Welcome to the GitHub repository for Awesome-RL-for-Video-Generation! This repository serves as a curated collection of research, resources, and tools related to Reinforcement Learning (RL) for Video Generation. Our goal is to provide an up-to-date and comprehensive overview of RL techniques used in video generation, focusing on the latest advancements. We aim to bridge the gap between RL theory and real-world applications in video generation tasks, offering a solid foundation for future research and development in this field. We hope this repository will serve as a valuable resource for anyone interested in exploring RL applications in video generation!
- [February 14, 2025] We have developed an agent that automatically collects and analyzes the latest papers in the RL-based Video Generation field. It will update the Related Papers daily at 1:00 AM UTC+8.
We are committed to offering researchers the latest advancements in the field. By regularly reviewing and evaluating recent research studies, we ensure that the list of papers stays up-to-date.
| Date | Paper | Contribution | Available Link |
| Jul 2025 | EchoMimicV3: 1.3B Parameters are All You Need for Unified Multi-Modal and Multi-Task Human Animation |
 |
  |
|
• Affiliation: Terminal Technology Department, Alipay, Ant Group • Method Name: EchoMimicV3, Base Model: Wan2.1-FUN-inp-480p-1.3B, Strategy: DPO |
|||
| Jul 2025 | EchoMimicV3: 1.3B Parameters are All You Need for Unified Multi-Modal and Multi-Task Human Animation |
|
|
|
• Affiliation: Terminal Technology Department, Alipay, Ant Group • Method Name: EchoMimicV3, Base Model: Wan2.1-FUN-inp-480p-1.3B, Strategy: DPO |
|||
| Jul 2025 | LongAnimation: Long Animation Generation with Dynamic Global-Local Memory |
|
|
|
• Affiliation: University of Science and Technology of China • Method Name: LongAnimation, Base Model: CogVideoX-1.5-5B, Strategy: NGR |
|||
| Jun 2025 | Video Perception Models for 3D Scene Synthesis |
|
 |
|
• Affiliation: Tsinghua University |
|||
| Jun 2025 | RDPO: Real Data Preference Optimization for Physics Consistency Video Generation |
|
|
|
• Affiliation: Fudan University • Method Name: Real Data Preference Optimization (RDPO), Base Model: LTX-Video-2B, Strategy: DPO |
|||
| Jun 2025 | VQ-Insight: Teaching VLMs for AI-Generated Video Quality Understanding via Progressive Visual Reinforcement Learning |
|
|
|
• Affiliation: School of Electronic and Computer Engineering, Peking University • Method Name: VQ-Insight, Base Model: Qwen-2.5-VL-7B-Instruct, Strategy: GRPO |
|||
| Jun 2025 | Toward Rich Video Human-Motion2D Generation |
 |
 |
|
• Affiliation: Tongji University • Method Name: RVHM2D, Base Model: None, Strategy: Fine-tuning with an FID-based reward • Benchmark Name: Motion2D-Video-150K, Data Number: 150000, Evaluation Metric: R-Precision, FID, MM Dist, Diversity |
|||
| Jun 2025 | AlignHuman: Improving Motion and Fidelity via Timestep-Segment Preference Optimization for Audio-Driven Human Animation |
|
|
|
• Affiliation: ByteDance • Method Name: AlignHuman, Base Model: , Strategy: Timestep-Segment Preference Optimization (TPO) |
|||
| Jun 2025 | Multimodal Large Language Models: A Survey |
|
|
|
