Prompt-A-Video: Prompt Your Video Diffusion Model via Preference-Aligned LLM

TL;DR

Prompt-A-Video leverages preference-aligned LLMs with reward-guided evolution, boosting video quality metrics by over 0.2 on average across models.

cs.CV 🔴 Advanced 2024-12-20 21 citations 64 views
Yatai Ji Jiacheng Zhang Jie Wu Shilong Zhang Shoufa Chen Chongjian GE Peize Sun Weifeng Chen Wenqi Shao Xuefeng Xiao Weilin Huang Ping Luo
video generation prompt optimization large language models preference alignment deep learning

Key Findings

Methodology

This work introduces a two-stage automatic prompt optimization framework combining reward-guided evolution and preference alignment. First, a multi-dimensional reward system evaluates generated videos, constructing a pool of high-quality prompts. Then, leveraging evolutionary algorithms and GPT-4o, the system automatically refines prompts to align with model preferences. Fine-tuning the large language model via supervised fine-tuning (SFT) and direct preference optimization (DPO) further enhances its ability to generate video-centric prompts. Extensive experiments on WebVid and VBench datasets demonstrate significant improvements in multiple quality metrics, with average gains exceeding 0.2 points over baseline prompts. The framework effectively addresses challenges such as modality inconsistency, high annotation costs, and model unawareness, providing a scalable solution for diverse text-to-video models.

Key Results

  • On WebVid, prompts refined through DPO-2 improved the average visual quality (VQ) score from 3.079 to 3.254 for Open-Sora 1.2, and similar improvements were observed in temporal consistency and dynamic degree metrics. The overall video quality metrics showed an average increase of over 0.2 points, indicating clearer, more coherent, and more engaging videos.
  • In VBench, the proposed method enhanced subject and background consistency, motion smoothness, and aesthetic quality, with average improvements exceeding 0.15. Human subjective evaluations revealed over 80% win rates for prompts optimized with Prompt-A-Video, confirming its superiority in perceptual quality.
  • The cross-model generalization experiments demonstrated that prompts optimized for one model could effectively transfer to others, indicating robustness. The multi-round DPO training further refined the prompts, leading to consistent performance gains across different architectures and datasets.

Significance

This research marks a significant advancement in automated prompt engineering for video synthesis. By integrating reward-based evaluation and preference alignment within a large language model framework, it reduces reliance on manual prompt crafting and domain expertise. The approach addresses longstanding issues such as modality mismatch and high annotation costs, paving the way for more accessible, scalable, and high-quality video content creation. Its implications extend to industries like entertainment, virtual reality, and gaming, where rapid, high-fidelity video generation is crucial. Furthermore, the methodology offers a blueprint for future multi-modal content optimization, fostering innovation in AI-driven creative workflows.

Technical Contribution

The core technical innovation lies in the combination of multi-dimensional reward evaluation, evolutionary prompt refinement, and large-scale language model fine-tuning. The reward system assesses video quality across visual, temporal, and semantic dimensions, guiding the evolution of prompts. The use of GPT-4o as an evolutionary operator enables automatic, model-specific prompt enhancement. The two-stage training pipeline—supervised fine-tuning with curated prompt pairs and DPO for preference alignment—ensures that the language model learns to generate prompts that are both effective and aligned with model preferences. This integrated framework surpasses existing static or manually crafted prompt methods, offering a dynamic, scalable solution for multi-model video synthesis.

Novelty

This work is the first to introduce an automated, preference-aware prompt optimization system specifically tailored for video generation tasks. Unlike prior image-focused prompt refinement methods, it explicitly considers the dynamic and temporal aspects of videos. The integration of reward-guided evolution with large language models for prompt enhancement is a novel approach, addressing the core challenges of modality mismatch, high annotation costs, and model unawareness. The framework's ability to generate model-preferred prompts without manual intervention represents a significant leap forward in AI content creation.

Limitations

  • The computational cost of multi-round evolution and reward evaluation remains high, especially for large-scale video datasets, limiting real-time applications.
  • The framework's effectiveness depends on the quality and diversity of the reward models; in scenarios with limited or biased rewards, performance may degrade.
  • While the method generalizes across different models, extreme or novel content types may still pose challenges, requiring further adaptation or retraining.

