Everything direktor does
direktor is deliberately small: a resumable pipeline you can read, edit and script. Here is the full surface, grouped by what it gives you.
Pipeline architecture
Six discrete stages
script → narration → transcript → prompts → images → compose. Each stage is an isolated step with a defined input and output.
Per-stage checkpointing
Every stage writes its result to a temp directory keyed off your input file. Nothing gets recomputed unless you ask.
Resume after failure
A crash in stage 5 leaves stages 1–4 intact. Re-run and direktor skips the finished work and continues.
Stop at any stage
direktor build input.txt --stage 3 halts after the transcript, ready for you to inspect or edit before continuing.
Editability & control
Plain-file intermediates
The script, transcript, image_prompts.json and stills are ordinary files. Open them, diff them, commit them.
Hand-edit prompts
Regenerate a weak image by rewriting its prompt in image_prompts.json, then resume stage 5 for that segment.
Keyword overlays
Optional FFmpeg drawtext overlays surface key terms from each segment on the stills.
Deterministic assembly
Stage 6 is pure FFmpeg — the same intermediates always compose to the same MP4.
Model & provider flexibility
Env-var model selection
BARK_MODEL, FLUX_MODEL, GPT4_MODEL and DISTIL_MODEL choose the model per stage — no code edits.
OpenAI for language
GPT-4 handles both script generation and per-segment image prompting.
Replicate for media
BARK narration, Distil-Whisper transcription and FLUX-schnell stills all run on Replicate by default.
Your own storage
An S3-compatible bucket (Cloudflare R2, Backblaze B2) holds the intermediate audio that feeds transcription.
Developer experience
CLI first
direktor build input.txt is the whole interface. Flags control stages, overlays and resume behaviour.
Python API
generate_video() and per-module helpers let you embed direktor in a larger Python program.
CI-friendly
The CLI drops cleanly into a Makefile or a GitHub Action for repeatable, scripted video builds.
MIT licensed
Fork it, vendor it, wrap it. pip install direktor and the source is on GitHub.