• Affiliation: School of Architecture, Technology and Engineering, University of Brighton • Method Name: Video Diffusion Alignment via Reward Gradients, Base Model: , Strategy: Reward Gradients • Method Name: Diffusion Model Alignment Using Direct Preference Optimization, Base Model: , Strategy: Direct Preference Optimization • Method Name: VADER, Base Model: , Strategy: Backpropagating Reward Gradients • Benchmark Name: MJ-VIDEO, Data Number: , Evaluation Metric: Fine-Grained Benchmarking and Rewarding Video Preferences • Benchmark Name: VideoScore, Data Number: , Evaluation Metric: Simulating Fine-grained Human Feedback for Video Generation |
|||
| Jun 2025 | Multimodal Large Language Models: A Survey |
|
|
|
• Affiliation: School of Architecture, Technology and Engineering, University of Brighton, United Kingdom • Method Name: Video Diffusion Alignment via Reward Gradients, Base Model: , Strategy: Reward Gradients • Method Name: Diffusion Model Alignment Using Direct Preference Optimization, Base Model: , Strategy: Direct Preference Optimization (DPO) • Method Name: VADER, Base Model: , Strategy: Backpropagating Reward Gradients • Benchmark Name: MJ-VIDEO, Data Number: , Evaluation Metric: Fine-Grained Video Preferences • Benchmark Name: VideoScore, Data Number: , Evaluation Metric: Simulating Fine-Grained Human Feedback |
|||
| Jun 2025 | Seedance 1.0: Exploring the Boundaries of Video Generation Models |
|
|
|
• Affiliation: ByteDance • Method Name: Human Feedback Alignment (RLHF), Base Model: , Strategy: Reward feedback learning with multiple reward models (Foundational Reward Model, Motion Reward Model, Aesthetic Reward Model) |
|||
| Jun 2025 | Seedance 1.0: Exploring the Boundaries of Video Generation Models |
|
|
|
• Affiliation: ByteDance • Method Name: Human Feedback Alignment (RLHF), Base Model: , Strategy: Reward Maximization with Multi-Dimensional Reward Models |
|||
| Jun 2025 | ContentV: Efficient Training of Video Generation Models with Limited Compute |
|
|
|
• Affiliation: ByteDance Douyin Content Group • Method Name: Reinforcement Learning from Human Feedback (RLHF), Base Model: Stable Diffusion 3.5 Large (SD3.5L), Strategy: RLHF |
|||
| Jun 2025 | ContentV: Efficient Training of Video Generation Models with Limited Compute |
|
|
|
• Affiliation: ByteDance Douyin Content Group • Method Name: Reinforcement Learning from Human Feedback (RLHF), Base Model: Stable Diffusion 3.5 Large (SD3.5L), Strategy: Optimizing conditional distribution pθ(x1|c) with reward model r(c, x1) and KL-divergence regularization |
|||
| May 2025 | Photography Perspective Composition: Towards Aesthetic Perspective Recommendation |
|
|
|
• Affiliation: East China University of Science and Technology • Method Name: Photography Perspective Composition (PPC), Base Model: , Strategy: DPO |
|||
| May 2025 | Scaling Image and Video Generation via Test-Time Evolutionary Search |
|
|
|
• Affiliation: Hong Kong University of Science and Technology • Method Name: EvoSearch, Base Model: , Strategy: |
|||
| May 2025 | InfLVG: Reinforce Inference-Time Consistent Long Video Generation with GRPO |
|
 |
|
• Affiliation: MAPLE Lab, Westlake University • Method Name: InfLVG, Base Model: , Strategy: GRPO • Benchmark Name: CsVBench, Data Number: 1000, Evaluation Metric: HPSv2, Aesthetic Score, CLIP-Flan, ViCLIP, ArcFace-42M, ArcFace-360K, QWen |
|||
| May 2025 | AvatarShield: Visual Reinforcement Learning for Human-Centric Video Forgery Detection |
|
 |
|
• Affiliation: School of Electronic and Computer Engineering, Peking University • Method Name: AvatarShield, Base Model: Qwen2.5-VL-7B, Strategy: GRPO • Benchmark Name: FakeHumanVid, Data Number: 15000, Evaluation Metric: AUC |
