Cinemagic.AI

AI Powered Content Creation

Vision
Cinemagic.AI is a system that can predict movie success, measured by return on investment. It uses various factors to provide a detailed breakdown and evaluation of potential future success.

Cinemagic.AI will make for better investment decisions in the long term.

It seeks to reduce a level of uncertainty, resulting in more informed investment decisions.
Unfortunately, human mind is not able to comprehend the entirety of the world, with its analytical data and relationships to take into account; that’s the job for a machine.

Many people trust their gut, honed through years of experience; this gut feeling guides their decisions. We try to find a pattern in random events, which leads us further away from the right decision.
AL and ML in the movie industry and stock market prediction are cut from virtually the same cloth: you need to identify trends and patterns based on empirical data.

The movie industry is something that has a set of features we can use to look for patterns which have a close connection to reality.
How do we
reach it?
Stage 1 (research)
Status: completed
Stage 1 (application)
Status: WIP
6
Use MVP to discover a viable monetization strategy and development model
5
Build an MVP to launch it and map out our product offer
4
Research into the ways we can leverage models for new formats, languages and data sources
3
Write an effective case study which proves that the algorithm actually works
2
Test those models and identify the best-performing parameters
1
Do research and collect all models we can possibly find that might carry the day
Research Result
Research requirements
  • The U.S. movie market only
  • Feature movies only
  • English-language movies only
  • Movies that made it to theaters only
  • Movies with budget and box office data available only
  • We analyzed the IMDb Pro database and news datasets as sources
  • We created a structured repository of data to be collected, with data processing algorithms
  • We developed an algorithm for employing metrics to find out whether a decision is viable
Research process
Research result
ROI =
REVENUE – BUDGET
BUDGET
≥
a
where a means a parameter
  1. Numerical metrics:
Movie budget, box office receipts

2. Category metrics:
Movie genre, sub-genre

3. Rating metrics:
IMDb user reviews
4. Text metrics:
Title, slogan, synopsis

5. Draft metrics:
Relationship among the cast

6. News metrics:
Headlines and emotions behind the printed word
1
We designed a neural network to build a model, which will help predict the box-office receipts of a particular feature movie
To evaluate financial performance of a movie, the classifier uses the following formula of return on investment (ROI):
2
Analytically-wise, the classifier builds on six key metrics:
3
Measuring the Performance of the Neural Network
The neural network is now set to measure a movie’s success using the ROI formula.

We use two variables to measure its performance, with each reporting that we are on the right track, and the model is functioning like clockwork.

According to the classifier, the algorithm has an 88% accuracy of movie success prediction.

For reference, the previous team couldn’t get it higher than 0.65 (65%) success rate.
F1 SCORE = 0.836
ROC AUC = 0.88
Score F shows the number of successful forecasts and the number of metrics the model won’t identify.
ROC AUC: Receiver Operating Characteristic curve. The higher it is, the more capable the classifier is.
Classifier
features
Building on the resulting data and model improvements, we can foster new features:
  • Predicting a movie’s financial success

  • Adjusting and leveraging the technology in decisions by putting the “tune” in investment opportunity

