Spaces:

---

junchenfu

/

llmpopcorn-demo

like 0

Running on Zero

LLMPopcorn: LLM-Assisted Popular Micro-Video Generation (ICASSP 2026)

We are a research group from the University of Glasgow. This Space is the official demo for our ICASSP 2026 paper "LLMPopcorn: Exploring LLMs as Assistants for Popular Micro-video Generation" (arXiv: 2502.12945).
What the demo does:
The demo takes a user's video idea as input and automatically generates a trending-optimized title, and a 3-second video clip using a text-to-video diffusion model. It showcases two modes:
Basic — direct LLM generation
PE (Prompt Enhancement) — RAG + Chain-of-Thought reasoning over 19,560 real-world TikTok-style videos from the MicroLens dataset

Why we need GPU resources:
The pipeline requires running two large models simultaneously:
Text-to-video model: Lightricks/LTX-Video (2B parameters) for generating the actual video clip
LLM: meta-llama/Llama-3.1-8B-Instruct (8B parameters) for title and prompt generation with RAG + CoT reasoning
Both models require GPU for any reasonable inference speed. Without GPU, the demo is non-functional for end users.

Impact:
This is an academic open-science project. The demo, datasets, and code are fully open-sourced on GitHub (GAIR-Lab/LLMPopcorn) and Hugging Face. We hope this demo helps the research community explore LLM-driven video content creation. The paper has already attracted interest from the HF team (we were reached out to by @NielsRogge for submission to hf.co/papers).
We would greatly appreciate a ZeroGPU grant to make this demo publicly accessible.