|||
| May 2025 | RLVR-World: Training World Models with Reinforcement Learning |
|
|
|
• Affiliation: School of Software, BNRist, Tsinghua University • Method Name: RLVR-World, Base Model: , Strategy: GRPO |
|||
| May 2025 | Temporally-Grounded Language Generation: A Benchmark for Real-Time Vision-Language Models |
|
|
|
• Affiliation: University of Michigan • Benchmark Name: Temporally-Grounded Language Generation (TGLG), Data Number: 16487, Evaluation Metric: TRACE |
|||
| May 2025 | Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models |
|
 |
|
• Affiliation: MMLab, CUHK, Hong Kong • Method Name: Negative Preference Optimization (NPO), Base Model: , Strategy: Diffusion-NPO |
|||
| May 2025 | DanceGRPO: Unleashing GRPO on Visual Generation |
|
|
|
• Affiliation: ByteDance Seed • Method Name: DanceGRPO, Base Model: , Strategy: GRPO |
|||
| May 2025 | VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding |
|
|
|
• Affiliation: University of Maryland, College Park • Benchmark Name: VideoHallu, Data Number: 3000, Evaluation Metric: |
|||
| Apr 2025 | TesserAct: Learning 4D Embodied World Models |
|
|
|
• Affiliation: UMass Amherst • Method Name: TesserAct, Base Model: , Strategy: |
|||
| Apr 2025 | Reasoning Physical Video Generation with Diffusion Timestep Tokens via Reinforcement Learning |
|
|
|
• Affiliation: Zhejiang University • Method Name: Phys-AR, Base Model: Llama3.1-8B, Strategy: GRPO |
|||
| Apr 2025 | SkyReels-V2: Infinite-length Film Generative Model |
|
 |
|
• Affiliation: Skywork AI • Method Name: SkyReels-V2, Base Model: Qwen2-VL-7B, Strategy: DPO |
|||
| Apr 2025 | FingER: Content Aware Fine-grained Evaluation with Reasoning for AI-Generated Videos |
|
|
|
• Affiliation: AMAP, Alibaba Group • Method Name: FingER, Base Model: Qwen2.5-VL, Strategy: GRPO • Benchmark Name: FingER-Instruct-60k, Data Number: 60000, Evaluation Metric: |
|||
| Apr 2025 | Aligning Anime Video Generation with Human Feedback |
|
|
|
• Affiliation: Fudan University • Method Name: Gap-Aware Preference Optimization (GAPO), Base Model: , Strategy: Direct Preference Optimization (DPO) • Benchmark Name: AnimeReward, Data Number: 30000, Evaluation Metric: multi-dimensional reward scores |
|||
| Apr 2025 | Discriminator-Free Direct Preference Optimization for Video Diffusion |
|
|
|
• Affiliation: Zhejiang University • Method Name: Discriminator-Free Video Preference Optimization (DF-VPO), Base Model: , Strategy: DPO |
|||
| Apr 2025 | Morpheus: Benchmarking Physical Reasoning of Video Generative Models with Real Physical Experiments |
|
|
|
• Affiliation: University of Trento, Italy • Benchmark Name: Morpheus, Data Number: 80, Evaluation Metric: Dynamical Score |
|||
| Apr 2025 | OmniCam: Unified Multimodal Video Generation via Camera Control |
|
|
|
• Affiliation: Zhejiang University • Method Name: OmniCam, Base Model: Llama3.1, Strategy: PPO • Benchmark Name: OmniTr, Data Number: 1000 trajectories, 10,000 descriptions, 30,000 videos, Evaluation Metric: Mstarttime, Mendtime, Mspeed, Mrotate, Mdirection |
|||
| Mar 2025 | VPO: Aligning Text-to-Video Generation Models with Prompt Optimization |
|
 |
|
• Affiliation: The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University • Method Name: VPO, Base Model: LLaMA3-8B-Instruct, Strategy: DPO |
|||
