Text
Screenplay PDF
Industry-format LaTeX screenplay compiled from Fountain-structured scenes — ready for readers and producers.
Research Project · June 2026
Your book.
A full screenplay.
A narrated film.
A hierarchical pipeline of local AI agents turns a PDF novel into industry-format screenplay artifacts and a 4K picture-video — entirely on your machine, with no cloud APIs.
120min
Feature runtime
~70
Scenes generated
4K
Video output
100%
Local & private
Architecture
Eleven stages. One pipeline.
Specialized local agents handle global structure, parallel scene writing, and multimedia synthesis — with checkpointing at every step.
01
Extract
PDF → structured text chunks
02
Director
Story bible & 3-act beat sheet
03
Scene Planner
~70 scene skeletons with dramatic goals
04
Writers × N
Parallel Fountain-structured scenes
05
Continuity
Cross-scene consistency patches
06
Polish
Dialogue & action refinement
07
Screenplay
Fountain, LaTeX & PDF output
08
Images
Cinematic still per scene + portraits
09
Narration
Per-character voice synthesis
10
4K Video
Ken Burns, captions & chapters
Deliverables
What the pipeline produces
Every run yields a complete adaptation package — from formatted screenplay to a watchable narrated film, without sending your manuscript to the cloud.
Structure
Story Bible
Characters with voice & arc, locations, themes, and a 24–40 beat three-act treatment distilled from the source book.
Visual
Scene Stills
One cinematic 1920×1080 still per scene, plus character portrait thumbnails for rich-caption mode.
Audio
Narrated Audio
Per-line WAV tracks with distinct macOS voices assigned to each speaking character in reading order.
Video
4K Picture-Video
3840×2160 MP4 with Ken Burns motion, per-character coloured captions, slug cards, act breaks, and chapter metadata.
Accessibility
Subtitle Sidecar
SRT file emitted alongside every render — usable independently of on-video overlay presets.
screenplay.pdf — sample excerpt
Adaptation Draft
Int. Study — Night
Rain hammers the window. A manuscript lies open, pages curling at the edges. The lamp throws a warm cone of light across the desk.
Protagonist
(quietly, to the empty room)
Every story wants to become something else.
They close the book. On the monitor, a progress bar inches forward.
· · · Generated scene — illustrative preview · · ·
Operational Benchmarks
Built for real hardware
Observed on a 96 GB Mac Studio. End-to-end wall-clock for a full two-hour pipeline: several hours — dominated by LLM generation and 4K ffmpeg encodes.
Target runtime
120 min
≈120 script pages, ~70 scenes
Peak memory
~48 GB
96 GB Mac Studio, throttled cap
Text model
14B params
Director, planner, continuity
Writer model
8B params
3 parallel scene agents
Beats
24–40
Three-act structure
Disk (2 h run)
≥30 GB
Intermediates + final outputs
| Ollama instances | Peak RAM | Notes |
|---|---|---|
| 1 | ~30 GB | Single server, model swapping |
| 2default | ~48 GB | Default — text + image servers |
| 4 | ~80 GB | Higher writer throughput |
Key Findings
What we learned
Role-specialized model sizing
14B models handle global structure; 8B writers run in parallel with half the KV-cache footprint per slot.
Explicit memory orchestration
Text models are evicted before image generation; image models before ffmpeg — reclaiming ~10 GB between stages.
Checkpointed resumability
Every stage caches under work/. Crashes and Ctrl-C are safe; re-runs skip completed steps.
Robust JSON recovery
Lenient parsing with auto-close handles truncated LLM output — critical for long scene-plan responses.
Best-effort continuity
Parallel writers drift; the continuity pass fixes naming and anachronisms but cannot invent plot facts absent from the bible.
Platform constraints
macOS-only TTS today; Ollama image API shapes still evolving; voice availability varies by install.
Research Paper
Read the full paper
A hierarchical multi-agent pipeline for local book-to-screenplay and narrated video adaptation — architecture, benchmarks, and future work.
Download PDFbook2screenplay: A Hierarchical Multi-Agent Pipeline for Local Book-to-Screenplay and Narrated Video Adaptation