TL;DR
LLMPopcorn is a novel pipeline that leverages Large Language Models (LLMs) to act as "creative directors" for generating viral micro-videos. By combining Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) prompting, the system produces AI-generated content that rivals or even surpasses human-created videos in predicted popularity.
What Problem Does This Paper Address?
Micro-videos (like those on TikTok and YouTube Shorts) are a multi-billion-dollar industry, but creating "viral" content is traditionally labor-intensive, requiring professional scripting, filming, and editing. While AI video generation models (like Sora or LTX-Video) have made massive strides in visual quality, they lack the "creative spark" to know what content will actually engage an audience.
Current research typically focuses on technical metrics like resolution or duration. This paper addresses the "popularity gap": Can LLMs be used to autonomously plan and prompt videos specifically designed to be popular?
Methods
The authors introduce the LLMPopcorn pipeline, which treats the LLM as a video generation assistant. The process follows these steps:
- Input: A user provides a query (either "Concrete," like a specific product description, or "Abstract," like a general theme).
- Prompt Enhancement (PE): This is the core innovation. Instead of just passing the user query to a video generator, the system uses:
- Retrieval-Augmented Generation (RAG): It searches a database of existing micro-videos to find examples of what has succeeded (positive examples) and what has failed (negative examples) in that specific genre.
- Chain-of-Thought (CoT): The LLM follows a structured reasoning process—examining the retrieved data, brainstorming unique topics, and developing a logical script—before generating the final video prompt and title.
- Video Generation: The enhanced prompt is fed into a pre-trained video model (e.g., LTX-Video, HunyuanVideo).
- Popularity Prediction: An offline evaluator (the MMRA model) predicts the video's engagement potential based on the generated video and its title.
Datasets
- Microlens: The largest available micro-video dataset, used as the source for the RAG component and as a human-created baseline for comparison.
- Custom Prompt Datasets: The authors used ChatGPT-4o to generate 200 unique user prompts (100 concrete, 100 abstract) across five categories: Anime, Delicacy, Daily Sharing, Film & Television, and Gaming.
Key Insights and Ideas
- LLMs as Strategic Planners: The study finds that LLMs are highly effective at structuring descriptions. By analyzing past successes and failures via RAG, they can steer video generators away from "static" or boring content toward more dynamic, narrative-driven scenes.
- DeepSeek Dominance: In the benchmarking phase, DeepSeek-V3 and DeepSeek-R1 consistently outperformed other models like Llama-3.3 and GPT-4o in generating high-popularity prompts.
- Abstract vs. Concrete: LLMs showed a particular strength in handling "Abstract" prompts. While humans often struggle to visualize a vague concept, the LLM could transform a general idea into a diverse range of imaginative, high-engagement scenes.
- The "Human-Level" Threshold: The most striking finding is that DeepSeek-V3-generated videos achieved median popularity scores (0.73 for abstract prompts) that slightly exceeded the human-crafted videos in the Microlens dataset (0.71).
Key Results
- Popularity Scores: DeepSeek-V3 achieved median popularity scores of 0.56 (concrete) and 0.73 (abstract), compared to human-crafted scores of 0.44 and 0.71, respectively.
- Prompt Enhancement (PE) Impact: PE significantly boosted results. In a user study, participants preferred videos generated with PE over the basic pipeline 66.67% of the time for concrete prompts and 62.50% for abstract prompts.
- Video Model Benchmarking: LTX-Video and HunyuanVideo were found to be more reliable and higher-performing than CogVideoX-5b for this specific task.
Limitations and Future Work
- Video Model Constraints: While the LLM "scripts" are excellent, the actual visual output is still limited by the current state of open-source video generation models, which can sometimes produce low-resolution or short-duration clips.
- Evaluation Proxy: The study relies on an offline popularity predictor (MMRA). While robust, it is a proxy for real-world viral dynamics which are notoriously fickle.
- Future Directions: The authors suggest exploring Reinforcement Learning from Human Feedback (RLHF) to further align LLM prompts with real-world engagement metrics and contextual cues.
Summary updated