| Mar 2025 | Zero-Shot Human-Object Interaction Synthesis with Multimodal Priors |
|
|
|
• Affiliation: The University of Hong Kong • Method Name: Physics-based HOI Refinement, Base Model: , Strategy: Actor-Critic with Gaussian Policy |
|||
| Mar 2025 | Judge Anything: MLLM as a Judge Across Any Modality |
|
|
|
• Affiliation: Huazhong University of Science and Technology • Benchmark Name: TASKANYTHING, Data Number: 1500, Evaluation Metric: • Benchmark Name: JUDGE ANYTHING, Data Number: 9000, Evaluation Metric: Agreement, Pearson correlation, Spearman correlation, MAE, Accuracy |
|||
| Mar 2025 | MagicID: Hybrid Preference Optimization for ID-Consistent and Dynamic-Preserved Video Customization |
|
|
|
• Affiliation: Zhejiang University • Method Name: MagicID, Base Model: , Strategy: DPO |
|||
| Mar 2025 | Unified Reward Model for Multimodal Understanding and Generation |
|
 |
|
• Affiliation: Fudan University • Method Name: UnifiedReward, Base Model: LLaVA-OneVision-7B, Strategy: DPO |
|||
| Feb 2025 | Pre-Trained Video Generative Models as World Simulators |
|
|
|
• Affiliation: Hong Kong University of Science and Technology • Method Name: Dynamic World Simulation (DWS), Base Model: , Strategy: PPO |
|||
| Feb 2025 | Harness Local Rewards for Global Benefits: Effective Text-to-Video Generation Alignment with Patch-level Reward Models |
|
|
|
• Affiliation: Carnegie Mellon University • Method Name: HALO, Base Model: , Strategy: DPO |
|||
| Feb 2025 | IPO: Iterative Preference Optimization for Text-to-Video Generation |
|
|
|
• Affiliation: Shanghai Academy of Artificial Intelligence for Science • Method Name: Iterative Preference Optimization (IPO), Base Model: , Strategy: DPO |
|||
| Feb 2025 | MJ-VIDEO: Fine-Grained Benchmarking and Rewarding Video Preferences in Video Generation |
|
|
|
• Affiliation: UNC-Chapel Hill • Benchmark Name: MJ-BENCH-VIDEO, Data Number: 5421, Evaluation Metric: |
|||
| Feb 2025 | HuViDPO:Enhancing Video Generation through Direct Preference Optimization for Human-Centric Alignment |
|
|
|
• Affiliation: Zhejiang University • Method Name: HuViDPO, Base Model: , Strategy: DPO |
|||
| Feb 2025 | Zeroth-order Informed Fine-Tuning for Diffusion Model: A Recursive Likelihood Ratio Optimizer |
|
 |
|
• Affiliation: Guanghua School of Management, Peking University • Method Name: Recursive Likelihood Ratio (RLR) optimizer, Base Model: , Strategy: |
|||
| Jan 2025 | Improving Video Generation with Human Feedback |
|
|
|
• Affiliation: The Chinese University of Hong Kong • Method Name: Flow-DPO, Base Model: , Strategy: DPO • Method Name: Flow-RWR, Base Model: , Strategy: RWR • Method Name: Flow-NRG, Base Model: , Strategy: Reward Guidance • Benchmark Name: VideoGen-RewardBench, Data Number: 26500, Evaluation Metric: |
|||
| Dec 2024 | VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation |
|
 |
|
• Affiliation: Tsinghua University • Method Name: Multi-Objective Preference Optimization (MPO), Base Model: , Strategy: DPO |
|||
| Dec 2024 | OnlineVPO: Align Video Diffusion Model with Online Video-Centric Preference Optimization |
|
|
|
• Affiliation: The University of Hong Kong • Method Name: OnlineVPO, Base Model: , Strategy: DPO |
|||
| Dec 2024 | VideoDPO: Omni-Preference Alignment for Video Diffusion Generation |
|
|
|
• Affiliation: HKUST • Method Name: VideoDPO, Base Model: , Strategy: DPO |
|||
| Dec 2024 | FLIP: Flow-Centric Generative Planning for General-Purpose Manipulation Tasks |