Future Work

Future research will focus on improving the efficiency of the prompt evolution process, possibly through reinforcement learning or meta-learning techniques. Expanding the reward system to incorporate more nuanced dynamic and semantic metrics could further enhance video quality. Additionally, integrating multimodal cues such as audio and text descriptions will enable richer content generation. Exploring real-time prompt adaptation and extending the framework to longer, more complex videos are promising directions. Ultimately, the goal is to develop a fully autonomous, scalable system capable of producing diverse, high-fidelity videos tailored to user preferences and application needs.

AI Executive Summary

The rapid development of AI-driven video synthesis has revolutionized digital content creation, yet challenges remain in producing high-quality, user-aligned videos efficiently. Traditional methods rely heavily on manual prompt engineering, which is labor-intensive and often yields inconsistent results. Existing automatic approaches tend to struggle with modality mismatch, high annotation costs, and model unawareness, especially in the complex domain of text-to-video generation. These limitations hinder the widespread adoption of AI-generated videos in entertainment, education, and virtual reality applications.

In response, this study introduces Prompt-A-Video, a novel framework that leverages large language models (LLMs) guided by reward-based evaluation and preference alignment to automatically generate optimized prompts for video diffusion models. The core innovation is a two-stage process: first, constructing a multi-dimensional reward system that assesses video quality across visual fidelity, temporal coherence, and semantic alignment; second, employing a reward-guided evolutionary algorithm combined with GPT-4o to iteratively refine prompts. This process produces a high-quality prompt pool tailored to specific models, which is then used to fine-tune the LLM via supervised learning and direct preference optimization (DPO). The result is a model capable of autonomously generating prompts that are both effective and aligned with the preferences of different video diffusion architectures.

Extensive experiments on WebVid and VBench datasets demonstrate the effectiveness of Prompt-A-Video. Quantitative metrics such as VQ, TC, and DD show average improvements exceeding 0.2 points after DPO-2 optimization, indicating clearer, more coherent, and more dynamic videos. Human evaluations further confirm that videos generated with prompts refined by Prompt-A-Video outperform those from baseline prompts and other state-of-the-art methods, with win rates surpassing 80%. These results highlight the framework’s robustness, generalizability, and potential to significantly advance automated video content creation.

The broader impact of this work lies in democratizing high-quality video synthesis, reducing reliance on expert knowledge, and enabling scalable content production. It opens avenues for AI-assisted filmmaking, personalized virtual environments, and real-time content adaptation. Nonetheless, challenges such as computational costs and the need for more nuanced reward systems remain. Future efforts will aim to optimize efficiency, incorporate multimodal feedback, and extend the approach to longer, more complex videos. Overall, Prompt-A-Video represents a pivotal step toward autonomous, preference-aware AI content generation, promising to reshape the landscape of digital media production.

Deep Dive

Abstract

Text-to-video models have made remarkable advancements through optimization on high-quality text-video pairs, where the textual prompts play a pivotal role in determining quality of output videos. However, achieving the desired output often entails multiple revisions and iterative inference to refine user-provided prompts. Current automatic methods for refining prompts encounter challenges such as Modality-Inconsistency, Cost-Discrepancy, and Model-Unaware when applied to text-to-video diffusion models. To address these problem, we introduce an LLM-based prompt adaptation framework, termed as Prompt-A-Video, which excels in crafting Video-Centric, Labor-Free and Preference-Aligned prompts tailored to specific video diffusion model. Our approach involves a meticulously crafted two-stage optimization and alignment system. Initially, we conduct a reward-guided prompt evolution pipeline to automatically create optimal prompts pool and leverage them for supervised fine-tuning (SFT) of the LLM. Then multi-dimensional rewards are employed to generate pairwise data for the SFT model, followed by the direct preference optimization (DPO) algorithm to further facilitate preference alignment. Through extensive experimentation and comparative analyses, we validate the effectiveness of Prompt-A-Video across diverse generation models, highlighting its potential to push the boundaries of video generation.

cs.CV cs.CL cs.MM

References (20)

CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

Zhuoyi Yang, Jiayan Teng, Wendi Zheng et al.

2024 2312 citations ⭐ Influential View Analysis →

Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers

Qingyan Guo, Rui Wang, Junliang Guo et al.

2024 376 citations ⭐ Influential

Differential Evolution – A Simple and Efficient Heuristic for global Optimization over Continuous Spaces

R. Storn, K. Price

1997 28266 citations ⭐ Influential

JADE: Adaptive Differential Evolution With Optional External Archive

Jingqiao Zhang, A. Sanderson

2009 3496 citations ⭐ Influential

GPT-4 Technical Report

OpenAI Josh Achiam, Steven Adler, S. Agarwal et al.

