Frequently asked questions
The honest answers on what direktor is, what it is not, and how to run it.
What exactly does direktor produce?
A single 1920x1080 MP4: a narration track with AI-generated still images that change at transcript-segment boundaries, plus optional keyword overlays. It is a podcast-style narrated video, not a motion film.
Does direktor generate motion video?
No. Stages 5 and 6 produce still images from FLUX-schnell and compose them under the audio with FFmpeg. There is no per-frame video model, no camera motion, no lipsync. If you need genuine motion, direktor is the wrong tool.
What are the prerequisites?
Python 3.11 or newer, FFmpeg on your PATH, an OpenAI API key, a Replicate API token, and an S3-compatible bucket (Cloudflare R2, Backblaze B2) for the intermediate narration audio.
How is it licensed and distributed?
direktor is MIT-licensed, published on PyPI as the direktor package, and developed in the open at github.com/Skelf-Research/direktor.
Can I use different models?
Yes. The model for each stage is chosen by an environment variable: GPT4_MODEL, BARK_MODEL, DISTIL_MODEL and FLUX_MODEL. Swapping the TTS, image model or LLM is a config change, not a code change.
What happens if a stage fails?
Each stage checkpoints its output to a temp directory keyed off your input file. If stage 5 fails, stages 1 to 4 are already on disk — re-running skips them and resumes from the failure.
Can I edit the script or image prompts?
Yes, and this is the point. Stop after any stage with --stage N, edit the script or image_prompts.json by hand, then resume. Only the stages after your edit re-run.
How long can the output video be?
direktor is built for long-form: roughly 3 to 20 minutes of narrated content. Length follows the length of your input text and the resulting narration.
Where does my data and spend go?
direktor runs locally and calls your own OpenAI and Replicate accounts and your own bucket. There is no direktor SaaS in the middle — your keys, your spend, your storage.
Is it wired to Sora, Veo or Runway?
No. The default image model is FLUX-schnell on Replicate. direktor is a pipeline you assemble from your own model accounts, not a wrapper around a frontier video model.
Can I run it in CI?
Yes. The CLI drops into a Makefile or a GitHub Action, and the Python API (generate_video()) lets you embed the pipeline in a larger program.
How do keyword overlays work?
During composition, FFmpeg drawtext can render key terms from each transcript segment onto the corresponding still. Overlays are optional and toggled per run.