PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation

Conference: ECCV 2026
arXiv: 2512.24551
Code: https://github.com/caiyuanhao1998/Open-PhyGDPO
Area: Video Generation
Keywords: Text-to-Video Generation, Physical Consistency, Direct Preference Optimization, Plackett-Luce Model, Group-Wise Preference Learning

TL;DR

PhyGDPO proposes a group-wise direct preference optimization framework for physically consistent video generation. It utilizes real-world videos as winning samples and a Plackett-Luce group-wise probability model instead of pairwise comparisons to capture global preferences. Co-designed with VLM-guided physical reward reweighting and a LoRA-switching reference mechanism, it efficiently achieves superior physical consistency over Sora2 and Veo3.1 at the 14B parameter scale.

Background & Motivation

Text-to-Video (T2V) generation has made breakthrough progress recently, with models like Wan2.1, Sora2, and Veo3.1 capable of generating videos with extremely high visual quality. However, accurately and consistently modeling physical laws in videos—such as joint coordination in human motion, trajectories in ball collisions, and stress propagation during glass shattering—remains a highly challenging and under-explored problem. Improving the physical reasoning capabilities of video generation models to bring them closer to general-world physics simulators has profound implications for games, film production, autonomous driving, and even robotics control.

Currently, technical routes for improving physical modeling are roughly divided into two categories. The first is graphics engine-based approaches, which rely on simulators to specify physical parameters in simple scenes (such as perfect elastic collisions or rigid body dynamics), but struggle to parameterize as environment complexity scales. The second is LLM prompt expansion-based methods, which let LLMs explicitly include descriptions of physical laws in the input prompt and simulate physical effects through iterative generation or fine-tuning. These methods outsource physical reasoning to the LLM. However, T2V models themselves have limited prompt-following capabilities, and LLMs' physical knowledge is far from perfect, which may introduce erroneous guidance. More fundamentally, even Sora2 and Veo3.1, pre-trained on massive datasets, still frequently fail on complex human motions and physical phenomena because the training data lack negative exemplars to provide the contrastive signal of "what is physically wrong."

Direct Preference Optimization (DPO) offers a promising solution, but directly applying it incurs three specific difficulties. First, there is a lack of paired training data that can comprehensively capture various physical interactions. Second, standard DPO treats conditionally generated videos as winning samples, but the physical realism of generated videos is inherently limited. More critically, DPO relies on the Bradley-Terry model for pairwise comparisons, which cannot capture global preference signals like physical plausibility that require a multi-candidate global perspective. Third, standard DPO requires duplicating a full 14B model as a reference, which imposes an unbearable burden on GPU memory and computational efficiency. This work systematically addresses these issues from the three dimensions of data, modeling, and engineering. Core Idea: Extend physical preference optimization in video generation from pairwise Bradley-Terry comparison to a Plackett-Luce group-wise probability model using real-world videos as winning samples. It performs DPO post-training of a 14B model efficiently under a 25GB VRAM footprint using VLM physical score-guided hard-case reweighting and a LoRA-Switch Reference mechanism.

Method

Overall Architecture

PhyGDPO consists of two main parts: data construction (PhyAugPipe) and preference optimization (PhyGDPO). On the data side, starting from a million-scale T2V data pool, PhyAugPipe uses Qwen2.5-72B VLM with Chain-of-Thought (CoT) rules to filter 135K text-video pairs rich in physical interactions. It then clusterizes action semantics and evaluates generation difficulty using VideoCon-Physics, sampling a final 17K training subset based on difficulty index. On the training side, real-world videos serve as winning samples, and multiple candidate generations as losing samples, which are optimized under the PL group-wise probability model. Physics-Guided Rewarding (PGR) dynamically adjusts loss weights of losing samples based on scores from VideoCon-Physics. LoRA-Switch Reference (LoRA-SR) freezes the backbone as the reference model, appends trainable LoRA modules to the attention layers, and flexibly switches between reference and training modes through an environment manager.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Million-scale T2V Data Pool"] --> B["PhyAugPipe:<br/>VLM + CoT Filtering"]
    B --> C["135K Physics-Rich<br/>Text-Video Pairs"]
    C --> D["Action Semantic Clustering"]
    D --> E["PGR Difficulty Sampling"]
    E --> F["17K Training Subset"]

    F --> G["Model Generates<br/>Multiple Candidate Videos"]
    G --> H["Plackett-Luce<br/>Group-wise Probability Model"]
    H --> I["PGR: VLM Scoring<br/>Dynamic Reweighting"]
    I --> J["LoRA-SR:<br/>Frozen Backbone + LoRA Switch"]
    J --> K["PhyGDPO Model"]