  • Evaluating factors affecting the model performance
Classifier
features
Screen writers and producers will capitalize on our neural network before the actual development by optimizing the resources available.
Mechanics behind the model:
1
The collaboration with the Classifier will be key when it comes to writing a synopsis
2
The neural network will create a brand-new synopsis using a default prompt
3
But it is not just about the creation: the model needs improving, so it will also employ such factors as hybrid genres, i.e., a “barrier”.
Future Work
The second stage features the following targets:
  1. Create a real MVP
  2. Build a sustainable business model
USD 3 million
Project funding required:
Series and cinematic universes
The model is required to test well with TV shows, mini-series and cinematic universes, and updated accordingly
Enhanced news parsing
Marketing campaigns, i.e., news, are the key factor that contributes to the success of movies. However, we have collected all data available, which means their usage is limited
Time frame
AI will factor in data over time, for example, changes in genre, director popularity and the like
Data collection and analysis
We are now mostly employing IMDb in our analyses and breakdowns, but there are plans to include other sources and databases down the road: it will improve analytics drastically
Difference makes all the difference
The model is supposed to learn what near-identical movies are and find its way around remakes
Uncertainty quantification
Uncertainty quantification is the staple here, so we need to quantify both aleatoric and epistemic uncertainty (uncertainty of the model and data).
Advanced architecture model
Our graph database only features data on relationships between movie makers and movies. We need to allow a natural representation of higher-order relationships, so advancement is a must.
Generator optimization
The generator now produces a sample, which is then uploaded to the classifier; it’s far from perfect, but that’s something we’re working on
Stru
CTURE
1. Market (Case Study)
2. About project
3. Vision
4. Research
5. Results
6. Strategy
Case
study
Artificial Intelligence
in the Movie Industry
For 2019, the artificial intelligence market was valued at $39 billion globally and is projected to grow at a compound annual growth rate of 42% from 2020 to 2027. From 2016 to 2019, the number of businesses leveraging AI in their day-to-day operations skyrocketed by 270%; movie companies were no exception.
Companies offering AI and ML assistance are flourishing; their number is higher than ever. They do almost everything under the sun: from streamlining workflow to selecting actors based on their online presence, social media persona and media mentions.
However, because of the social stigma surrounding AI and movie-making, companies are not exactly rushing to share this particular detail about their day-to-day activities, so the data is incomplete.
What we know is, NETFLIX values its algorithm at $1 billion, while Qloo raised a handsome sum of $14 million (from AXA venture partners, Leonardo DiCaprio, Elton John and others); Cinelytic raised $2.3 million from T&B in 2018, ScriptBook, $1.4 million in 2016; Vault AI, $200 thousand in 2015 and $1.2 million in 2018 from Azteca.
That said, annual sales revenue for 2018 (from what we know) is $420K for Qloo and $620K for Legendary Entertainment.
Script Book
Independent company
Features
Assessing a movie’s financial potential
Predicting movie box office receipts by country
Researching and evaluating a script (by 400 parameters; over 30k of scripts in the database)
Predicting the audience by gender, race and age
Analyzing character types (emotions, good guys/bad guys)
Supporting decision-making
Writing a movie script (10 quality pages)
Successful case study: Passengers
The algorithm estimated the box office receipts at USD 118.1 during the pre-production; the movie ended up at USD 100 million.
Algorithm flexibility case study: La La Land
Receipts forecast: $59 million.
Final box-office figures: $446 million.

The small budget came into play, which drove the algorithm to approve the release of the movie (allowing for experimentation).
Successful algorithm accuracy case study
In 2015–2017, Sony released 62 movies, 32 of which were flops. The AI instantly rejected 22 movies as soon as they got into its database. Experts were accurate in 36% of cases, as opposed to the 80% accuracy of the algorithm.
Merlin Video
With contributions made by 20th Century Fox and Google
Features
Assessing a movie’s financial potential
Analyzing trailers
Analyzing characters in a trailer
Predicting audiences based on a trailer
Providing marketing strategy tips
How it works
The service analyzes a trailer and assigns a specific word to every character of the movie. For example, Logan from X-Men got “beard”, “car”, “man”, “forest”, “light”, “atmosphere”, “tree”, “facial hair” assigned to him. And this data is further used to make predictions and suggestions.
Successful case study
The AI predicted the target audience for The Greatest Showman using just the text description and video trailer
Cinelytic
Features
Assessment of a movie’s financial potential
Project management
Evaluation of project participants
(500k actors in the database)
Content evaluation
(95k in the database)
How it works
The AI measures a movie’s financial performance based purely on the cast, with their past projects’ receipts and social media taken into account.
The algorithm does not stop there: it also factors in the number of theaters releasing the movie and finds out just how much the movie revenue will lose/win from an actor getting replaced.
Successful case study
The algorithm nailed it with La La Land, Get Out and A Quiet place defying all odds.
CLIENTS
About Project
Vision
Unfortunately, human mind is not able to comprehend the entirety of the world, with its analytical data and relationships to take into account; that’s the job for a machine.

Many people trust their gut, honed through years of experience and common market practice; this gut feeling guides their decisions. People try to find a pattern in random events, which leads them further away from the right decision.
AL and ML in the movie industry and stock market prediction are cut from virtually the same cloth.

You need to identify trends and patterns based on empirical data.

The movie industry is something that has a set of features we can use to look for patterns which have a close connection to reality.
We are looking to come up with a system that will make for better investment decisions in the long term.

It seeks to reduce a level of uncertainty, resulting in more informed investment decisions.
The AI service market is still in its early stage, especially when it comes to movies. This means that we have our work cut out for us.
Our vision
Our goal is to design a perfect system.
This system is supposed to pave the way for attractive investment and better investment opportunities.
Criteria the system will use to come up with relevant keywords:
Example: drama, comedy
Genre and genre elements
01
  • Time and place of the action

  • Protagonist / antagonist (gender, age, social status)

  • Individual characteristics of the protagonist (deviations)

  • For example: an alcoholic

  • Basic premise. The protagonist’s obstacles on his journey

  • Story arc. Key events that trigger character development (make them face a choice)
Components that unlock the potential
of a movie
02
For example: environment
Associated topics
03
Locations
04
Seasonal release
05
Director, CGI and cast
require additional looking into
06
OUR
GOAL
To successfully apply machine learning algorithms to the
business decisions of media holdings by generating movie
loglines and predictions about a movie’s success
The Result. What Will It Be Like?
The project model is a system that can predict movie success, measured by return on investment. It uses various factors to provide a detailed breakdown and evaluation of potential future success.

