AI scene generation for interactive stories

In a Pominis story, each scene can carry a background image, and that image does real work: it tells the player where they are, who is present, and what the mood is before anyone speaks. This guide covers the three tools you have for that job, in the order worth reaching for them: generated scene images, plain-language edits, and short video moments.

The running example is an invented scene from a small mystery: a tram depot at dawn, where a night-shift driver finds the last tram of the night parked with its doors open and nobody aboard.

Where scene images come from

After the agent writes a chapter's dialogue scenes, it prepares an image prompt for each scene and generates the images in the project's image style, framed for the canvas target you chose: 16:9 for PC landscape, 9:16 for Mobile portrait. When a character appears in a scene, their portrait or design sheet goes to the image model as an identity reference, which helps the same face and outfit carry across locations.

You can steer this before and after. Before: describe the fixed features of recurring locations in your first message, or in Creator notes, the standing instructions the agent reads on every turn. After: open the scene with Edit scene and use the Scene Media section to generate a new image from your own prompt, refresh the visuals, or upload an image you made elsewhere. Typing @ in the prompt pulls in images already in the project as references.

Describe the place once, then the change

• Fixed: a brick tram depot, three parallel tracks, green-painted trams, a clock above the exit, sodium lamps still lit.
• Time and light: the first gray light through a high window, the lamps not yet switched off.
• This scene only: one tram at the platform with both doors open and every seat empty, the driver in the foreground with her back to the camera.

Keeping those layers apart makes the next request easier. The same depot at midnight, for an earlier scene, changes one line instead of the whole description.

Fix an image with a plain-language edit

When an image is almost right, do not regenerate it from scratch. Ask the agent for an edit in ordinary words: “In the depot scene, open the tram doors.” “Same shot, but at night with the lamps on.” “Remove the second driver standing under the clock.” The edit sends the current image to the image model with your change, keeps everything you did not mention, and replaces the scene image; the previous version is kept in the project.

Regenerate instead when the composition itself is wrong, for example when the camera is on the wrong side of the platform. An edit is for a detail; a new image is for a new shot. Scene edits also carry the characters' identity references, so fixing a lamp does not quietly change the driver's face.

Motion first, video when it earns it

Many scenes need a little movement, not a clip. Scene Motion applies an effect to the still image in the player: camera moves such as Slow push-in or Drift left, particles such as Rain, Snow, or Fireflies, and accents such as Signal glitch or Memory glow. You can preview each one on the scene image before applying it. For the depot, Soft fog over the tracks is enough for most of the chapter.

Save real video for a moment the player should remember. In the scene editor, Turn to video creates a short clip from the scene image, and you can describe the motion you want, such as “the tram doors slowly close on their own.” Scene videos can be 5, 8, 10, or 15 seconds long, with 5 as the default, at up to 768p. The clips have no audio of their own, and there is no option to build a clip from reference images or to edit an existing video. A scene with a video plays the video; its motion setting only applies if the player falls back to the still image. Phone chat scenes are not video targets.

What it costs

Image generation, image edits, and scene videos all spend credits, and richer media and higher image-quality settings use more of them. Unless you switch the agent to Full auto, it pauses for your approval before credit-spending steps. A good habit is to finish the dialogue and play the route before polishing visuals, then spend on the handful of scenes that carry the story's turns.

Check each scene in the player

• Does the first image of each location tell the player where they are?
• Does a recurring character still read as the same person after the location changes?
• Is the important part of the frame visible behind the dialogue box?
• Does the video or motion effect support the line on screen, or compete with it?
• Is the image readable at the screen size your readers will use?