PhysMaster: Mastering Physical Representation for Video Generation via Reinforcement Learning

1 INTRODUCTION

• Video generators often struggle to adhere to physical laws, limiting physically plausible generation and their potential to serve as world models.The paper identifies appearance-focused MSE supervision and the difficulty of directly supervising physical performance as central challenges.Source: §1 · p. 1
• PhysEncoder extracts physical representations from input-image cues, including object configurations, relative positions, and potential interactions, to guide image-to-video generation.The input image is treated as a source of physical priors for predicting subsequent dynamics.Source: §1 · p. 2
• PhysMaster uses a three-stage RLHF pipeline: SFT followed by two-stage DPO on pairwise video preferences to optimize the physical representation.Generated-video plausibility provides feedback for learning representations when they lack a straightforward direct supervision target.Source: §1 · p. 2
• The authors report that PhysEncoder guides improved physical performance from a free-fall proxy task to open-world scenarios governed by diverse physical laws.They present the representation-learning approach as plug-in conditioning intended to generalize beyond specific phenomena or motion modalities.Source: §1 · p. 2
• PhysMaster is presented as a generalizable solution for capturing physical knowledge across diverse phenomena and as a potential foundation for physics-aware video generation.The conclusion also points to potential broader applications.Source: §1 · p. 2

2 RELATED WORKS

• Simulation-based approaches guide video generation with simulation results, while simulation-free approaches use large-scale video finetuning or feedback from annotators or vision-language models.Examples include PhysGen and PhysMotion for simulation-based methods, and PhyT2V and WISA for simulation-free methods.Source: §2 · pp. 2–3
• The paper characterizes existing approaches as constrained to fixed physical categories or limited in physical comprehension.Its introduction also notes that simulation methods may be limited in simulable processes and modalities, while data fitting can depend heavily on similar training samples.Source: §2 · p. 3
• Unlike prior video-generation RLHF methods, PhysMaster uses generated-video feedback to optimize a physical encoder rather than the whole generation model.The paper presents this design as a way to mitigate overfitting to specific physical processes and promote generalizable physical knowledge.Source: §2 · p. 3

3 METHOD

• PhysMaster formulates image-to-video generation as predicting plausible dynamics from an initial image and text, with physical representations providing extra guidance.The method overview describes both ground-truth supervision through SFT and pairwise supervision from generated videos through DPO.Source: §3 · p. 3

3.1 PHYSICAL REPRESENTATION

• Input images provide explicit physical states and implicit cues such as gravity, which PhysEncoder converts into embeddings concatenated with image embeddings for DiT conditioning.PhysEncoder uses a semantic-perception encoder and a physical head to adapt features to the dimension required by DiT.Source: §3.1 · pp. 3–4

3.2 TASK FORMULATION

• The proxy task is free-fall, where deterministic trajectories under gravity support synthetic data generation and automatic preference-pair construction against ground truth.The task also covers object properties such as density, elasticity, and hardness.Source: §3.2 · p. 4
• For open-world scenarios, domain-specific prompt prefixes identify the relevant physics, while human annotators provide pairwise preference labels for DPO data and evaluation.The scenarios are intended to cover common physical phenomena observed in the real world.Source: §3.2 · p. 4

3.3 TRAINING SCHEME

• Stage I uses SFT to condition the I2V model on PhysEncoder representations, establishing an initial ability to predict subsequent frames.The embeddings are concatenated with visual embeddings as an additional model condition.Source: §3.3 · p. 4
• Stage II applies DPO to DiT using preferred and less-preferred video pairs; Stage III freezes DiT and optimizes PhysEncoder’s physical head using the same objective.Stage II adapts the generated-video distribution before its feedback is used to optimize the encoder in Stage III.Source: §3.3 · p. 5

4.1 IMPLENTATION DETAILS

• Evaluation uses Kubric free-fall data and WISA-80K’s 17 event types across three physics branches; open-world testing uses 344 VIDEOPHY prompts and reports PC and SA.Source: §4.1 · pp. 5–6

4.2 EVALUATION ON PROXY TASK

• On the real-world PisaBench subset, PhysMaster outperforms PhysGen and PISA overall, with stronger IoU than PISA while maintaining competitive trajectory accuracy.Source: §4.2 · p. 6, Figure 2 · p. 6, Table 2 · p. 7
• SFT gives the free-fall model preliminary motion prediction, while later DPO stages improve physical plausibility and PhysEncoder’s guidance; SFT alone is insufficient for PhysEncoder.Source: §4.2 · pp. 6–7, Table 1 · p. 6, Figure 3 · p. 7
• Stage III physical features group objects under the same external forces and distinguish materials more clearly than Stage I features.Source: §4.2 · p. 7, Figure 4 · p. 7

4.3 GENERALIZATION ON GENERAL OPEN-WORLD SCENARIOS

• 70x faster than PhyT2V and 8x faster than WISA; Stage III achieves state-of-the-art SA and PC, generating a 5-second video in 26 seconds on one A800 GPU.Source: Table 3 · p. 9, §4.3 · p. 9
• Ablations on open-world scenarios find that DPO after Stage I improves Stage III physical coherence, and annotators prefer Stage III to Stage I and the I2V base model.Source: Figure 6 · p. 8, §4.3 · p. 9, Table 4 · p. 9

• PhysMaster learns physical representations from input images to guide image-to-video models toward physically plausible videos.Source: §5 · p. 9
• Generative feedback from a pretrained video model and DPO optimize the physical encoder, enhancing physical accuracy and generalizability across physical processes.Source: §5 · p. 9
• Injecting physical knowledge is presented as a generic, plug-in approach for physics-aware video generation and broader applications.Source: §5 · p. 10
• Real-world DPO preference datasets rely on costly, time-consuming human annotation; the authors report that 500 human-labeled examples mitigate this limitation.Source: §5 · p. 10