PaperlayerView exact revision on arXiv

Source-linked AI summary

PresentAgent: Multimodal Agent for Presentation Video Generation

Jingwei Shi, Zeyu Zhang, Biao Wu, Yanjie Liang, Meng Fang, Ling Chen, Yang Zhao

arXiv:2507.04036v1cs.CV

TL;DR

PresentAgent targets the gap between static document summaries and fully narrated presentation videos by coordinating slide planning, narration, and audio-visual synchronization. PresentEval assesses these outputs through factual comprehension and subjective quality judgments, and the reported experiments show performance approaching human-level quality. The paper also notes limited evaluation diversity and the current absence of dynamic slide animations.

  • Problem

    Existing methods produce static visual summaries or generic video clips without structured narration, leaving long-form document-to-presentation video generation insufficiently addressed.

    Source: §1 · pp. 1–2

  • Method

    PresentAgent segments documents, plans layout-guided slides, generates oral-style narration, and synthesizes synchronized audio-visual presentation videos, while PresentEval uses quizzes and vision-language-model scoring.

    Source: §1 · p. 2

  • Results

    PresentAgent approaches human-level performance and significantly outperforms competitive variants across experiments and ablations.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The results support controllable and reusable multimodal presentation generation across diverse document types.

    Source: §1 · p. 2, §8 · p. 9

    Evaluation was limited to five academic papers, and PresentAgent currently generates static slides without dynamic animations or effects.

    Source: §7 · p. 9

Abstract

from arXiv · show

We present PresentAgent, a multimodal agent that transforms long-form documents into narrated presentation videos. While existing approaches are limited to generating static slides or text summaries, our method advances beyond these limitations by producing fully synchronized visual and spoken content that closely mimics human-style presentations. To achieve this integration, PresentAgent employs a modular pipeline that systematically segments the input document, plans and renders slide-style visual frames, generates contextual spoken narration with large language models and Text-to-Speech models, and seamlessly composes the final video with precise audio-visual alignment. Given the complexity of evaluating such multimodal outputs, we introduce PresentEval, a unified assessment framework powered by Vision-Language Models that comprehensively scores videos across three critical dimensions: content fidelity, visual clarity, and audience comprehension through prompt-based evaluation. Our experimental validation on a curated dataset of 30 document-presentation pairs demonstrates that PresentAgent approaches human-level quality across all evaluation metrics. These results highlight the significant potential of controllable multimodal agents in transforming static textual materials into dynamic, effective, and accessible presentation formats. Code will be available at https://github.com/AIGeeksGroup/PresentAgent.

1 Introduction

PresentAgent addresses document-to-presentation video generation by combining structured slides, oral narration, and synchronized audio-visual rendering. PresentEval evaluates these multimodal outputs across factual comprehension and subjective quality, while experiments on 30 document–presentation pairs show performance approaching human-level quality.

  • Existing methods typically produce static visual summaries or generic video clips without structured narration, limiting their use for structured communication.Source: §1 · p. 1
  • PresentAgent segments documents into semantic blocks, generates layout-guided slides and oral-style narration, then synthesizes aligned audio and visuals into a presentation video.Source: Figure 1 · p. 1, §1 · p. 2
  • PresentEval combines factual comprehension assessment with vision-language-model preference scoring to measure objective correctness and subjective video quality.Source: §1 · p. 2, Figure 3 · p. 3
  • The evaluation uses a curated set of 30 human-authored document-video pairs spanning diverse domains.Source: §1 · p. 2
  • The paper formulates document-to-presentation video generation as converting long-form documents into narrated videos with synchronized slides and speech.Source: §1 · pp. 1–2
  • PresentAgent approaches human-level performance and significantly outperforms competitive variants across experiments and ablations.Source: §1 · pp. 2–3

2 Related Work

Prior work supports document-to-slide generation and general-purpose vision-language agents, but it does not fully address long-form document understanding with narrated, temporally synchronized presentation videos.

