Docker

The repo ships a Dockerfile and a compose.yaml as an alternative to the local Poetry setup. The stack is fully containerized: the image bundles Python 3.13, FFmpeg, DejaVu fonts and all Python dependencies, and Ollama runs as a compose service — nothing to install on the host besides Docker itself.

The local workflow (poetry install + python -m shortificator) keeps working as-is; Docker is optional.

Requirements

• Docker (Docker Desktop or docker-ce)
• NVIDIA Container Toolkit so the containers can use the GPU (transcription and LLM inference)

Build and verify

make docker-build
make docker-pull MODEL=mistral-small   # pull an LLM into the Ollama service
make docker-check-env

docker-check-env runs inside the container and checks CUDA visibility (CTranslate2) and the Ollama connection, listing the pulled models.

Run

Through the Makefile, mirroring the local make run variables:

make docker-run INPUT=my_video.mp4 MODEL=mistral-small MAX=5 DYNAMIC=1

Or with Docker Compose directly — anything after the service name is passed to the shortificator CLI:

docker compose run --rm shortificator \
  --input /videos/my_video.mp4 \
  --model mistral-small \
  --max-shorts 5 \
  --dynamic-subtitles

The Ollama service is started automatically when needed. Stop it when you are done:

make docker-down

Volumes and paths

Host	Container	Purpose
./ (or VIDEOS_DIR)	/videos	input videos (read by --input /videos/...)
./output	/app/output	generated Shorts, transcripts, candidates
./models	/app/models	YuNet face detection model (auto-downloaded)
whisper-cache (named volume)	/root/.cache/huggingface	Whisper model (~3 GB for large-v3), downloaded once
ollama-models (named volume)	/root/.ollama	LLMs pulled with make docker-pull, kept across runs

Set VIDEOS_DIR=/path/to/videos to mount a different input directory:

VIDEOS_DIR=~/videos make docker-run INPUT=gameplay.mp4

Reusing an Ollama on the host

If you already run Ollama on the host and prefer it over the bundled service, point the container at it:

OLLAMA_HOST=http://host.docker.internal:11434 make docker-run INPUT=my_video.mp4

Host Ollama must listen beyond loopback

By default Ollama binds to 127.0.0.1, which containers cannot reach — docker-check-env will report connection refused. Make it listen on all interfaces:

# one-off
OLLAMA_HOST=0.0.0.0 ollama serve

Or persistently for the systemd service:

sudo systemctl edit ollama
# add:
#   [Service]
#   Environment="OLLAMA_HOST=0.0.0.0"
sudo systemctl restart ollama

host.docker.internal is resolved on Linux/WSL2 via the extra_hosts: host-gateway entry in compose.yaml; on Docker Desktop it works out of the box.