    F2["Real Video<br/>(Winning Sample)"] --> H

Key Designs

1. PhyAugPipe: Filtering Physics-Rich Interaction Videos from Massive Data via VLM+CoT

Physics-interaction videos are scarce because determining whether a video contains meaningful physical interaction is not trivial to automate—one cannot simply check if something is moving, but must evaluate whether the motion-object interaction embeds mechanical logic. The core idea of PhyAugPipe is to delegate this decision to a VLM (Qwen2.5-72B) with chain-of-thought capabilities, processing each video via a 5-step CoT. First, the VLM parses entities, their materials, actions, and force interactions from the prompt and video frames. Second, it cross-validates parsing results with video frames to eliminate hallucinations. Third, it writes physical reasoning descriptions based on reliable parsing, explaining how mechanical interactions cause outcomes. Fourth, it calculates a physical richness score (\(0\)-\(1\)) based on entity count, interaction complexity, force clarity, and causal transparency, performing binary classification with a \(0.6\) threshold. Fifth, it expands original prompts with causal physical details. Unlike rule-based templates, this CoT-driven filtering understands distinct physical mechanisms (e.g., a curved soccer kick vs. a shattering glass falling off a table). Following this step, 135K pairs are retained from 1 million. Then, sentence Transformer action semantic clustering is used to balance coverage across motion categories, and VideoCon-Physics evaluates action generation difficulty. Data budget is tilted toward difficult classes using exponential sampling weights \(r_k = \exp(\tau d_k)\), where \(d_k = 1 - \mathcal{S}_f(k)\) represents category-level physical score. Notably, PhyAugPipe is a filtering pipeline—the VLM only influences physical richness determination and sampling ratios, and the original prompt and real-world video frames are not rewritten/modified by the VLM, avoiding VLM-induced bias.

2. Groupwise DPO: PL Groupwise Probability Model with Real-world Videos as Winning Samples

Standard DPO has two fundamental limitations for physical plausibility optimization. First is contrastive granularity: the Bradley-Terry model can only handle one-to-one comparisons, requiring averaging over pairwise combinations for multiple candidates, which discards global ranking information. Second is incorrect winning signals: standard DPO treats generated videos as winning samples, but the physical realism of generations is unreliable. PhyGDPO solves both problems with a single design: using real-world videos as winning samples (which naturally obey physical laws, ensuring flawless learning signals), and describing the probability of selecting the winner from a candidate set of 1 winner and \(m\) losers via the Plackett-Luce (PL) model: $$ p_{\text{PL}}(x^w_0 \mid \mathcal{G}^l(c), c) = \frac{\exp(r(c, x^w_0))}{\exp(r(c, x^w_0)) + \sum_{j=1}^m \exp(r(c, x^{l_j}_0))} $$ This probability directly learns the global preference that "real-world videos should be chosen as best." By substituting DPO's implicit reward formulation into the PL model, the loss function extends from pairwise to group-wise. To control training efficiency, the paper approximates the group-wise training upper bound into a step-wise, pairwise form using Jensen's inequality and a constructed inequality \(\log(1 + \sum e^{x_j}) \le \sum \gamma_j \log(1 + e^{\alpha_j x_j})\). The final loss is equivalent to encouraging the predicted velocity field \(v_\theta\) of real-world videos at each step of flow matching to be closer to the oracle velocity field \(x_1 - x_0\) than that of generated videos. The core mechanism shifts from "which generated video is less bad" to "whose predicted velocity field is closer to physical ground truth."

3. PGR (Physics-Guided Rewarding): Adaptive Hard-Sample Reweighting Driven by VLM Scores

Not all physical errors should be treated equally: a slightly tilted chair vs. a person jumping anti-gravitationally—the latter contains physical errors that must be heavily penalized. PGR utilizes a physics-aware VLM (VideoCon-Physics) to output two scores for each losing candidate: semantic alignment \(s^{sa}_j \in [0,1]\) and physical common sense \(s^{pc}_j \in [0,1]\)—and constructs a physical difficulty index \(v_j = 1 - (s^{sa}_j + s^{pc}_j)/2\). Under a two-layer sigmoid mapping, \(v_j\) determines the contrastive sharpness \(\alpha_j\) and loss weight \(\gamma_j\): physically implausible samples receive larger \(\gamma_j\) (scaled-up loss contribution) and smaller \(\alpha_j\) (smoother contrast). Here, \(\alpha_j\) controls the "strictness" of the preference comparison, while \(\gamma_j\) controls sample importance, governed by hyperparameters like \(\lambda\), \(\kappa_{\alpha,\gamma}\), \(b_{\alpha,\gamma}\), and \(\alpha_{\text{min}}\). This adaptive allocation of optimization budget is data-driven, yet a key constraint remains: VLM scores only affect sample weights and contrastive sharpness without modifying the core target of "real-world videos are superior to generated videos." Ground-truth physics is strictly guaranteed by real videos.

