • Swap a product

    Replace the bottle on the table with the one in image 1, matching its scale and the room lighting. Keep the camera move and the hand exactly as filmed.

    Why it works: Objects work the same way as people, and matching the scale and light is what stops a swap looking pasted on.

  • Restyle the setting

    Keep every performer and their choreography exactly as filmed, and replace the location with the one in image 1, matching the source camera moves shot for shot.

    Why it works: Leading with what to keep, then naming the single change, keeps a big location swap from disturbing the performances.

Advanced

Advanced Magejutsu prompting

Techniques for more control once the basics work.

Mapping references to subjects

A Magejutsu request carries your clip as the source and your photos as references, numbered from 1 in the order you added them. The model resolves "image 1" by position, so the mapping you write is the mapping it follows.

Where a clip holds more than one person or object, the mapping is the whole job. Describe each subject by something visible and unambiguous — left and right, foreground and background, or by a garment — and give each one its own sentence.

Saved Characters and saved references can be attached like any other photo and mentioned by @handle in your request, so a character you use often is one token away.

  • Count your photos before you write, and match the numbers to the slot order you see in the picker.
  • For a busy shot, say what to leave alone: "everyone else in the frame stays as filmed".
  • One photo per subject is enough; adding several of the same person does not raise the count of subjects.

Four people, four photos

Replace the four skaters with image 1 (leftmost), image 2, image 3 and image 4 (rightmost) in that order. Each keeps their own position, timing and choreography; the street and the crowd stay exactly as filmed.

Why it works: Ordering the mapping to a visible sequence keeps four subjects from being assigned at random.

Protecting the source footage

Because Magejutsu edits footage rather than generating it, the most common disappointment is not a wrong face — it is the model re-rendering something you wanted kept. The fix is to say what is protected, explicitly, every time.

Name the elements that matter to you: the camera move, the cuts, the timing, the light, the background, a held prop. Anything you name tends to survive; anything you leave unsaid is the model's to reinterpret.

The result keeps the source length and shape automatically, so you never need to ask for a duration or an aspect ratio.

  • A phrase like "everything else stays exactly as filmed" is worth more than a long description of the scene.
  • If a cut or a wipe is important, say "keep every cut and transition in place".
  • When the source sound matters, it is preserved without being asked for.

Choosing a source clip

A clean, well-lit, steady clip gives the best result. Heavy noise, motion blur, or constant camera shake give the model less to hold on to, and small, distant faces are the hardest case of all.

A talking head facing the camera is the easiest source and the most reliable for likeness, because the face is large and the motion is simple. Crowded or fast action is the hardest, because there is more to keep track of.

The clip is the master for sound as well as picture: the audio you supply comes back with the result, unchanged.

  • Prefer a clip where the subject is clearly visible for most of the run.
  • For likeness, a source with a large, front-facing face beats a wide, busy shot.
  • Very fast action can make a face drift; if it does, try a shorter section of the clip.

Scenes with several people

Multi-person edits are the case where a plain request most often goes wrong: subjects get swapped, or a reference bleeds onto a bystander. Two habits fix most of it.

First, map each reference to a named person, as above. Second, add a line that limits each mapping to one person — "each reference replaces one person only; do not repeat anyone". That single sentence is what stops a face from being duplicated across a crowd.

Background people are also worth protecting by name. If someone in the distance should not change, say so.

  • Map first, protect second, then add any other change.
  • For a two-person scene, one clear photo each is easier for the model than a card showing both.

A fight scene, mapped and bounded

Replace the fighter on the left with image 1 and the fighter on the right with image 2. Each reference replaces one person only and is never duplicated elsewhere. Keep every punch, the camera move and the timing exactly as filmed; the crowd stays untouched.

Why it works: The "one person only" line is what keeps a busy scene from duplicating a face onto bystanders.

Products, outfits, and locations

The same mechanism covers non-people: a product on a table, an outfit on a performer, a location behind them. Give a clear photo of the thing and say where it belongs.

Two details separate a convincing swap from a pasted-on one: scale, and light. Say that the replacement matches the size of what it replaces and the lighting of the room.

For a location, the useful instruction is that the camera keeps doing what it did — shot for shot, move for move — with only the surroundings changed.

  • Name the surface or position: "on the table, in the same spot".
  • For an outfit, say the outermost layer stays on in every frame.
  • A location plate with no people in it is the cleanest reference for a setting change.

When to split a request

A single request can carry several changes, but they compete. When a result comes back with one thing right and another wrong, the efficient move is usually to split it rather than to write a longer request.

Recast first on the original footage, then run the result again for the setting or the outfit. Each pass then has one job, and you can see which one is failing.

This also helps with length: a short section of a long clip is easier to get right than the whole thing at once, and the pieces can be joined afterwards.

Common mistakes

Avoid this, do this instead

  • Avoid: Describing your references in words instead of numbering them.

    Replace the dancer with the woman from my second photo.

    Instead: Say "image 1" and say which subject it replaces. The model matches by position.

    Replace the dancer with image 2.

  • Avoid: Asking for a change without protecting the rest.

    Put my character into this clip.

    Instead: Name what must stay as filmed: the camera move, the cuts, the background, the timing.

    Replace the performer with image 1. Keep the choreography, the camera move and the timing exactly as filmed.

  • Avoid: Mapping several people with one vague sentence.

    Swap the two fighters with my two photos.

    Instead: Give each subject its own sentence, described by a visible position, and limit each reference to one person.

    Replace the fighter on the left with image 1 and the fighter on the right with image 2. Each replaces one person only.

  • Avoid: Stacking a recast, a new location, and a new outfit in one pass.

    Put image 1 in the scene, move it to a beach at sunset, and change the outfit to image 2.

    Instead: Change one thing at a time, and split the request if the result drifts.

    Replace the performer with image 1, keeping everything else exactly as filmed.

  • Avoid: Reaching for a busy action clip when what you want is likeness.

    Instead: Use a source where the face is large and clearly visible; save fast action for when motion is the point.

FAQ

Magejutsu prompting questions

How long should the source clip be?
Four to thirty seconds. The result keeps the source length, so a longer source makes a longer edit rather than a longer generation.
Do I get the original audio back?
Yes. The clip you supply is the master for sound as well as picture, and the audio returns with the result.
How many reference photos should I attach?
One clear photo per subject is the sweet spot: a single sharp, front-facing photo matched as well as a multi-view card in testing.
Can I use it on someone else’s footage?
Only with their permission and the rights to the material. Magejutsu carries the same consent requirement as every other face-editing tool on Mage.
Why did my result change something I did not ask about?
The model reinterprets whatever you leave unsaid. Name the elements that must survive — the camera move, the cuts, the background, the timing — and they tend to stay.
Does a bigger reference set make a better result?
Not for people. More photos of the same subject do not count as more subjects, and one clean photo is usually enough. Several references help when there are several distinct things to place.

Make your first video with Magejutsu

Start free, then run Magejutsu on Gems or a plan. No install.

Try Magejutsu