PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models

Conference: ECCV2026
Authors: Zeqing Wang; Keze Wang; Lei Zhang
Affiliations: Sun Yat-sen University; Hong Kong Polytechnic University; OPPO Research Institute
Paper: ECCV paper page
PDF: Conference full text
Code: https://github.com/Zeqing-Wang/PhyDetEx
Area: Video Generation
Keywords: physical plausibility, anomaly detection, vision-language models, paired supervision, interpretability

Identity check: the conference record and cached paper have the same title. Zeqing Wang is affiliated with Sun Yat-sen University and OPPO, Keze Wang with Sun Yat-sen University, and Lei Zhang with Hong Kong Polytechnic University and OPPO; the latter two are corresponding authors.

TL;DR

PhyDetEx fine-tunes Qwen2.5VL 7B with LoRA on 2,588 plausible/implausible video pairs, combining physical judgments and explanations in a language-modeling task and achieving 75.6% implausible-class accuracy, a reported F1 of 83.5, and a Reasoning Score of 4.24 on PID Test.

Background & Motivation

Text-to-video (T2V) models can produce clear, smooth videos that follow instructions while still making objects penetrate one another or leaving unsupported people suspended in midair. Visual quality and physical plausibility are different objectives: convincing cups and water do not imply that the water level changes correctly during pouring. Assessing progress toward world models therefore requires checking dynamic constraints rather than relying only on text-video similarity.

General-purpose vision-language models (VLMs) also miss these errors. The preliminary experiments show that telling a model that a video may be AI-generated makes it more likely to detect violations, but also more likely to reject plausible videos. Failure may therefore reflect not only missing physical knowledge but also assumptions about provenance: real videos are presumed plausible, whereas generated videos invite suspicion. Prompt changes alone cannot ensure that judgments follow visual evidence.

The authors address both training data and evaluation distributions: preserve scene semantics while changing physical events during training, and include plausible generated videos in the plausible test class. Core Idea: use semantically similar video pairs with opposite physical plausibility to supervise both judgments and physical explanations, shifting evaluation from source recognition toward event inspection.

Method

Overall Architecture

PhyDetEx is an evaluator for video generation, not a new generator or a detector of whether a video was made by AI. It takes a video and an evaluation instruction, then produces a textual sequence beginning with [YES] or [NO] and explaining which motions conform to or violate real-world physics. Training captions help construct supervision; deployment should not depend on a human revealing the violation in advance.

The workflow has three connected designs: construct training examples through Physical-Event Pairing, obtain a detector through Joint Judgment-and-Explanation Tuning, and evaluate unintended failures under normal prompts through Provenance-Shortcut-Controlled Evaluation. PID Test is constructed independently and is not used for LoRA training.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Real videos and captions"] --> B["Physical-Event Pairing"]
    B --> C["Joint Judgment-and-<br/>Explanation Tuning"]
    C --> D["PhyDetEx detector"]
    E["Independent PID Test<br/>and generated videos"] --> F["Provenance-Shortcut-<br/>Controlled Evaluation"]
    D --> F
    F --> G["Judgments, explanations,<br/>and generator evaluation"]

Key Designs

1. Physical-Event Pairing: similar contexts with opposite physical events

The authors select real videos and captions with clear physical interactions from VIDGEN-1M, then ask a large language model to rewrite the physical event. The rewrite should retain the environment, background objects, and scene layout while introducing an impossible motion or interaction. A T2V model generates the counterexample. Each training pair thus contains a plausible real event and a contextually similar implausible event, rather than two unrelated video topics.

Generators may fail to realize an implausible caption, so candidates require filtering. A VLM is explicitly told that a candidate is generated, exploiting the more sensitive behavior observed in the preliminary study to retain pairs that exhibit the intended violation. The resulting training set contains 2,588 video pairs, not 2,588 individual videos. The cached main paper does not specify the full filtering VLM and generator configurations or filtering thresholds, so these cannot be supplied as verified reproduction settings.

Pairing reduces irrelevant semantic differences. The presence of water, people, or a particular background should not distinguish the positive and negative members of the same pair; motion becomes more informative. This is a contrastive structure in the data, not an additional InfoNCE loss, and it does not mean inference requires feeding both plausible and implausible videos together.

The design intent should be distinguished from guaranteed properties. Preserving background descriptions does not make a regenerated video pixel-identical to a real one. Plausible training examples are still real and implausible examples still generated. Pairing can control semantic confounds without proving that every provenance shortcut has disappeared.

2. Joint Judgment-and-Explanation Tuning: put the decision and its evidence in one sequence

With only binary labels, a model can guess correctly from generation artifacts without revealing what it observed. PhyDetEx therefore first emits a judgment and then describes the physical content supporting it: the first token represents [YES] or [NO], and subsequent tokens form an explanation. Rewritten captions provide known violation semantics from which labels and explanatory supervision can be constructed.