4. LoRA-SR (LoRA-Switch Reference): Lightweight Reference Model with Frozen Backbone and LoRA Switching

Standard DPO requires maintaining both the active model \(\theta\) (training) and the reference model \(\psi\) (frozen full copy). For Wan2.1-14B, the reference model alone demands ~50GB of GPU VRAM. LoRA-SR resolves this by freezing the backbone as reference model \(\psi\), appending LoRA modules to the Q/K/V/O linear projections of self-attention layers, and controlling LoRA activation through an environment manager: $$ \mathbf{Y} = \mathbf{X}(\mathbf{W} + \mathbf{1}_{action} \cdot \frac{\alpha}{r} \mathbf{BA})^\top $$ When \(\mathbf{1}_{action}=0\), the output is equivalent to the reference model; when \(\mathbf{1}_{action}=1\), the output acts as the active model (backbone + LoRA). Consequently, DPO only stores one additional set of LoRA weights (84MB) on top of the backbone, avoiding full duplication. Experiments demonstrate that LoRA-SR reduces GPU VRAM from 48.7GB to 25.3GB (44% reduction), compresses storage from 5.3GB to 84MB (over 60x reduction), and because low-rank updates restrict parameter drift, prevent active models from deviating too rapidly—improving Hard Action scores (0.0444) by 14% over the full-duplication version (0.0389), indicating that training stability aids hard-sample learning.

Loss & Training

The final training objective forces the flow matching denoising MSE (the \(\ell_2\) distance between predicted velocity field \(v_\theta\) and oracle velocity field \(x_1 - x_0\)) of real-world videos to be strictly less than that of generated candidate videos. Each candidate's weight and contrastive sharpness are dynamically adjusted via PGR based on VLM scores. Built upon the Wan2.1-14B base model, training is conducted on 8 H100 GPUs with a batch size of 8 for 10K steps (approx. 6 days) using BF16 mixed-precision and sublinear memory techniques. The optimizer is AdamW (\(\beta_1=0.9\), \(\beta_2=0.999\), weight decay 0.01), with the learning rate cosine-annealed from 1e-5 to 1e-6. Video resolution is set to 480×832. Key hyperparameters: temperature \(\tau=3\) (difficulty index sampling), \(\alpha_{\text{min}}=0.5\), \(\lambda=0.6\), and LoRA rank \(=48\).

Key Experimental Results

Main Results

Dataset	Dimension	Ours	Prev. SOTA	Gain
VideoPhy2	Hard	0.0500	Veo3.1 0.0444	+13%
VideoPhy2	Activity	0.1571	Veo3.1 0.1405	+12%
VideoPhy2	Interaction	0.1761	Veo3.1 0.1887	-7%
VideoPhy2	Overall	0.1627	Veo3.1 0.1525	+7%
PhyGenBench	Mechanics	0.55	VideoDPO 0.48	+15%
PhyGenBench	Thermal	0.58	VideoDPO 0.47	+23%
PhyGenBench	Avg	0.55	VideoDPO 0.54	+2%

User study (104 participants, 48 rounds/person): PhyGDPO outperforms all baselines (including Sora2 and Veo3.1) with a preference rate of 64.4%–94.2%, with the 95% confidence interval for each result not exceeding ±2.4%.

Ablation Study

Configuration	VideoPhy2 Overall	Description
Baseline (Wan2.1-14B)	0.1288	Original base model
+ CoT Data Filtering	0.1525	Physics richness filtering, +18%
+ Action Clustering & Balanced Sampling	0.1575	More balanced action distribution
+ Physics Reward Reweighting (Full PhyAugPipe)	0.1627	Complete data pipeline
+ LoRA-SR	0.1458	GPU VRAM reduced by 44%, Hard +150%
+ Groupwise PL Model	0.1559	Significant improvement over pairwise baseline
+ PGR (Full PhyGDPO)	0.1627	Complete method

Key Findings

• Hard Action is the biggest beneficiary: The complete method achieves a 4.5x improvement on Hard cases (0.0111 → 0.0500), showing that group-wise contrast and hard-case reweighting effectively target the "difficult actions" bottleneck. Improvements on Activity and Interaction are comparatively moderate.
• LoRA-SR reduces VRAM while boosting Hard Score: Compared to the full duplication version (Hard 0.0389), the LoRA-SR version achieves a 14% higher Hard score (0.0444), suggesting that training stability brought by low-rank parameter updates has a positive effect on hard-sample learning—the active model does not "drift" too far during optimization.
• Cross-model and cross-evaluation validation confirms generalization: When replacing the base model from Wan2.1-14B to Vcrafter2, or replacing the evaluator from VideoPhy2-AutoRater to Gemini-2.5-pro, PhyGDPO consistently maintains its lead, demonstrating resistance to specific T2V architectures or VLM scorers.

