Glossary

The direktor vocabulary

The words that show up across the docs, the CLI output and this site — defined once, plainly.

Pipeline
The end-to-end chain of six stages direktor runs to turn text into video. Each stage has one job and one output file.
Stage
A single step in the pipeline (script, narration, transcript, prompts, images, compose). Stages run in order and checkpoint independently.
Checkpoint
The on-disk output a stage writes to a temp directory keyed off your input file. Checkpoints are what make runs resumable.
Resume
Re-running direktor after a stop or failure. Finished stages are detected from their checkpoints and skipped.
Script
The podcast-style, single-narrator text GPT-4 produces from your input in stage 1. Everything downstream is faithful to it.
Narration
The spoken audio track synthesised from the script by BARK in stage 2, stored as narration.wav.
Transcript
The timestamped, chunked text Distil-Whisper produces from the narration in stage 3. Timestamps drive image timing.
Segment
A ~30-second window of the transcript. One image prompt and one still are produced per segment.
Image prompt
The text description GPT-4 writes for each segment in stage 4, stored in image_prompts.json — the most useful file to hand-edit.
Still
A single 16:9 image FLUX-schnell renders from an image prompt in stage 5. Stills are held for their segment duration in the final video.
Composition
Stage 6: FFmpeg stitching stills and narration (plus optional overlays) into the final 1920x1080 MP4.
Overlay
Optional keyword text drawn onto a still by FFmpeg drawtext during composition.
BARK
The text-to-speech model (suno-ai/bark on Replicate) direktor uses for narration by default. Overridable via BARK_MODEL.
Distil-Whisper
A distilled Whisper speech-recognition model used for transcription. Overridable via DISTIL_MODEL.
FLUX-schnell
The fast FLUX image model (black-forest-labs/flux-schnell on Replicate) direktor uses for stills. Overridable via FLUX_MODEL.
FFmpeg
The local command-line media tool direktor shells out to for the final composition. Must be installed and on your PATH.
Replicate
The hosted model-running platform direktor calls for narration, transcription and image generation.
S3-compatible bucket
Object storage (Cloudflare R2, Backblaze B2, etc.) that holds the intermediate narration audio so the transcription model can fetch it.