2.1 Document-to-Multimodal Generation

  • Document-to-slide systems construct layout-aware static outputs, whereas PresentAgent adds temporal planning and audio-visual synchronization for narrated presentation videos.Source: §2.1 · p. 3

2.2 Vision-Language Agents

  • General-purpose agents support reasoning, code execution, API calls, interface manipulation, and motion generation, but still struggle with professional tasks requiring accuracy and domain-specific knowledge.Source: §2.2 · p. 3
  • Unlike slide automation from short prompts, the proposed task processes entire long-form documents and requires content understanding, multimodal alignment, speech generation, and video synthesis.Source: §2.2 · p. 3

3 Presentation Benchmark

The Doc2Present Benchmark pairs diverse documents with human-crafted presentation videos, while PresentEval combines objective quiz comprehension with subjective vision-language scoring. Together, they assess both information fidelity and presentation quality.

3.1 Doc2Present Dataset

  • The Doc2Present Benchmark contains document–presentation video pairs across business, educational, policy, research, and instructional domains.Source: §3.1 · p. 4, §3 · p. 4
  • The benchmark includes 30 video samples with synchronized slide visuals and voiceover, aligned to source documents whose videos preserve structure and compactly convey information.Source: §3.1 · p. 4
  • Documents average 3,000–8,000 words, while corresponding presentations last 1–2 minutes and contain 5–10 slides.Source: §3.1 · p. 4

3.2 PresentEval

  • PresentEval gives vision-language models complete slide images and narration transcripts to assess generated presentations without ground-truth references.Source: §3.2 · p. 4
  • Objective Quiz Evaluation uses five document-specific multiple-choice questions, producing a 0–5 comprehension score from correct answers.Source: §3.2 · p. 4, Table 1 · p. 5
  • Subjective Scoring rates narration coherence, visual or audio quality, and comprehension clarity on a 1–5 scale using dimension-specific prompts.Source: §3.2 · p. 5, Table 2 · p. 5

4 PresentAgent

PresentAgent converts documents into synchronized narrated presentation videos through a modular pipeline that plans semantic segments, composes slides, generates narration, and assembles aligned audiovisual output.

4 PresentAgent

  • PresentAgent transforms long-form documents into presentation videos by segmenting content, generating slides and narration, and aligning both modalities over time.The pipeline is designed for controllability, interpretability, and multimodal alignment.Source: §4 · p. 5, §4.1 · pp. 5–6

4.1 Problem Formulation

  • The problem formulation treats the whole document as a globally structured input rather than directly generating slide elements from isolated chunks.This design first identifies high-level structure before producing slide visuals and narration from scratch.Source: §4.1 · pp. 5–6

4 PresentAgent

  • PresentAgent’s evaluation pipeline uses designed prompts and a VLM-based scoring process to produce dimension-specific metrics for generated videos.The overview includes this evaluation stage after video composition.Source: Figure 4 · p. 6

4.2 Slide Planning and Composition

  • Document processing produces semantic content blocks, assigns slide types and layouts using structural cues, and populates them with editable operations such as replace_text, insert_image, and add_list.Slides are rendered into static visual frames using python-pptx or HTML-based renderers.Source: §4.2 · p. 6

4.3 Narration and Audio Synthesis

  • Each slide receives concise oral-style narration generated from its key message, converted to speech, paired with the slide, and timestamped for synchronized rendering.Narration length is controlled to a target duration typically between 30 and 150 seconds.Source: §4.3 · p. 7

4.4 Video Assembly

  • The final video displays each slide for the duration of its corresponding audio segment and combines the tracks into a standard video file.Optional transitions can be inserted between segments during FFmpeg-based composition.Source: §4.4 · p. 7

5 Experiments

Experiments compare PresentAgent videos with human-created references using quiz-based comprehension and dimension-specific VLM scoring. Results show strong factual performance and competitive but uneven subjective quality across model variants.

