This document discusses using generative AI models for creative work such as generating movie scripts and content. It describes several generative models including restricted Boltzmann machines, deep generative models, variational autoencoders, and hierarchical LSTMs. It proposes using these models to help optimize the movie production process by automatically generating schedules, budgets, storyboards, and dialog from input scripts to help filmmakers explore creative options. An example project of a movie conceived by humans but with AI-generated dialog is briefly described.
2. Generative Models
Model the Joint Probability
of Inputs and Outputs
Can generate both Inputs and Outputs given hidden paramete
Restricted Boltzmann Machine:
Learns probability distribution ov
set of inputs.
7. Long-Short Term Memory Models
RNNs Cannot encode
long-range dependencies
LSTMs encode long-range
dependencies with gated
memory
8. Maximum Likelihood
The _ quick _ brown _ fox
Learn conditional probabilities: P (fox | “The quick brown
Weight matrix W is large. ~ 100,000 tokens per corpus.
Generate text: Maximize likelihood of generated sentences
13. PreProduction : Story Board
Text -> Image Classifier
Script -> Suggested Shot; Find image in archive that best match
14. PreProduction : Story Board
Trained on: Scripts -> Film Images (canonical shots in scenes)
Archival footage can be used for training algorithms. Licensed for Storyboar
16. Variables that impact the optimal schedule:
• Cast schedules
• Location availability
• Equipment and Props (e.g. vehicles)
• Weather
• Cost and budget constraints
AI to solve the combinatorial optimization problem:
We realized up to 20% cost and time efficiencies. Data: Run optimization re
PreProduction : Scheduler
19. Ingest Script Automatically generate a Story Interaction Graph
Agile StoryWriter
20. Remove Node: Action Sequence was adding significantly to budget
Agile StoryWriter
Butterfly Effects: What other nodes are impacted from the change?
21. Explore What If Scenarios
Agile Content Creation
Script
Greenlighting
Decision
Pre-Production
Financing
Forecasting
Schedules,
Budget
Distribution
PredictionsIterations
Eliminate
Scenes
22. Language Models
Why is language hard? Discrete space.
Cats are good pets
Bats are good pets
Images & Sound : Continuous space
Gradient search is not effective for language learning.
24. Generative Adversarial
Networks
Drawbacks:
- Cannot control what to generate (some advances with Conditional GANs)
- Cannot generate categorical data
- Cannot access latent features
Not very successful in discrete space e.g. text generation
- Learning converges to a specific part of the distribution
26. Variational Autoencoders
Calculating is infeasible, so we use variational lower
KL-divergence with the variational posterior
A normal distribution
parameterized by a
Deep Neural Network.
30. AI Generated Content
Assistant Writer, Assistant Producer & Assistant Director
• Writer sets up the context. AI generates the narrative and dialogues.
• Producer enters ATL parameters. AI generates budget, schedule.
• AI recommends Shot Lists and Storyboards.
32. Thank You!
• AI team @ End Cue
• Walter Kortschak and Andrew Kortschak @ End Cue.
• We are Hiring! Email me at deb@endcue.com
• Production Team at End Cue
• Production Team at GreenCard Pictures
Nathan Crockett