Highlights & Insights

• "Real-world video as winning sample" is a game-changing decision: Shifting DPO from "comparing two imperfect generated videos" to "perfect real video vs. imperfect generated video" qualitatively improves the learning signal. This is the most insightful design, likely to become a standard practice in future DPO-based video generation research.
• Mathematical inequality construction makes groupwise training practical: Starting from the PL model and decoupling \(\log(1 + \sum e^{x_j})\) into step-wise/pairwise forms using Jensen's and constructed inequalities not only reduces training inference costs per group from \(2m+2\) to \(1\), but also naturally introduces sample reweighting through \((\alpha_j, \gamma_j)\)—unifying mathematical formulation and physical motivation.
• The "VLM score only changes weight, not target" constraint in PGR is elegant: VLM scores only control which samples learn more or adapt more aggressively, while the "what is correct" signal is consistently supplied by real videos. This prevents VLM bias from polluting the core optimization of DPO while retaining the benefits of adaptive reweighting.
• LoRA-SR is a highly transferable engineering contribution: Not limited to video generation, any scenario requiring reference model loading in DPO can reuse this design—saving 23GB VRAM on a 14B model represents a qualitative leap in efficiency.

Limitations & Future Work

• Dependence on upper limits of VLM evaluation quality: PGR's effectiveness is constrained by VideoCon-Physics' evaluation. VLMs may struggle to accurately evaluate extremely rare physical phenomena. Though cross-evaluator validation using Gemini-2.5-pro alleviates this, the risk is not entirely eliminated.
• High computational cost of CoT filtering: Processing million-scale data using Qwen2.5-72B is a costly, one-time offline step, which could pose a barrier for teams trying to reproduce or transfer this to other domains (e.g., medical physical simulation).
• Lagging behind Veo3.1 on the Interaction dimension: PhyGDPO (0.1761) performs worse than Veo3.1 (0.1887) in VideoPhy2 Interaction. The paper does not analyze this extensively—group-wise DPO may still fall short in object-interaction tasks requiring precise temporal alignment.
• Capacity limits of LoRA rank: Whether LoRA rank=48 has sufficient capacity to express physical parameter variations under longer training or larger datasets, and the performance of LoRA-SR compared to full fine-tuning (non-LoRA SFT), warrant further exploration.

Related Work & Insights

• vs VideoDPO: VideoDPO is the first work to apply DPO to T2V, but relies on pairwise comparisons and generated videos as winning samples. PhyGDPO uses real-world videos as winning samples with the PL group-wise model, achieving an 18.5% improvement in VideoPhy2 Overall (0.1373 → 0.1627) and an 80% improvement in Hard Actions.
• vs PhyT2V: PhyT2V adopts an LLM prompt expansion route, rewriting prompts at inference to include physical descriptions. PhyGDPO conducts post-training and does not rely on external modules at inference. PhyGDPO leads by 9% on VideoPhy2 Overall (0.1627 vs. 0.1492) with lower inference latency/cost.
• vs Flow-DPO: Flow-DPO integrates DPO into flow matching models but relies on pairwise comparison. Under the same settings (Wan2.1-1.3B, same data), PhyGDPO comprehensively outperforms Flow-DPO, leading by 50% in Hard Actions.

Rating

• Novelty: ⭐⭐⭐⭐☆ Introducing group-wise PL models to video physical preference optimization, utilizing real-world videos as winning samples, dual-weighted PGR, and LoRA-SR are individually innovative. Together, they constitute a systematic adaptation rather than a paradigm shift.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ The ablation studies are comprehensive and progressive (covering data components, algorithmic components, DPO variants, cross-VLM evaluation, cross-backbone, parameter sensitivity), paired with a solid 104-participant, 48-round user study.
• Writing Quality: ⭐⭐⭐⭐☆ The mathematical derivations are clear, notations are consistent, and ablation details align with the tables. The introduction is slightly long, and some equation derivations could reside in the appendix rather than the main text.
• Value: ⭐⭐⭐⭐⭐ Physical consistency is a primary bottleneck in T2V generation. LoRA-SR can be directly applied to other DPO scenarios, and the experimental design along with the user study is highly convincing.

Related Papers

• [CVPR 2026] LocalDPO: Direct Localized Detail Preference Optimization for Video Diffusion Models
• [ECCV 2026] Physics Question Scene Graph: Fine-grained Evaluation of Physical Plausibility in Text-to-Video Generation
• [ICLR 2026] NewtonGen: Physics-consistent and Controllable Text-to-Video Generation via Neural Newtonian Dynamics
• [ICLR 2026] Dual-IPO: Dual-Iterative Preference Optimization for Text-to-Video Generation
• [ICLR 2026] Consistent Noisy Latent Rewards for Trajectory Preference Optimization in Diffusion Models