Research map / Video / Video-language understanding
TeaserGen: Generating Teasers for Long Documentaries
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DocumentaryNet (1,269 documentary-teaser pairs) and TeaserGen, a two-stage LLM-narration-plus-visual-matching system for generating teasers from long documentaries.
Problem
Generating teasers for long documentaries requires long-range multimodal modeling, audiovisual alignment, smooth scene transitions and factual accuracy, but no public dataset existed to study it.
Approach
Narration-centered pipeline: GPT-4o rewrites transcribed narration into a story-like teaser script with an ending question, then selects matching video clips either by thresholding a pretrained video temporal grounding score (TeaserGen-PT) or via a trained transformer with CLIP diffusion prior, beam search and a diversity regularizer (TeaserGen-LR).
Key results
- TeaserGen-LR with beam search and diffusion prior reaches F1 1.88% and CLIPScore 0.63, with SCR 46.56% versus 27.6% for ground truth; the two baselines show SCR of roughly 90-99%.
- In human evaluation, TeaserGen-PT (title query) scores highest on coherence (3.14) and engagingness (2.81) among compared systems, beating UniVTG and CLIP-it.
- Finely-tuned GPT-4o narration prompts beat naive summarization on organization (3.88 vs 3.58), informativeness (3.82 vs 3.72) and engagingness (3.70 vs 3.60).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use title rather than narration as the query for TeaserGen-PT | TeaserGen-PT · Documentary teaser test set | F1 (%): 1.07 -> 1.85 | Helped |
| Use title rather than narration as the query for TeaserGen-PT | TeaserGen-PT · Documentary teaser test set | CLIPScore: 0.58 -> 0.56 | Hurt |
| Use TeaserGen-PT instead of UniVTG | Documentary teaser generation · Documentary teaser test set | F1 (%): 1.82 -> 1.85 | Helped |
| Use TeaserGen-PT instead of UniVTG | Documentary teaser generation · Documentary teaser test set | SCR (%): 89.68 -> 13.16 | Helped |
| Use beam search instead of greedy decoding (ablation) | TeaserGen-LR without diffusion prior · Documentary teaser test set | F1 (%): 1.56 -> 1.88 | Helped |
| Add smoothing (ablation) | TeaserGen-LR greedy decoding without diffusion prior · Documentary teaser test set | F1 (%): 0.53 -> 1.56 | Helped |
| Add smoothing (ablation) | TeaserGen-LR greedy decoding without diffusion prior · Documentary teaser test set | REP (%): 81.32 -> 31.97 | Helped |
| Add diffusion prior (ablation) | TeaserGen-LR greedy decoding with smoothing · Documentary teaser test set | F1 (%): 1.56 -> 1.38 | Hurt |
| Add diffusion prior (ablation) | TeaserGen-LR greedy decoding with smoothing · Documentary teaser test set | CLIPScore: 0.58 -> 0.62 | Helped |
| Add diffusion prior (ablation) | TeaserGen-LR beam search with smoothing · Documentary teaser test set | SCR (%): 41.97 -> 46.56 | Hurt |
| Use finely tuned prompts instead of naive summarization (ablation) | LLM teaser narration generation · Teaser narration subjective evaluation | Organization: 3.58 ± 0.57 -> 3.88 ± 0.44 | Helped |
| Finetune the pretrained video temporal grounding model (ablation) | TeaserGen-PT with narration query · Documentary teaser test set | F1 (%): 1.07 -> 1.23 | Helped |
Surprising findings
- The CLIP-NN baseline achieves the highest CLIPScore but has very high repetitiveness (92.73%), and the higher-F1 TeaserGen-LR model scores lower than TeaserGen-PT in human coherence and engagingness.
- Finetuning UniVTG on about 10k in-domain samples worsened performance, attributed to the small finetuning set relative to pretraining data.
Limitations
- Cannot accurately match interview scenes to names and faces because the language-vision model does not link names to faces.
- Assumes a scene change at every sentence boundary and that narration matters more than visuals, which may not hold for vlogs or silent movies.
- Generated teaser narrations fall short of professional scripts in artistic quality and creativity.
Experimental setup
Datasets: DocumentaryNet ActivityNet
Benchmarks: DocumentaryNet
Metrics: F1 REP (repetitiveness) SCR (scene change rate) CLIPScore VTGHLS subjective coherence/alignment/engagingness/realness Likert
Baselines: Random selection CLIP-NN CLIP-IT UniVTG
Builds on: UniVTG CLIP-IT CLIP Whisper GPT-4o
Closest papers by meaning
- REGen: Multimodal Retrieval-Embedded Generation for Long-to-Short Video Editing (NeurIPS 2025 · similarity 0.87)
- LVD-2M: A Long-take Video Dataset with Temporally Dense Captions (NeurIPS 2024 · similarity 0.85)
- MovieBench: A Hierarchical Movie Level Dataset for Long Video Generation (CVPR 2025 · similarity 0.84)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.84)
- Is Your World Simulator a Good Story Presenter? A Consecutive Events-Based Benchmark for Future Long Video Generation (CVPR 2025 · similarity 0.84)
- SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.