It involves:
Do research and collect all models we can possibly find that might carry the day
1
Test those models and identify the best-performing parameters
2
Write an effective case study which proves that the algorithm actually works
Research into the ways we can leverage models for new formats, languages and data sources
3
4
First and foremost, they are designed to help screen writers find their inspiration and an original idea.
Action Plan
Stage 1 (research)
Status: completed
Stage 1 (application)
Status: WIP
6
Use MVP to discover a viable monetization strategy and development model
5
Build an MVP to launch it and map out our product offer
4
Research into the ways we can leverage models for new formats, languages and data sources
3
Write an effective case study which proves that the algorithm actually works
2
Test those models and identify the best-performing parameters
1
Do research and collect all models we can possibly find that might carry the day

Research Results (Stage 1)

Research Result (1/2)
  • The U.S. movie market only

  • Feature movies only

  • English-language movies only

  • Movies that made it to theaters only

  • Movies with budget and box office data available only
Features
  • We analyzed the IMDb Pro database and news datasets as sources

  • We created a structured repository of data to be collected, with data processing algorithms

  • We developed an algorithm for employing metrics to find out whether a decision is viable
Research
  • We sent the resulting data to experts to improve and fine-tune algorithms

  • We finished infrastructure design to be able to further handle data and adapt seamlessly to ever-changing conditions

  • We built a model to engineer the resulting data
The steps we took set a certain benchmark for all metrics, which was instrumental for identifying the most popular trends; it became the boon to our development of the MVP
Result
Research Result (2/2)
Classification system
We designed a neural network to build a model, which will help predict the box-office receipts of a particular feature movie
To evaluate financial performance of a movie, the classifier uses the following formula of return on investment (ROI):
Analytically-wise, the classifier builds on six key metrics:
ROI =
a
≥
REVENUE – BUDGET
BUDGET
where a means a parameter
  1. Numerical metrics:
Movie budget, box office receipts

2. Category metrics:
Movie genre, sub-genre

3. Rating metrics:
IMDb user reviews
4. Text metrics:
Title, slogan, synopsis

5. Draft metrics:
Relationship among the cast

6. News metrics:
Headlines and emotions behind the printed word
Analytical model of
the classifier
Classifier Architecture
Movie data

Metrics
identification
Category metrics
Categorical
embedding
Data-based
training
Classifier training
Text metrics
BERT embedding
Draft metrics
GNN embedding
Training / split testing
News data
Cluster analysis
Numerical metrics
Standardization
Data-based
testing
Determination of the classifier quality
Emotion-based classification
BERT: bidirectional neural network
GNN: three-tier convolutional neural network
Measuring the Performance of the Classifier
The classifier neural network is now set to measure a movie’s success using the ROI formula.

We use two variables (F1 Score Roc Auc) to measure the performance of the neural network, with each reporting that we are on the right track, and the model is functioning like clockwork.

According to the classifier, the algorithm has an 88% accuracy of movie success prediction.

For reference, the previous team couldn’t get it higher than 0.65 (65%) success rate.
F1 SCORE = 0.836
ROC AUC = 0.88
Score F shows the number of successful predictions and the number of metrics the model won’t identify.
ROC AUC: Receiver Operating Characteristic curve. The higher it is, the more capable the classifier is.
Classifier
features
Building on the resulting data and model improvements, we can foster new features:
Predicting a movie’s financial success
Adjusting and leveraging the technology in decisions by putting the “tune” in investment opportunity
Impact assessment
1. Casting the best actors for a particular movie, thus achieving the most effective performance

2. Simulating the powerful information campaign for the project

3. Casting the best actors for a particular movie, thus achieving the most effective performance

4. Optimizing movie-making costs
Neural Networks Ensemble
Mechanics behind the synopsis:
Screen writers and producers will capitalize on our neural network before the actual development by optimizing the resources available.:
  1. The collaboration with the Classifier will be key when it comes to writing a synopsis.
  2. The neural network will create a brand-new synopsis using a default prompt
  3. But it is not just about the creation: the model needs improving, so it will also employ such factors as hybrid genres, i.e., a “barrier”.
Mechanics behind the synopsis:
  • We made headway when we were also able to create text generated following a particular subject area

  • We can’t give a definite answer as to whether the generator will actually work

  • We need more data, as it will set the stage for modified architecture

  • Further development involves the model generating anchor points (key tags)
Ensemble
Results (1/2)
The classifier neural network is now set to measure a movie’s success using the ROI formula.
Input data:
Title:
1
2
3
Harry Potter and the Goblet of Fire
Logline:
When Harry Potter's name emerges from the Goblet of Fire, he becomes a competitor in a grueling battle for glory among three wizarding schools—the Triwizard Tournament. But since Harry never submitted his name for the Tournament, who did?
Genres:
Drama, fantasy
When Harry Potter's name emerges from the Goblet of Fire, he becomes a competitor in a grueling battle for glory among three wizarding schools—the Triwizard Tournament.