|
|
|
• Affiliation: National University of Singapore • Method Name: FLIP, Base Model: , Strategy: model-based planning |
|||
| Dec 2024 | The Matrix: Infinite-Horizon World Generation with Real-Time Moving Control |
|
|
|
• Affiliation: Tongyi Lab • Method Name: The Matrix, Base Model: , Strategy: Shift-Window Denoising Process Model (Swin-DPM) |
|||
| Dec 2024 | Improving Dynamic Object Interactions in Text-to-Video Generation with AI Feedback |
|
|
|
• Affiliation: The University of Tokyo • Method Name: RL-Finetuning for Text-to-Video Models, Base Model: , Strategy: RWR, DPO |
|||
| Nov 2024 | Free$^2$Guide: Gradient-Free Path Integral Control for Enhancing Text-to-Video Generation with Large Vision-Language Models |
|
|
|
• Affiliation: Kim Jaechul Graduate School of AI, KAIST • Method Name: Free2Guide, Base Model: , Strategy: Path Integral Control |
|||
| Nov 2024 | A Reinforcement Learning-Based Automatic Video Editing Method Using Pre-trained Vision-Language Model |
|
|
|
• Affiliation: SSE, The Chinese University of Hong Kong, Shenzhen • Method Name: RL-based editing framework, Base Model: , Strategy: actor-critic |
|||
| Oct 2024 | Video to Video Generative Adversarial Network for Few-shot Learning Based on Policy Gradient |
|
|
|
• Affiliation: Northwestern University • Method Name: RL-V2V-GAN, Base Model: , Strategy: Policy Gradient |
|||
| Oct 2024 | WorldSimBench: Towards Video Generation Models as World Simulators |
|
|
|
• Affiliation: The Chinese University of Hong Kong, Shenzhen • Benchmark Name: WorldSimBench, Data Number: 35701, Evaluation Metric: Human Preference Evaluator |
|||
| Oct 2024 | Animating the Past: Reconstruct Trilobite via Video Generation |
|
|
|
• Affiliation: AI Lab, Yishi Inc. • Method Name: Automatic T2V Prompt Learning Method, Base Model: , Strategy: KTO |
|||
| Oct 2024 | VideoAgent: Self-Improving Video Generation |
|
|
|
• Affiliation: University of Waterloo • Method Name: VideoAgent, Base Model: , Strategy: self-improvement through online finetuning |
|||
| Oct 2024 | E-Motion: Future Motion Simulation via Event Sequence Diffusion |
|
 |
|
• Affiliation: Xidian University • Method Name: Event-Sequence Diffusion Network, Base Model: , Strategy: PPO |
|||
| Oct 2024 | DART: A Diffusion-Based Autoregressive Motion Model for Real-Time Text-Driven Motion Control |
|
|
|
• Affiliation: ETH Z__rich • Method Name: DART, Base Model: , Strategy: PPO |
|||
| Oct 2024 | SePPO: Semi-Policy Preference Optimization for Diffusion Alignment |
|
 |
|
• Affiliation: University of Rochester • Method Name: SePPO, Base Model: , Strategy: DPO |
|||
| Jul 2024 | Video Diffusion Alignment via Reward Gradients |
|
|
|
• Affiliation: Carnegie Mellon University • Method Name: VADER, Base Model: , Strategy: Reward Gradients |
|||
| Dec 2023 | InstructVideo: Instructing Video Diffusion Models with Human Feedback |
|
|
|
• Affiliation: Zhejiang University • Method Name: InstructVideo, Base Model: , Strategy: reward fine-tuning |
|||
| Nov 2023 | AdaDiff: Adaptive Step Selection for Fast Diffusion Models |
|
|
|
• Affiliation: Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University • Method Name: AdaDiff, Base Model: , Strategy: policy gradient |
|||
If you have a paper or are aware of relevant research that should be incorporated, please contribute via pull requests, issues, email, or other suitable methods.