video-generator

Define each recurring character with a unique identifier and appearance description, then generate a primary reference image before any keyframes. Use that reference for all additional angles and pass it when generating first keyframes so the character stays consistent across clips.

Segment the video into 3–10 second clips with one action and one scene each. For every clip specify narrative purpose, pacing, a detailed 2–4 sentence transition description, camera movement, framing, and audio cues before generating anything.

All first keyframes must use 16:9 or 9:16 and be upright, never 1:1 or other ratios. This applies equally to frames extracted from a previous clip's rendered output when reusing continuity between clips.

No, never use TTS for on-screen dialogue or singing. The video model generates that audio with lip sync directly. TTS is reserved for off-screen narration, generated per narration span with a consistent voice profile.

Audio tracks get lost when one track replaces another during assembly. The mixing rules require overlaying all tracks—video audio, narration, and separate BGM—while keeping narration clearly audible at consistent volume across clips.

Use embedded BGM for music videos or diegetic music with a visible source, described directly in the video prompt. Use separate generation for background mood music, driven by a second-by-second emotional arc blueprint built during clip planning.