Prompting edits and references

An edit takes an existing image and an instruction and returns the same image with that change made. Edits are how you:

• make the end frame of a shot, by editing its start frame into the state the shot should finish in,
• refine or correct an image,
• bring a specific person, object or style into a picture with references.

Prompts can be up to 4000 characters. Edits are priced by render time: 2 credits at 1 MP, 5 at 2 MP, a little more per reference mentioned.

The short version #

1. Write an instruction. Lead with the operation: Add, Remove, Replace, Change, Move, Turn, Open, Make it night.
2. Say exactly what changes, where, and what it looks like afterwards.
3. Say once that everything else stays as it is.
4. Refer to the image being edited as @frame and to a reference as @label.
5. Leave the seed unset.

Replace the paper cup in the man's right hand in @frame with a white ceramic mug with a thin blue rim. Keep his pose, face, clothing and the background exactly as they are.

Writing the instruction #

Be specific about the change. If the request is vague, the model fills the gaps for you. Give category, colour, size, orientation and position.

Make the change unmistakable. A timid instruction gives a result that looks like the input. Describe the end state fully: The window is now wide open, both panes swung inward, the curtain blown into the room.

State what stays, positively, and once. Keep the framing, the background and the subject's size and position exactly as they are. One blanket clause works better than listing every detail: the more you describe something you meant to keep, the more likely it is to be redrawn. Avoid lists of prohibitions.

Name the style by its features. Restyle as 1970s disco: mirrored walls, a disco ball, coloured spotlights, warm saturated tones. If you also change content, put the style part last.

Text edits. Quote the text exactly and phrase it as a replacement: Replace "SALE" with "SOLD OUT". The original lettering style is kept unless you ask for another: Change "Harbour Cafe" to a black hand-painted brush script.

Viewpoint. Show the same scene from a low angle, Turn her to face left, Pull back to a wide shot of the same room.

Do not reuse the seed of the image you are editing. Editing an image with the same seed that generated it tends to return it almost unchanged. Leave seed unset, or pick a different one.

End frames for video #

An end frame answers one question: what does the last frame of this shot look like? The video model then finds the motion between the start frame and it.

• Change only what moves or transforms. Same scene, same props, same framing unless the camera is meant to move. Things that appear in only one of the two frames pop in or out.
• Keep the change achievable in the clip length. A few seconds is enough for a door to open, a figure to cross a room, a flower to bloom, daylight to turn to dusk. It is not enough for a whole story.
• Camera moves are end frames too. Pull back to a wide shot of the same room, the woman now small in the centre. The clip becomes a move with a defined destination.
• Good subjects: time-lapses, light and weather changes, colour and state changes, growth and decay, slow reveals, an object assembling or dissolving, a character changing pose or expression.
• Edit from the step's selected start frame, or refine an existing end-frame candidate.

Change @frame to late evening: the sky is deep blue with the first stars, the street lamps are lit and cast warm pools of light on the wet pavement, and the shop window glows from inside. Keep the camera position, the buildings and the parked bicycle exactly as they are.

References #

A reference is a supporting image with a label. Mentioning @label in a prompt hands that image to the model alongside your words. References work in text-to-image prompts and in edits.

• Up to 4 references per generation, plus @frame in an edit. Only references you mention are used and charged.
• A reference belongs to a sequence. Add one by uploading it, or by copying any image you own (MCP add_reference, REST POST /projects/:id/references).
• Labels are 1 to 24 characters: lowercase letters, digits, - and _.

What makes a reference carry over #

Say exactly what to take from it. Name the thing and describe it:

Give the gnome in @frame the exact same wizard hat shown in @hat, tall and pointed, midnight blue with silver stars.

The hat pictured in @hat on its own tends to give a generic hat.

Give each image a role. Which is the scene that stays, and what comes from each reference:

Place the woman from @anna on the bench in @frame, wearing the red raincoat shown in @coat. Keep the park, the bench and the lighting of @frame.

Crop busy references to the object. A screenshot or a crowded photo used whole often gets copied in wholesale. Crop it to the item (MCP add_reference and update_reference take a crop; in the web app use the crop tool).

For a likeness, point at the image. The same man as in @sam preserves a face better than describing his features in words, which makes the model redraw him from the description.

Adding to a picture you have beats generating from scratch. If you need a referenced object in a specific scene, generate or upload the scene first and then edit the object in, rather than generating the whole scene with the reference.

Things that do not help: quality suffixes, negative prompts, a generic reproduce it exactly sentence. Small references are fine; they are scaled up internally.

Consistent characters across a film #

Make one clean, well-lit image of the character, add it as a reference (for example @mira), and mention it wherever she must appear in a new start frame. Inside a sequence, continuity comes for free: each step starts from the previous take's last frame.

When an edit goes wrong #

Each attempt is a new candidate; nothing is overwritten. Try again with a sharper instruction and select the best.