The backbone is Qwen2.5VL 7B, adapted with LoRA rather than training a video encoder from scratch or repeating large-scale pretraining. Lightweight here means parameter-efficient adaptation and a comparatively small dataset. The main paper does not report GPU models, training time, or memory consumption, so it does not support a concrete cost claim.

Unified language modeling lets the judgment and explanation share visual conditioning and autoregressive context. The intended behavior is not merely to flag an error but to identify the object, motion, and physical expectation that conflict. Nevertheless, an explanation remains generated language, not a physics simulation or an automatically verifiable causal proof.

3. Provenance-Shortcut-Controlled Evaluation: detect accidental errors under normal requests

PID Test starts from real user prompts in VidProM. The pipeline excludes descriptions involving cartoons, magic, or science fiction that already permit unrealistic content, and removes static scenes without physical interactions. Multiple T2V models generate videos from the remaining normal prompts; human annotators identify violations and provide explanations. The failures arise unintentionally rather than because users explicitly requested impossible events.

The plausible class combines physically plausible generated videos with real VIDGEN-1M videos. The real clips are filtered to last under 10 seconds and contain clear physical interactions. The final set has 500 videos, with 250 in each class. Labeling every generated video implausible now causes errors on plausible generations, making this a stronger test of provenance bias than an entirely real plausible class.

The detector is also used to benchmark T2V models on 50 prompts with clear physical interactions from VIDGEN-1M. Plausibility Rate is the fraction of videos classified as plausible. Plausibility Score adds the corresponding prediction logit for plausible judgments and subtracts it for implausible judgments. This is a cumulative internal confidence signal, not a physical unit, probability, or percentage; the main text does not fully specify logit extraction and calibration.

A Worked Example

In the paper's Figure 5, a dolphin above the sea resembles a normal leap but remains at approximately the same height over time. General VLMs describe familiar dolphin-jumping behavior instead of checking the expected subsequent descent.

PhyDetEx instead returns an implausible judgment and identifies persistent suspension without supporting or propulsive evidence and without the expected effect of gravity. The example shows why an explanation should identify temporal evidence rather than merely mention gravity, and why a plausible single-frame appearance can coexist with implausible video dynamics.

Loss & Training

Section 4.2 defines supervision as standard negative log-likelihood over the output sequence. Equation (7) is corrupted in the text extraction; the following standard form is reconstructed from its surrounding description, rather than interpreted as a novel loss:

\[ \mathcal{L}_{\mathrm{LM}}(\mathbf{s}\mid v,t) =-\sum_{u=1}^{T}\log p_{\theta}(s_u\mid s_{<u},v,t). \]

Here \(v,t\) denote the video and text inputs in that section's notation, \(\mathbf{s}\) includes both judgment and explanation, and \(\theta\) denotes trainable parameters. The main paper does not introduce a separate pairwise distance term, contrastive temperature, or ranking loss. Its phrase contrastive fine-tuning should therefore not be interpreted as adding such an objective.

Training lasts 3 epochs with a learning rate of \(10^{-4}\) and a cosine schedule. LoRA rank is 8 and adaptation is applied to all target modules. Their exact names, batch size, and video frame-sampling policy are not detailed in the cached main text.

Key Experimental Results

Main Results

Table 1 evaluates two datasets. Impossible Videos uses 535 implausible videos from the Physical Laws category and all 650 plausible real videos. PID Test contains 250 implausible and 250 plausible videos. These are different distributions, and the former is not the complete collection of implausibility categories in Impossible Videos.

Both class accuracies below are percentages. F1 retains the paper's percentage-scale reported values, while Reasoning Score retains the raw scores. The main paper describes the latter as explanation correctness and depth, but places its scale, judging protocol, and complete F1 protocol in the unavailable Supplement D. It is therefore not labeled a five-point score or macro-F1 here.

Dataset	Model	Implausible accuracy (%)	Plausible accuracy (%)	F1 (reported)	Reasoning Score
Impossible Videos	Qwen2.5VL 7B	64.9	69.8	70.3	1.73
Impossible Videos	Qwen2.5VL 72B	51.2	88.5	77.4	3.22
Impossible Videos	PhyDetEx 7B	86.4	87.1	87.8	4.20
PID Test	Qwen2.5VL 7B	57.2	86.4	75.4	1.66
PID Test	Qwen2.5VL 72B	54.0	91.2	76.9	3.13
PID Test	PhyDetEx 7B	75.6	89.2	83.5	4.24

Relative to the Qwen2.5VL 7B backbone in Table 1, PID implausible accuracy increases by 18.4 percentage points, reported F1 by 8.1 points, and Reasoning Score by 2.58. Plausible accuracy also improves from 86.4% to 89.2%, so these gains are not simply obtained by pushing every judgment toward implausibility.

