Papers
Topics
Authors
Recent
Search
2000 character limit reached

SkyScript-100M: 1,000,000,000 Pairs of Scripts and Shooting Scripts for Short Drama

Published 18 Aug 2024 in cs.CL | (2408.09333v2)

Abstract: Generating high-quality shooting scripts containing information such as scene and shot language is essential for short drama script generation. We collect 6,660 popular short drama episodes from the Internet, each with an average of 100 short episodes, and the total number of short episodes is about 80,000, with a total duration of about 2,000 hours and totaling 10 terabytes (TB). We perform keyframe extraction and annotation on each episode to obtain about 10,000,000 shooting scripts. We perform 100 script restorations on the extracted shooting scripts based on our self-developed large short drama generation model SkyReels. This leads to a dataset containing 1,000,000,000 pairs of scripts and shooting scripts for short dramas, called SkyScript-100M. We compare SkyScript-100M with the existing dataset in detail and demonstrate some deeper insights that can be achieved based on SkyScript-100M. Based on SkyScript-100M, researchers can achieve several deeper and more far-reaching script optimization goals, which may drive a paradigm shift in the entire field of text-to-video and significantly advance the field of short drama video generation. The data and code are available at https://github.com/vaew/SkyScript-100M.

Summary

  • The paper introduces SkyScript-100M, a massive dataset comprising 1 billion paired scripts and shooting scripts specifically targeting AI short drama video generation.
  • The dataset was built by extracting shooting scripts from 6,660 short drama episodes and using an AI model to restore 100 script variations per shooting script.
  • SkyScript-100M aims to foster advancements in AI script generation, video synthesis, character consistency, and highlight detection for more nuanced drama production.

An Overview of SkyScript-100M: Revolutionizing Short Drama Production with a Billion Script Pairs

The paper, "SkyScript-100M: 1,000,000,000 pairs of scripts and shooting scripts for Short Drama," introduces an ambitious dataset designed to transform the short drama video generation domain. SkyScript-100M stands as a colossal repository comprising one billion paired scripts and shooting scripts, intended for applications in the burgeoning field of AI-powered drama production. This paper meticulously details the dataset's construction process, intrinsic characteristics, and its potential implications for advancing AI methodologies in script generation for video production.

Dataset Construction and Methodology

The authors have compiled SkyScript-100M by collecting 6,660 popular short drama episodes from the Internet, resulting in approximately 80,000 episodes with over 2,000 hours of content. From these episodes, 10,000,000 shooting scripts were carefully extracted and annotated through keyframe analysis. The process involved deploying a self-developed large short drama generation model, SkyReels, to restore 100 script variations from each extracted set of shooting scripts, culminating in the creation of the massive dataset. This construction involved multiple iterations of machine learning models for feature extraction, data cleansing, and refinement, invoking advanced multimodal LLMs like InternVL2 for initial object and character detection.

Contributions and Comparison to Existing Work

SkyScript-100M is notably positioned within the paper as having a pivotal role in multimodal datasets, particularly targeting the short drama domain. While existing datasets like MSR-VTT and HowTo100M cater to generalized video-text tasks, SkyScript-100M distinguishes itself by its specificity and scale, tailored explicitly for short drama content. By focusing on refined shooting scripts, the dataset introduces an enriched paradigm for narrative articulation and world-building, whereby detailed shooting scripts are aligned with AI models to facilitate coherent video generation tasks.

Implications for AI and Short Drama Generation

The practical implications of SkyScript-100M are vast, potentially fostering a new wave of innovations in automatic script and video generation. By setting a foundation for a new short drama generation paradigm—described within the paper as aligning AI models more closely with the thematic and emotional nuances prevalent in short dramas—this work anticipates improvements in areas such as character consistency, plot coherence, and emotional impact. Moreover, the paper elucidates novel applications in implicit character relationship mining and video highlight detection, areas not adequately addressed by existing datasets.

Future Prospects and Research Directions

The dataset not only lays a cornerstone for the advancement of short drama video generation but also opens avenues for related research fields such as multimodal dialogue systems, interactive storytelling, and AI-driven content creation. The complex interplay between narrative structure and machine understanding proposed in SkyScript-100M could inspire further investigations into how AI models interpret and generate contextually rich video content consistent with human storytelling techniques.

The paper suggests that future work will involve expanding SkyScript and optimizing the SkyReels model to leverage the extensive data for finer script and video synthesis. As such, this dataset is expected to act as a catalyst for both academic inquiries and practical applications concerning AI's role in media and entertainment.

In summary, SkyScript-100M constitutes a significant contribution to AI research in drama production, heralding a shift towards more nuanced and autonomous video generation systems. The methodologies, results, and implications outlined in the paper provide a roadmap for the creative and technical challenges yet to be addressed in this evolving intersection of artificial intelligence and digital storytelling.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Continue Learning

We haven't generated follow-up questions for this paper yet.

Collections

Sign up for free to add this paper to one or more collections.