video-forge

Define the video type, target duration, tone, audience, and voiceover plan first, then freeze a storyboard and video-spec JSON. The pipeline dubs the full script in one TTS pass, aligns timestamps with a forced aligner, renders a preview, and runs review gates before final export.

Generate the voiceover from the complete script in a single pass, then use a forced alignment tool like Qwen3-ForcedAligner or WhisperX to derive word-level timestamps. Do not rely on native TTS timestamps, which are unreliable for segment timing.

Yes, generate each segment separately within the API duration limit, download the clips, and concatenate them with FFmpeg using concat or xfade. Multi-segment outputs claiming to be one continuous shot must pass a seam review covering character, scene, color temperature, camera direction, and motion continuity.

Follow the retiming priority: trim excess footage first, then freeze the last frame with styling, insert B-roll, or apply moderate slow motion no slower than 0.7x. Never use aggressive slow motion below 0.7x to fill large time gaps.

Static frames with only camera pans, zooms, or transitions count as packaging, not motion. Each major narrative beat needs real on-screen change such as recorded UI interactions or animated state transitions; otherwise the preview is blocked until genuine motion footage is captured.