AI Video Generation

Four model families do the work and they run in order, not all at once. A language model drafts the script and splits it into scenes. An image or video model renders frames for each scene. A speech model reads the narration aloud. A music model writes a track sized to the finished runtime. An assembly pass burns in the captions and cuts everything to length.

The sequence is what separates a usable tool from a demo. The script decides the scene count, the scene count decides how much footage gets generated, and the narration length decides how long each shot has to hold on screen. Break that chain anywhere and you get a video that ends on five seconds of silence over a still image.

Cost follows the same logic. Orange prices the visual half by target runtime, plus a small flat fee covering the script, the standard voice, the music, and the render. A 30-second video costs the same whether the writer returned four scenes or six. Per-scene pricing used to punish a longer outline. Charging by duration does not, and the exact rates sit on the pricing page, quoted before any render starts.

The limits are worth knowing before you plan a series around this. Character continuity across scenes is unreliable, so the same person in scene one and scene five can come back with a different jaw. Text rendered inside a generated image is a coin flip on most models, which is why titles get composited afterward instead of prompted in. Anything that has to show a specific real product or a real building needs a reference image rather than a description.

Where it earns its keep is volume. One idea, ten variants, all rendering while you write the next batch.