But since Harry never submitted his name for the Tournament, who did? And why did the Goblet of Fire select Harry Potter to enter the competition? This fun-filled Rogue One adventure is brought to you by the hit film, "Harry Potter and the Goblet of Fire” (with Ron Weasley and Daniel Radcliffe)!
Result:
Synopsis:
The classifier neural network is now set to measure a movie’s success using the ROI formula.
Input data:
Title:
1
2
3
The Shawshank Redemption
Logline:
Two imprisoned men bond over a number of years, finding solace and eventual redemption through acts of common decency.
Genres:
Drama
Two imprisoned men bond over a number of years, finding solace and eventual redemption through acts of common decency. "The Shawshank Redemption" is set in the fictional Pennsylvania state prison at Shawshank, during the late 1930s, a time in which those seeking freedom from the federal government were routinely treated with indifference by the guards and the fellow inmates.
For Frank Martin (Jimmy Durante), a grifter, the only break from the doldrums is a chance encounter with the more aggressive but loveable Harvey Darrell (Ben
Result:
Synopsis:
Ensemble
Results (2/2)
The classifier neural network is now set to measure a movie’s success using the ROI formula.
Input data:
Title:
1
2
3
The Wheel of Time
Logline:
Follow Moiraine, a member of the shadowy and influential all-female organization called the “Aes Sedai” as she embarks on a dangerous, world-spanning journey with five young men and women. Moiraine believes one of them might be the reincarnation of an incredibly powerful individual, whom prophecies say will either save humanity or destroy it.
Genres:
Drama, fantasy
Synopsis:
Follow Moiraine, a member of the shadowy and influential all-female organization called the "Aes Sedai" as she embarks on a dangerous, world-spanning journey with five young men and women. Along the way, they'll cross thousands of miles of dangerous landscape and face a perilous journey through a hostile land as each of them faces an epic choice: abandon humanity and be sacrificed to the forces of evil or - as Moiraine believes - take on their destiny by saving themselves.

***Warning: this is not for the squeamish, but for the adventurous!***
Result:
Synopsis:
The classifier neural network is now set to measure a movie’s success using the ROI formula.
Input data:
Title:
1
2
3
Invasion
Logline:
Earth is visited by an alien species that threatens humanity’s existence. Events unfold in real time through the eyes of five ordinary people across the globe as they struggle to make sense of the chaos unraveling around them.
Genres:
Drama, sci-fi, fantasy
Earth is visited by an alien species that threatens humanity's existence. Events unfold in real time through the eyes of five ordinary people across the globe as they struggle to make sense of the chaos unraveling around them. The 20th of July marks the start of the alien invasion of Earth. It comes with a warning. When the humans on Earth detect an asteroid passing over, a portal opens in space that connects them to the alien species, but these travelers can only stay behind. In New York City, Audrey Crowes (Laura Linney), a 30-year-old businesswoman from the boomtown of Jackson Heights, discovers a gathering of space aliens in the skies. Audrey and the aliens are declared a military threat by President Adam Stein (Eric Bana). He orders them to evacuate New York.
Result:
Synopsis:
Classifier Results
Successful prediction of The Last Duel’s flop
Successful prediction of Titanic being a box-office sensation
Stage 2
The second stage features the following targets:
  1. Create a real MVP
  2. Build a sustainable business model
USD 3 million
Project funding required:
Series and cinematic universes
The model is required to test well with TV shows, mini-series and cinematic universes, and updated accordingly
Enhanced news parsing
Marketing campaigns, i.e., news, are the key factor that contributes to the success of movies. However, we have collected all data available, which means their usage is limited
Time frame
AI will factor in data over time, for example, changes in genre, director popularity and the like
Data collection and analysis
We are now mostly employing IMDb in our analyses and breakdowns, but there are plans to include other sources and databases down the road: it will improve analytics drastically
Difference makes all the difference
The model is supposed to learn what near-identical movies are and find its way around remakes
Uncertainty quantification
Uncertainty quantification is the staple here, so we need to quantify both aleatoric and epistemic uncertainty (uncertainty of the model and data).
Advanced architecture model
Our graph database only features data on relationships between movie makers and movies. We need to allow a natural representation of higher-order relationships, so advancement is a must.
Generator optimization
The generator now produces a sample, which is then uploaded to the classifier; it’s far from perfect, but that’s something we’re working on
Made on
Tilda