Cinematic video

Produce a short video with intentional shot language: a chosen framing and lens, a named camera move, lighting, location, and a color grade. The lever is a compact cinematic prompt that names the director's decisions in camera vocabulary the model reads directly, not a wall of adjectives.

Inputs to collect

• The shot brief. What is in frame and what it should feel like. (Infer from the request when stated.)
• Text-to-video or image-to-video. If the user has a still that fixes the look, that is image-to-video and the prompt directs motion only. If not, it is text-to-video and the prompt builds the whole frame.
• Aspect and length. Landscape (16:9) or vertical (9:16), and roughly how long. Drives dimension pair and duration.
• Optional: a reference for a locked character or brand look, an explicit grade or era ("1950s", "teal-and-orange").

Models

Confirm the live model and its schema via runware-models + runware-run before calling. Never hardcode a stale choice. All three below carry io:text-to-video and io:image-to-video.

• Text-to-video with native audio: Google Veo 3.1 (google:3@2). Reads the five-element scaffold and camera vocabulary directly, fills scene detail from world knowledge, and generates synchronized audio natively. Best when the shot needs sound (ambience, foley) baked into the clip.
• Premium text-or-image-to-video: Luma Ray 3.2 (luma:ray@3.2). Strong cinematic motion and grade. Supports HDR output on 5-second clips.
• Image-to-video, locked look: Runway Gen-4.5 (runway:1@2). The source still fixes subject, composition, palette, and lighting, the prompt directs only the motion. Gen-4 Turbo is the cheaper draft tier in the same family.