Ablation Study

The main paper does not provide numerical component-removal ablations; the training analyses are assigned to the unavailable Supplement B. The following prompt-condition analysis from Table 2 tests the motivation and is not presented as an ablation of LoRA rank, pairing, or explanation supervision.

InternVL2.5 26B / Impossible Videos	Provenance prompt	Implausible accuracy (%)	Plausible accuracy (%)	Reasoning Score
C1	No provenance information	65.23	78.00	2.90
C2	Video may be AI-generated	81.87	60.15	3.62
C3	Video is generated by a T2V model	90.47	51.54	3.76

From C1 to C3, implausible accuracy rises by 25.24 percentage points while plausible accuracy falls by 26.46 percentage points. This supports a provenance-induced shift in judgment, not a claim that prompting alone comprehensively solves detection.

Some Qwen baseline values differ between Tables 1 and 2. For example, Qwen2.5VL 7B has 64.9% implausible accuracy in Table 1 but 53.08% under C1 in Table 2. Values are retained separately without assuming identical settings or combining them into a single improvement trajectory.

Key Findings

• Task adaptation lets a smaller model outperform a larger backbone: PID reported F1 is 83.5 for PhyDetEx 7B and 76.9 for Qwen2.5VL 72B. This does not establish that scale is irrelevant to every physical-reasoning task.
• In Table 3's 50-prompt generator benchmark, Sora2.0 has Rate/Score of 86.0%/32.62 and Veo3.1 has 72.0%/12.98. These are detector-based evaluations, not human-certified physical correctness rates.
• Table 3 reports 66.0%/16.54 for WanX2.1 1.3B and 66.0%/12.96 for WanX2.2 14B. The authors suggest simpler motions may produce fewer violations. Section 5.5 lists the former as 5B in one place; this note follows the 1.3B designation in Table 3 and the result discussion.

Highlights & Insights

• Data controls are more concrete than asking a model to inspect physics carefully. Pairing focuses supervision on motion and interaction differences, but semantic-level controls still require independent tests against generator artifacts.
• Plausible generated videos are central to testing provenance shortcuts. The same principle can test whether other generated-content evaluators merely recognize a particular generation style.
• Explanations make wrong judgments inspectable by humans. Their value is diagnostic access, not an automatic guarantee of reliability when the prose is fluent.

Limitations & Future Work

• The authors acknowledge hallucinations on non-visual or complex physical events and suggest audio and additional object information. If the source of a force is invisible, video-only judgments may lack sufficient evidence.
• Training provenance remains correlated with the label, so pairing cannot logically eliminate every real/generated shortcut. A direct follow-up would include plausible generated training examples and evaluate unseen generators.
• The cache contains the main paper and references but not Supplements A-H. Explanation-scoring details, numerical ablations, frame sampling, and the implementation and results of physics-aware DPO cannot be fully verified from this material.
• Table 1's F1 definition requires Supplement D, and baseline values and model specifications differ across parts of the paper. Comparisons should preserve table numbers, datasets, and prompt conditions rather than silently harmonizing them.
• The 50-prompt benchmark is limited, and simpler motions can increase plausibility rates. Future evaluation should jointly control motion complexity, instruction following, and visual quality to avoid rewarding reduced motion as stronger world modeling.

Related Work & Insights

• vs Impossible Videos: it establishes an initial benchmark for impossible-event detection, but its violations largely follow implausible prompts and its plausible videos are real. PID tests accidental failures under normal prompts and includes plausible generations.
• vs VideoPhy / VideoPhy-2: these evaluate physical commonsense in video generation. PhyDetEx emphasizes readable VLM explanations beyond coarse scores; stronger cross-distribution generalization still requires matched comparisons.
• vs Morpheus: Morpheus uses real physical experiments and conservation-law metrics. PhyDetEx supports language-based diagnosis in open scenes but does not provide equivalent quantitative physical measurements.
• Implications for generator training: the paper mentions constructing preference examples for physics-aware DPO. Because Supplement H is unavailable in the cache, this is treated as an application direction without unverified improvement claims.

Rating

• Novelty: 4/5. The contribution combines task-specific data, provenance-bias analysis, and interpretable evaluation; the training objective itself is standard.
• Experimental Thoroughness: 3/5. Two detection datasets and generator benchmarking are complementary, but essential protocols and ablations reside in unavailable supplements.
• Writing Quality: 3/5. Motivation and data construction are clear, while some cross-table values and model specifications need clarification.
• Value: 4/5. Useful for screening and analyzing physical errors in generated videos, but not as a physics judge that needs no review.