Movilizer Generative Movie Studio

A local-first, server-capable repository for movie-grade generative production:

• SDXL identity/style LoRA training,
• batched multi-project image+video generation,
• dialog voice + music + mix + final MP4 packaging,
• modular judges and self-evolving optimization loops,
• non-destructive scene/shot patch tweaks,
• Slurm jobs (2x80GB GPUs),
• natural-language AI command interface.

What is available now

1. Local debug mode

• Fast iteration profile via configs/run/local_debug.yaml.
• Use low resolution/fps/steps while preserving pipeline behavior.

2. Server mode

• studio.server FastAPI app with endpoints for:
  • run orchestration,
  • AI plan/execute,
  • model list/pull/push.

3. Model retrieval and publishing

• Pull from local path or Hugging Face repo into local model cache.
• Push from local model folder to local path or Hugging Face repo.
• CLI: python -m studio.model_registry ...

4. Generic movie project baseline

• projects/feature_film_demo is the default movie-oriented sample.
• Previous sample (my_makeover) remains available as legacy reference.

5. Multi prompt-image/video references

• Supported across workspace/project/scene/shot levels.
• Also supported via shot.references.prompt_images[] and shot.references.prompt_videos[].

Local setup (conda)

conda env create -f environment.yml
conda activate movilizer
pip install -e .

Quickstart

1) Local debug run

python -m studio.run \
  --workspace workspace.yaml \
  --run_config configs/run/local_debug.yaml \
  --projects feature_film_demo \
  --scenes all \
  --resume

2) Full-quality run profile

python -m studio.run --workspace workspace.yaml --projects feature_film_demo --scenes all --resume

3) Compile only

python -m studio.run --workspace workspace.yaml --project feature_film_demo --scene scene_001_opening --compile_only

4) AI command prompt interface

python -m studio.ai --workspace workspace.yaml --project feature_film_demo --scene scene_001_opening \
  "make shot_001 more cinematic and regenerate frames 48-120"

5) Model registry sync

# Pull model from HF to local cache
python -m studio.model_registry pull --source stabilityai/stable-diffusion-xl-base-1.0

# Push local model folder to HF repo
python -m studio.model_registry push --source_dir models/cache/stabilityai__stable-diffusion-xl-base-1.0 --target your-org/your-model

# Push local model folder to local destination
python -m studio.model_registry push --source_dir models/cache/stabilityai__stable-diffusion-xl-base-1.0 --target /tmp/model_export

6) Server mode

python -m studio.server --config configs/server/default.yaml

Example requests:

curl http://127.0.0.1:8080/health

curl -X POST http://127.0.0.1:8080/models/pull \
  -H 'content-type: application/json' \
  -d '{"source":"stabilityai/stable-diffusion-xl-base-1.0"}'

curl -X POST http://127.0.0.1:8080/run \
  -H 'content-type: application/json' \
  -d '{"projects":["feature_film_demo"],"scenes":["scene_001_opening"],"run_config":"configs/run/local_debug.yaml"}'

PR artifact video (automatic)

Every pull request now generates a sample MP4 and uploads it as a GitHub Actions artifact:

• Workflow: .github/workflows/pr-sample-video-artifact.yml
• Artifact name: sample-prototype-video
• File: artifacts/sample_prototype_pr.mp4

How to view/download:

1. Open the PR on GitHub.
2. Open the Checks tab and click PR Sample Video Artifact workflow run.
3. Download sample-prototype-video from the run's Artifacts section.

Output layout

• Shot: outputs/<run_id>/<project>/<scene>/<shot>/...
• Scene: outputs/<run_id>/<project>/<scene>/scene.mp4
• Project final: outputs/<run_id>/<project>/final.mp4
• Eval: outputs/eval/<run_id>/...

Slurm

• slurm/train_identity_2gpu.sbatch
• slurm/run_studio_2gpu.sbatch
• slurm/eval.sbatch
• slurm/evolve.sbatch

Documentation

Human docs:

• docs/HUMAN/PROJECTS.md
• docs/HUMAN/SCENES.md
• docs/HUMAN/TWEAKS.md
• docs/HUMAN/PIPELINE.md
• docs/HUMAN/SLURM.md
• docs/HUMAN/SERVER.md
• docs/HUMAN/AUDIO.md
• docs/HUMAN/VIDEO.md
• docs/HUMAN/JUDGES.md
• docs/HUMAN/TROUBLESHOOTING.md
• docs/HUMAN/SAFETY.md

AI docs:

• docs/AI/*.schema.json
• docs/AI/PLAYBOOK.md
• docs/AI/COMMANDS.md
• docs/AI/EXAMPLES/

Graceful fallback behavior

When optional components are unavailable:

• generation uses synthetic frame fallback,
• training uses mock LoRA artifact fallback,
• advanced judges stay heuristic,
• server/model sync endpoints return actionable failure messages.