2023 26898 citations ⭐ Influential View Analysis →

Optimizing Prompts for Text-to-Image Generation

Y. Hao, Zewen Chi, Li Dong et al.

2022 290 citations ⭐ Influential View Analysis →

Dynamic Prompt Optimizing for Text-to-Image Generation

Wenyi Mo, Tianyu Zhang, Yalong Bai et al.

2024 54 citations ⭐ Influential View Analysis →

Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval

Max Bain, Arsha Nagrani, Gül Varol et al.

2021 1649 citations View Analysis →

CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers

Wenyi Hong, Ming Ding, Wendi Zheng et al.

2022 1169 citations View Analysis →

Design Guidelines for Prompt Engineering Text-to-Image Generative Models

Vivian Liu, Lydia B. Chilton

2021 747 citations View Analysis →

Learning Transferable Visual Models From Natural Language Supervision

Alec Radford, Jong Wook Kim, Chris Hallacy et al.

2021 54858 citations View Analysis →

MAP: Multimodal Uncertainty-Aware Vision-Language Pre-training Model

Yatai Ji, Junjie Wang, Yuan Gong et al.

2022 83 citations View Analysis →

DiffusionDB: A Large-scale Prompt Gallery Dataset for Text-to-Image Generative Models

Zijie J. Wang, Evan Montoya, David Munechika et al.

2022 506 citations View Analysis →

MagicVideo: Efficient Video Generation With Latent Diffusion Models

Daquan Zhou, Weimin Wang, Hanshu Yan et al.

2022 512 citations View Analysis →

Seeing What You Miss: Vision-Language Pre-training with Semantic Completion Learning

Yatai Ji, Rong-Cheng Tu, Jie Jiang et al.

2022 18 citations View Analysis →

Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation

Jay Zhangjie Wu, Yixiao Ge, Xintao Wang et al.

2022 1201 citations View Analysis →

Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators

L. Khachatryan, A. Movsisyan, Vahram Tadevosyan et al.

2023 872 citations View Analysis →

Visual Instruction Tuning

Haotian Liu, Chunyuan Li, Qingyang Wu et al.

2023 10911 citations View Analysis →

Align Your Latents: High-Resolution Video Synthesis with Latent Diffusion Models

A. Blattmann, Robin Rombach, Huan Ling et al.

2023 1768 citations View Analysis →

SUR-adapter: Enhancing Text-to-Image Pre-trained Diffusion Models with Large Language Models

Shan Zhong, Zhongzhan Huang, Wushao Wen et al.

2023 62 citations View Analysis →

Cited By (20)

Bridging Creative Intent and Visual Quality: Creator-Driven Recurrent Video Generation with Agentic Feedback Loops

2026 ⭐ Influential View Analysis →

Causally Steered Diffusion for Automated Video Counterfactual Generation

2025 6 citations ⭐ Influential View Analysis →

LINA: Learning INterventions Adaptively for Physical Alignment and Generalization in Diffusion Models

Empowering Video Translation using Multimodal Large Language Models

2026

Goku: Flow Based Video Generative Foundation Models

2025 78 citations View Analysis →

FlashVideo: Flowing Fidelity to Detail for Efficient High-Resolution Video Generation

2025 37 citations View Analysis →

VPO: Aligning Text-to-Video Generation Models with Prompt Optimization

2025 24 citations View Analysis →

Seedance 1.0: Exploring the Boundaries of Video Generation Models

2025 243 citations View Analysis →

Otta: One-Shot Video Tuning with Temporal Adapters on Diffusion

2025

VISTA: A Test-Time Self-Improving Video Generation Agent

2025 20 citations View Analysis →

Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization

2025 2 citations View Analysis →

A Reason-then-Describe Instruction Interpreter for Controllable Video Generation

Beyond Trial-and-Error: Agentic Optimization for Image-to-Video Adherence

Talking with Actionbits—A Part-Enhanced VLM for Action and Interaction Recognition in Animals

2026

PhyPrompt: RL-based Prompt Refinement for Physically Plausible Text-to-Video Generation

2026 1 citations View Analysis →

BrandFusion: A Multi-Agent Framework for Seamless Brand Integration in Text-to-Video Generation

2026 2 citations View Analysis →

VQQA: An Agentic Approach for Video Evaluation and Quality Improvement

2026 3 citations View Analysis →

P-Flow: Prompting Visual Effects Generation

When Cultures Move: Measuring and Improving Multicultural Text-to-Video Generation

Flux-OPD: On-Policy Distillation with Evolving Contexts