This presentation introduces a revolutionary approach to movie streaming and distribution, inspired by a research paper on diversity in Hollywood film. The core idea is to reconstruct movies entirely from publicly available, Creative Commons licensed face-detection metadata (bounding box coordinates, actor identities, and timestamps), rather than using the copyrighted video content itself. This "few things" method claims to overcome several major challenges in the streaming industry: eliminating the high cost of streaming subscriptions by creating copyright-free movie clones, achieving unprecedented video compression (reducing a 2-hour movie to ~5MB while maintaining "K+ resolution" and "visually lossless" quality), and offering dynamic, on-the-fly personalization to address societal representation issues (e.g., increasing LGBTQ+ visibility or customizing cast members). The presenter demonstrates these capabilities with examples from films like "Everything Everywhere All at Once" and "The Favorite," highlighting how background actors and even lead roles can be programmatically swapped. The reconstructed movies are presented as a scalable vector video format and are available for live viewing on the presenter's website.
Revolutionizing Movie Streaming and Representation with "Few Things"
0:00:14 Introduction to the "Few Things" Project: The speaker introduces a project aiming to revolutionize movie streaming and society by addressing key problems: the high cost of streaming subscriptions (estimated at €919 annually for just three movies), poor video quality due to slow internet, and the pervasive lack of diversity (overrepresentation of white men) in Hollywood cinema.
0:03:36 Inspiration from Diversity Research: The project is inspired by David Bam's "Measuring Diversity in Hollywood" paper. This research involved legally acquiring and analyzing over 2,000 DVDs (top box office and award nominees from 1980-2022) to computationally track the representation of gender and ethnicity on screen using a sophisticated facial recognition pipeline.
0:05:09 Facial Recognition and Diversity Metrics: Bam's team used facial recognition to identify faces, their locations, and match them to actors' IMDb and WikiData profiles. They carefully accounted for evolving public perceptions of actors' genders (e.g., Elliot Page) through timestamped data. The research revealed persistent biases, such as women receiving only 25-30% of screen time and white men being significantly overrepresented in leading roles.
0:09:11 The "Few Things" Concept – Copyright-Free Reconstruction: The speaker's innovative idea is to reconstruct movies solely from the Creative Commons licensed face detection metadata provided by Bam's research (rectangles, actor IDs, timestamps), thereby creating "copyright-free clones" of films.
0:10:19 Leveraging Research Data: The project utilizes raw data including sequential IDs for face tracks, face recognition confidence scores for IMDb-identified actors, bounding box coordinates for each frame, and scene transition timestamps.
0:11:32 Visualizing with Actor Images: To make the reconstructed movies watchable, the project overlays transparent PNG images of recognized actors (sourced from IMDb) onto the bounding boxes. For unrecognized faces, random stock photos from Unsplash are used.
0:13:41 Pre-processing for Smoother Animation: Initial reconstructions showed flickering; this was resolved by pre-processing the data. This involves interpolating face rectangles for short disappearances (assuming a temporary occlusion) and consistently assigning sprites to unrecognized faces that reappear nearby, ensuring a smoother visual experience.
0:15:01 Copyright-Free Movie Demo: A demonstration of "Everything Everywhere All at Once" using this method shows that the film's "vibe," conversations, and special effects are still discernible, along with randomly assigned "extras" in the background.
0:16:05 Revolutionary Video Compression: The project also tackles file size. The "few things" method reduces a 2-hour movie like "The Favorite" to an astonishing 5 megabytes. This compares favorably against torrents (22GB), DVDs (4GB), Netflix HD (approx. 3GB), and even an extremely compressed FFmpeg encoding (11MB), while being described as "visually lossless."
0:19:01 K+ Resolution and Scalable Vector Video: The "few things" approach is characterized as a "scalable vector video format" that can achieve "K+ resolution." This means resolution is only limited by the quality of the individual actor images used, not the original video's resolution.
0:20:11 Fixing Societal Issues with Dynamic Content: The system allows for dynamic content modification to address representation issues and user preferences.
0:20:50 Increased LGBTQ+ Representation: Demonstrates "The Favorite" with all background actors replaced by images of queer and trans individuals.
0:21:26 Personalized Casting (Emma Stone Example): Shows how users can customize their viewing experience, such as watching "The Favorite" where every background actor is replaced by Emma Stone.
0:22:53 Modifying Lead Character Relationships: Illustrates a more complex modification by replacing Queen Anne in "The Favorite" with King Charles, altering the film's central queer relationships to be "more straight."
0:24:11 Conclusion and Live Availability: The speaker concludes that the project has successfully achieved copyright-free movies, revolutionary compression, and a solution for societal representation issues. The three example movies are currently available for live viewing on the presenter's website.