blender-cinematic-video-factory

Files (34)

• .env.example
• CHANGELOG.md
• LICENSE.md
• MANIFEST.sha256
• PRODUCT-SHEET.md
• QUICKSTART.md
• README.md
• SKILL.md
• SUPPORT.md
• THIRD_PARTY_NOTICES.md
• VERSION
• openai.yaml · agents/
• variant_comparison.png · assets/
• install.sh
• manifest.json
• README.md · pipeline/
• config.schema.json · pipeline/
• amber_gallery.json · pipeline/configs/
• base.json · pipeline/configs/
• cobalt_noir.json · pipeline/configs/
• jade_moon.json · pipeline/configs/
• ruby_rose.json · pipeline/configs/
• violet_silver.json · pipeline/configs/
• generate_variants.py · pipeline/
• pipeline.py · pipeline/
• physics.md · references/
• sound-design.md · references/
• requirements.txt
• coherent_fluid_cut_concept.py · scripts/
• coherent_fluid_cut_event_export.py · scripts/
• coherent_fluid_cut_production.py · scripts/
• coherent_fluid_cut_sound_design.py · scripts/
• render_scene.sh · scripts/
• validate_package.py · tools/

SKILL.md (first lines)

# Blender Cinematic Video Factory

Create deterministic 9:16 Blender videos from JSON configurations. The supported production template combines a curved glass vessel, a precision cut, Mantaflow liquid overflow, controlled rigid-body petals, and event-synchronized procedural sound.

## Mandatory Workflow

1. Run `python3 pipeline/pipeline.py doctor` and resolve every required check.
2. Run `plan` to inspect cache identity, output location, render profile, and selected encoder.
3. Run `concept` and inspect all four storyboard frames.
4. Continue with `preview --approve-concept` only after explicit concept approval.
5. Inspect all six preview frames, including cut, spill, settle, and glass-release states.
6. Continue with `render --approve-concept --approve-preview` only after explicit motion approval.
7. Run `final` to encode, export evaluated events, generate sound, mux, and execute QA.
8. Report the verified MP4 path and measured loudness/true peak.

Never bypass the approval gates or start a full render from an unapproved concept.

## Production Defaults

- 1080x1920, 30 fps, 450 frames, 15 seconds
- Blender Eevee with 64 final samples
- 720x1280 preview, 16 samples, ray tracing disabled
- APIC Mantaflow with cache manifest and per-hash bake lock
- H.264 `yuv420p`; automatic NVENC selection with `libx264` fallback
- AAC stereo, 48 kHz, event-driven procedural sound
- Black-frame, decode, frame-count, codec, pixel-format, loudness, true-peak, and faststart QA

## Physics Rules

- Render, material, camera, light, audio, and encoding changes must reuse the physics cache.
- Geometry, timing, source, viscosity, solver, mesh, or fluid-resolution changes must produce a new cache signature.
- Never run two first-time bakes for the same signature concurrently.
- Preserve sequential frame evaluation when exporting rigid-body or fluid events.
- Treat fluid contact timing as mesh-sampled rather than per-particle exact.

## Performance Rules

- Use the preview profile for approvals; do not reduce final quality before reviewing preview results.
- Mantaflow and rigid bodies are primarily CPU-bound. Eevee and NVENC use the GPU.
- On one GPU, begin with one render process. Benchmark two processes before increasing concurrency.
- Keep PNG masters until final QA succeeds.
- Use `codec: auto` for portable delivery. Force `h264_nvenc` only on validated NVIDIA systems.

## Supported Files

- `pipeline/pipeline.py`: doctor, planning, approval-gated rendering, cache management, encoding, and QA.
- `pipeline/generate_variants.py`: deterministic visual/audio variant generation without an LLM.
- `pipeline/configs/`: one base template and five production examples.
- `scripts/coherent_fluid_cut_concept.py`: four-frame concept storyboard.
- `scripts/coherent_fluid_cut_production.py`: Mantaflow and rigid-body production scene.
- `scripts/coherent_fluid_cut_event_export.py`: evaluated fluid and collision event export.
- `scripts/coherent_fluid_cut_sound_design.py`: synchronized procedural audio renderer.
- `references/physics.md`: simulation guidance.
- `references/sound-design.md`: event-to-audio guidance.

## Commands

Run from the installed skill directory:

```bash
python3 pipeline/pipeline.py doctor
python3 pipeline/pipeline.py plan pipeline/configs/cobalt_noir.json
python3 pipeline/pipeline.py concept pipeline/configs/cobalt_noir.json
python3 pipeline/pipeline.py preview pipeline/configs/cobalt_noir.json --approve-concept
python3 pipeline/pipeline.py render pipeline/configs/cobalt_noir.json --approve-concept --approve-preview
python3 pipeline/pipeline.py final pipeline/configs/cobalt_noir.json
```

Read `QUICKSTART.md` for installation and `pipeline/README.md` for every configurable performance field.

From ruota/open-sinapseed, MIT.

More from ruota/open-sinapseed