5.1 Evaluation Setup

  • The evaluation uses 30 long-form documents paired with manually created reference videos spanning education, product explanation, research overviews, and policy briefings.Each document also receives a generated presentation video from the full pipeline.Source: §5.1 · p. 7
  • PresentEval combines fixed multiple-choice comprehension questions with subjective scoring of content quality, visual quality, and comprehension accuracy.The split strategy evaluates full videos for quizzes and short video or audio segments for focused quality scoring.Source: §5.1 · p. 7

5.2 Implementation Details

  • The implementation parses documents into hierarchical topic trees, generates per-slide PowerPoint content and images, and produces synchronized Chinese or English narration assembled into 1080 p video.The pipeline supports PDF, Markdown, DOCX, and web snapshots, with optional captions and fade transitions.Source: §5.2 · p. 8

5.3 Main Results

  • Claude-3.7-sonnet achieves the highest quiz accuracy at 0.64, exceeding the human reference at 0.56, while Qwen-VL-Max and Gemini-2.5-flash score 0.52.The quiz evaluates factual alignment between generated content and source documents.Source: §5.3 · p. 8
  • Human-created presentations lead overall subjective video and audio scores, but GPT-4o-Mini reaches video content and visual appeal scores at or near 4.8.Claude-3.7-sonnet has the most balanced audio quality with a mean of 4.53.Source: §5.3 · p. 8
  • Gemini-2.5-flash achieves the highest visual quality score of 5.0 but lower comprehension, indicating a trade-off between aesthetics and clarity.This result illustrates why PresentEval reports multiple quality dimensions rather than a single score.Source: §5.3 · p. 8

5.4 Analysis

  • The demonstration shows a technical blog converted into slides with oral-style captions and synchronized speech while covering technical topics such as parallelization workflow and agent system architecture.The example is presented as evidence of coherent, conversational technical communication.Source: §5.4 · pp. 8–9

6 Discussion

The discussion identifies a limitation in evaluating multimodal presentation videos: assessing visual, textual, and audio modalities independently can miss their semantic and temporal coherence. It therefore motivates fusion-aware evaluation and multimodal reasoning.

6 Discussion

  • Independent assessments of visual clarity, textual relevance, and audio intelligibility may not capture semantic and temporal coherence across modalities.The paper argues that real-world communication effectiveness often depends on cross-modal coherence.Source: §6 · p. 9
  • Future evaluation should model interactions and alignment among image, audio, and text while reasoning over their combined meaning.The proposed direction goes beyond isolated modality scores toward fusion-aware understanding.Source: §6 · p. 9
  • Bridging aligned modality encoders with language models could support joint interpretation of multimodal inputs and detection of cross-modal inconsistencies.The paper gives examples including explaining visual concepts from audio and visual cues and identifying inconsistencies across modalities.Source: §6 · p. 9

7 Limitation and Future Work

The paper reports two principal constraints: evaluation covers only five academic papers, and PresentAgent currently produces static slides without dynamic animations or effects. Future work targets broader evaluation, dynamic synthesis, and more efficient rendering.

7 Limitation and Future Work

  • Evaluation was limited to five academic papers, potentially underrepresenting the broader document diversity in the benchmark.The limitation is attributed to the high computational costs of commercial LLM and VLM APIs.Source: §7 · p. 9
  • PresentAgent currently generates static slides without dynamic animations or effects because of video-synthesis constraints and speed–quality trade-offs.The paper connects this limitation to architectural constraints and proposes future dynamic-animation improvements.Source: §7 · p. 9
  • Future work will broaden model and document coverage, integrate dynamic animation, and explore lightweight distillation with physics-aware rendering.These directions aim to improve evaluation breadth, generation efficiency, realism, and hardware adaptability.Source: §7 · p. 9

8 Conclusion

PresentAgent introduces a modular approach and benchmark for transforming long-form documents into narrated presentation videos. Its experiments show coherent, informative outputs that closely approach human quality, supporting future controllable multimodal content generation.

8 Conclusion

  • PresentAgent addresses document-to-presentation video generation by coordinating slide planning, narration synthesis, and synchronized rendering.The system is presented as a modular framework for controllable and reusable multimodal outputs across diverse document types.Source: §8 · p. 9
  • The paper introduces document–video pairs and dual assessment strategies based on factual quizzes and preference-based vision-language scoring.These resources support rigorous evaluation of generated presentations.Source: §8 · p. 9
  • Experiments show that PresentAgent produces coherent, engaging, and informative presentations that closely approach human quality.The conclusion frames the system as a foundation for automated, controllable multimodal content generation in education, business, and accessibility.Source: §8 · p. 9
Loading 2507.04036v1…