ContentGeneration Pipeline

This project runs a 3-step video pipeline:

1. Generate shot videos from images + prompts.
2. Merge each generated video with its audio.
3. Concatenate merged clips into one final output.

The pipeline entrypoint is run_video_pipeline.py.

Quick Start

Local Python:

cp .env.example .env
uv sync --dev
uv run python run_video_pipeline.py

Docker (GPU):

cp .env.example .env
docker build -t content-generation:latest .
docker run --rm --gpus all --env-file .env \
  -v "$(pwd)":/app \
  -v "$HOME/.cache/huggingface":/root/.cache/huggingface \
  -w /app content-generation:latest

First run (skip S3 upload):

python run_video_pipeline.py --skip-s3-upload

Docker first run (skip S3 upload):

docker run --rm --gpus all --env-file .env \
  -v "$(pwd)":/app \
  -v "$HOME/.cache/huggingface":/root/.cache/huggingface \
  -w /app \
  content-generation:latest \
  python run_video_pipeline.py --skip-s3-upload

Project Layout

• run_video_pipeline.py: main entrypoint.
• src/: helper scripts used by the pipeline.
• HunyuanVideo-1.5/: Hunyuan inference code and model dependencies.
• reel_script.json: required script input with shots.
• images/, audios/, videos/, merged/, results/: working/output folders.
• .env.example: environment variable template.

Prerequisites

1. Linux with NVIDIA GPU and CUDA runtime.
2. ffmpeg and ffprobe available on PATH.
3. Python 3.10+.
4. uv installed (https://docs.astral.sh/uv/).
5. Hunyuan model checkpoints under HunyuanVideo-1.5/ckpts.
6. If using FLUX local download, access approved for black-forest-labs/FLUX.1-schnell.

Environment Variables

1. Create local env file:

cp .env.example .env

2. Fill required variables in .env:

• ELEVENLABS_API_KEY for audio generation.
• HUGGINGFACE_HUB_TOKEN if gated Hugging Face model access is needed.
• AWS_S3_BUCKET (+ optional AWS vars) if you want final output uploaded to S3.

Run Locally (Python)

1. Create and activate a virtual environment:

uv venv
source .venv/bin/activate

2. Install Python dependencies:

uv sync --dev

3. Install Hunyuan dependencies:

uv pip install -r HunyuanVideo-1.5/requirements.txt
uv pip install --upgrade tencentcloud-sdk-python
uv pip install sgl-kernel==0.3.18

4. Run full pipeline:

uv run python run_video_pipeline.py

5. Common options:

# Skip generation and only merge + concat
python run_video_pipeline.py --skip-generate

# Skip S3 upload
python run_video_pipeline.py --skip-s3-upload

# Override base directory
python run_video_pipeline.py --base-dir /absolute/path/to/workdir

# Change logging verbosity
python run_video_pipeline.py --log-level DEBUG

Run with Docker

1. Build image:

docker build -t content-generation:latest .

2. Optional build with extra attention backends:

docker build -t content-generation:latest --build-arg INSTALL_OPTIONAL_ATTENTION=1 .

3. Run pipeline in container (GPU required):

docker run --rm --gpus all \
  --env-file .env \
  -v "$(pwd)":/app \
  -v "$HOME/.cache/huggingface":/root/.cache/huggingface \
  -w /app \
  content-generation:latest

4. Pass extra pipeline args:

docker run --rm --gpus all \
  --env-file .env \
  -v "$(pwd)":/app \
  -v "$HOME/.cache/huggingface":/root/.cache/huggingface \
  -w /app \
  content-generation:latest \
  python run_video_pipeline.py --skip-s3-upload --log-level DEBUG

Input Expectations

1. reel_script.json must exist and contain a shots array.
2. images/shot_<n>.png and audios/output_<n>.mp3 should align by shot number.
3. Final output is written by default to results/final_output.mp4.

S3 Upload Behavior

1. If AWS_S3_BUCKET is set, the pipeline uploads final output to S3 using S3VideoStorage.
2. If AWS_S3_BUCKET is missing, upload is skipped with a warning.
3. Disable upload explicitly with --skip-s3-upload.

Troubleshooting

1. torch.cuda.is_available() is false in Docker.

• Run with GPU flags: docker run --gpus all ...
• Verify NVIDIA Container Toolkit is installed on host.
• Check host GPU visibility: nvidia-smi.

2. ffmpeg or ffprobe not found.

• Local: install ffmpeg with your package manager.
• Docker: ffmpeg is installed in the provided Dockerfile.

3. Hunyuan generate step fails due to missing checkpoints.

• Ensure checkpoints are available under HunyuanVideo-1.5/ckpts.
• Confirm mounted project path in Docker includes checkpoints.

4. Hugging Face model download fails (401/403).

• Accept model access terms for gated models (for example FLUX.1-schnell).
• Set HUGGINGFACE_HUB_TOKEN in .env.

5. S3 upload fails.

• Confirm AWS_S3_BUCKET is set.
• If needed, set AWS_REGION and credentials (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, optional AWS_SESSION_TOKEN).
• For S3-compatible providers, set AWS_S3_ENDPOINT_URL.

6. Permission issues when running Docker with mounted volumes.

• Use your host user mapping if needed: docker run --rm --gpus all -u "$(id -u):$(id -g)" ...

7. Out-of-memory during video generation.

• Keep PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:128.
• Reduce workload by skipping optional enhancements or lowering resolution/steps in generation scripts.

8. Verify syntax quickly before running.

uv run python -m py_compile run_video_pipeline.py src/*.py

Testing

Run tests with coverage:

uv run pytest