video-generator

Define each character's appearance and outfit in the global definitions phase, then generate a primary reference image followed by variation angles. Every clip's keyframe uses these reference images to maintain visual consistency across the video.

Segment the video into 3-10 second clips with one action and one scene each. Specify narrative purpose, pacing, camera movement, a detailed 2-4 sentence transition description, dialogue, sound effects, and narration cues per clip.

All first keyframes must use 16:9 or 9:16 aspect ratio in upright orientation. Square 1:1 or other ratios are not allowed, and frames extracted from previous clips must also match the chosen ratio.

Yes. Narration is generated via TTS per narration span using SSML guidance, while background music is generated separately from an emotional arc blueprint. All audio tracks are overlaid during mixing, never replaced.

On-screen dialogue and singing must be generated by the video model itself because it produces audio with lip sync. TTS is reserved only for off-screen narration where lip synchronization is not required.

Continuous clips require extracting the final frame from the previous clip's rendered video as the next keyframe, forcing sequential generation. Clips must also satisfy physical constraints, so impossible camera moves or distances must be split.