How it works

Drop in a Reddit post, original fiction, or any narrative text. The editor accepts up to 15,000 characters (with a soft warning at 10,000) and auto-saves your draft every 5 seconds, so your work is never lost.

Longer stories produce more scenes and use more Drops. The AI works best with clear narrative structure — characters, dialogue, and a progression of events.

Claude reads your story and performs a deep analysis. It segments the narrative into cinematic scenes based on story length — shorter stories get 2–8 scenes, longer ones up to 20. It identifies every character by name and appearance, maps each unique location with a detailed visual description, extracts the emotional tone of each scene, and determines the visual genre.

Pass 1 outputs a structured breakdown: scene narration text, character descriptions with physical details, a location bible (canonical environment descriptions reused across scenes for visual consistency), per-scene emotion tags (from 14 recognized emotions), and a genre classification (from 10 genres like noir, sci-fi, fantasy) that tints the visual style of the entire video. Scene count scales with story length — short stories get 2–8 scenes, while longer stories (1,500+ words) can produce up to 20.

Nano Banana 2 generates a reference portrait for each character identified in Pass 1. These portraits capture the physical details described in your story and serve as the visual anchor for that character across every scene.

On Creator plans and above, all characters get reference portraits. Free and Starter plans include the protagonist only. The reference images are stored and passed to the scene illustrator so faces and features stay consistent.

Inworld TTS-1.5-Max generates expressive, emotion-aware speech for each scene using your chosen voice. The emotion tag from Pass 1 is injected directly into the TTS request, so a "tense" scene sounds different from a "whimsical" one. Choose from 12 curated voices — 4 available on Free, all 12 on paid plans.

The audio duration of each scene drives the visual timing — every clip is shown for exactly as long as its narration lasts. This means the final video has no dead air and no rushing.

A second Claude pass takes each scene and optimizes its image prompt. It adds camera angles, lighting direction, atmosphere, genre-tinted style cues, and a motion prompt describing how the still image should be animated. Nano Banana 2 then generates the final 9:16 scene image using the character reference portraits and the canonical location description for cross-scene environment consistency.

On Creator plans and above, a vision judge compares the generated scene against the character reference and scores face consistency and artifact quality to ensure the character looks right.

Each scene image is animated into a short video clip using the motion prompt from Pass 2. On Starter plans and above, Wan 2.2 image-to-video AI generates real motion — subtle head turns, fabric ripple, camera push-ins, and ambient environmental movement. On Free, a smooth Ken Burns zoom-pan effect adds motion to the still image.

The motion prompt is under 25 words and describes only observable physical movement — never narration or abstract concepts. This keeps the animation subtle, realistic, and compatible with the short clip length (~5–15 seconds per scene).