Despite impressive visual quality, state-of-the-art video diffusion models often generate content that violates real-world physical laws. While existing solutions rely on external priors or specialized data, we investigate the root cause by exploring the internal mechanisms of these models. Specifically, we present the first interpretability study on the ''motion planning'' process of text-to-video diffusion models, revealing how motion trajectories form during early denoising stages. Building upon the ''first shape, then details'' finding, we combine cross-attention trajectory patterns with causal head contributions to identify a specific subset of attention heads driving motion planning. Further, our self-attention analysis shows that Rotary Position Embedding (RoPE) induces excessive spatial attention decay. This causes early candidate regions to prematurely lock into physically implausible positions, suppressing reasonable trajectories in adjacent frames and triggering generation failure modes. To address this fundamental flaw, we propose a lightweight architectural modification that scales the frequency of RoPE across different denoising steps. This strategy reduces excessive attention decay, helping the model explore better candidate regions to establish coherent physical motion. Finally, training-free and training-based experiments confirm the effectiveness of our approach in enhancing the physical commonsense of generated videos.
Recent advances in video diffusion models have greatly improved visual fidelity, yet their generated motions often violate physical plausibility. We observe a common kinematic failure, "motion entanglement", the unintended coupling of independent motion sources, such as camera movement and object motion. We identify that this issue stems from data bias and the reconstruction-based training design of diffusion models. Training on noisy videos that still retain coarse motion cues inadvertently encourages the model to replicate existing motion without an incentive to learn how to model kinematically-grounded motions. To address this, we propose a Self-Imagination Fine-Tuning (SIFT) paradigm, which enables the model to learn from its own generated videos rather than directly reconstructing real ones, breaking the reconstruction shortcut. We further employ motion-aware discriminative supervision and a progressive hard-case replay strategy to stabilize and accelerate learning. By leveraging freely-generated text prompts, our method can densely cover a broad motion space, including rare or finely-disentangled scenarios that would be costly to collect as video data. Extensive experiments demonstrate that our approach substantially improves the physical realism, motion disentanglement, and controllability of generated videos.
Modern video diffusion models generate increasingly realistic and temporally coherent videos, motivating their use as candidate world simulators. Yet it remains unclear whether these models internally encode physical structure, or merely reproduce motion patterns seen during training. We study this question by probing video diffusion models along latent trajectories corresponding to real videos with known physical plausibility. To obtain such trajectories, we approximately invert the deterministic sampling process by integrating the learned velocity field backward from a clean video latent to noise, giving access to the model's intermediate states and attention maps. Using these recovered trajectories, we show that physical plausibility is linearly decodable from diffusion transformer states across IntPhys and InfLevel, reaching around 81.27% average accuracy and outperforming dedicated representation-learning baselines such as V-JEPA and VideoMAE. Surprisingly, this signal is absent from the VAE latent input and emerges inside the denoising transformer itself, despite the model not being trained with a self-supervised predictive objective. These findings suggest that physically meaningful representations can arise as a byproduct of generative denoising.
Video diffusion models (VDMs) have demonstrated remarkable capabilities in synthesizing high-fidelity, photorealistic video content. However, they fundamentally lack an intrinsic understanding of physical laws and frequently produce visually appealing but causally illogical sequences characterized by structural hallucinations and physically implausible dynamics. Injecting physical awareness via training-free test-time optimization is a promising alternative, yet existing methods rely on global gradient updates and rigid scheduling heuristics that inadvertently corrupt passive backgrounds and fail to model complex dynamic state changes. To address this, we propose PhysPlan, a novel training-free guidance framework that shifts the paradigm from stochastic visual interpolation to agentic physics simulation. First, a VLM operates as an iterative cognitive simulator, decomposing multimodal inputs into a Chain-of-Visual-Thought to create a multimodal representation of kinematic trajectories and 3D depth geometries. Second, these signals drives an object-centric test-time optimization. Unlike prior training-free methods that rely on global gradients and rigid scheduling heuristics, PhysPlan introduces Object-Centric Gradient Routing to isolate kinematic modifications and completely lock the passive environment. Furthermore, our Kinetic Intensity Profiling dynamically parameterizes framework hyperparameters to accommodate the varying severity of physical deformations. Extensive evaluations on the PhyGenBench and Physics-IQ benchmarks demonstrate that PhysPlan significantly outperforms both foundational and controllable VDM baselines, offering a promising approach for improving the physical understanding of video generation.