Rapidata/world-models-physics-generation-human-feedback-benchmark

World Models: Physics Generation Benchmark v2

Size: 63,245 rows · 758,940 individual human votes in total (verified against the parquet files on 2026-11-03)

Built by Rapidata, collected with the Rapidata Python SDK.

To evaluate your own models and create a leaderboard, check out our MRI.

The usefulness of world and video models in robotics, autonomous driving and games depends heavily on how well they model physics, which has made physical accuracy a growing focus. Physics-IQ (Motamed et al., 2026) and its refined version Physics-IQ Verified (Rädsch et al., 2026) measure this automatically: a model continues a real recording of a physical experiment, and its video is compared with what actually happened using hand-crafted metrics on motion and pixels. We build on this work independently: on the same 198 real recordings across 66 physical scenarios, we rank models by human judgement instead of automated metrics. We evaluate 26 video and world models, which are each given an input frame taken from a real recording and a prompt describing the scene, without revealing the expected outcome.

The Blind mode presents two generated videos and asks human raters which of them looks more realistic.

The Referenced mode additionally shows the original recording and asks which video reproduces it more faithfully.

Using 758,940 human judgements and 63,245 pairwise comparisons, we compared 26 models. In our overall leaderboard, which weights both modes equally, FLUX 3 Max (post-trained by fal on Black Forest Labs' FLUX 3) emerged as the winner, followed by Gemini Omni 1.1 Flash and Dreamina Seedance 2.5.

This is the data behind World Models: Physics Generation Benchmark v2 on benchmark.ai.

• Live leaderboard: app.rapidata.ai
• Benchmark page and methodology: benchmark.ai/world-model-physics-v2

At a glance

Head-to-head comparisons (rows)	63,245
Human votes	758,940
Models compared	26
Recordings	198 (66 scenarios × 3 camera angles)
Votes per matchup	6 per question

Motivation

Physics-IQ (Motamed et al., 2026) evaluates physical understanding by comparing a model's continuation of a physical event with the real recording. Neither it nor Physics-IQ Verified collected human judgements. This benchmark adds a human ranking of 26 models on the same recordings.

Methodology

Physics-IQ introduced 66 scenarios, each filmed for 8 seconds from three angles (left, center and right) with two takes, giving a total of 396 recordings. The scenarios span solid mechanics, fluid dynamics, optics, thermodynamics and magnetism (Table 1). In Physics-IQ, the first take (from the 3 angles) was used to prompt a model for a prediction.

Following the Physics-IQ protocol, we generated predictions for take 1 from different angles, using the same switch frames. A switch frame is the last frame before the physical event starts, and the model continues the video from it. Every model received only the switch frame as visual input, which is the image-to-video protocol that Physics-IQ Verified evaluates. We did not use the 3 seconds of video conditioning that the original paper gave to video-to-video models.

The models were further supplied with a prompt describing the scenes. We used the improved best practice prompts from the verified paper as built by its default (base) templater, for every model. Besides the base template, the Physics-IQ Verified code includes templates for Sora 2 and P-Video, neither of which we benchmarked.

In Physics-IQ, the second take is used to estimate the physical variance between two real executions. In the automated metric, a prediction gets full marks on each of the four measures if it is at least as close to the first take as the second take is.

For the blind human metric, the reference does not matter. For the referenced mode, we also ran the question with the second take as the reference. The ranking was almost identical to the one with the first take (Spearman 0.995 across the 26 models), so showing the second take to raters as well would add little.

Table 1. Composition of the 198 prompts.

Composition	Prompts
Solid mechanics	114
Fluid dynamics	45
Optics	24
Thermodynamics	9
Magnetism	6
Per perspective, left / center / right	66 / 66 / 66

Video generation

The 21 video models were all generated through fal. The 5 world models were run from their open weights on Hugging Face: AlayaWorld (AlayaLab/AlayaWorld, 960×544, 24 fps), Cosmos 2.5 (nvidia/Cosmos-Predict2.5-2B, 1280×704, 16 fps), Cosmos 3 Nano (nvidia/Cosmos3-Nano, 832×480, 24 fps), Dream-X-World (GD-ML/DreamX-World-5B, 1280×704, 16 fps) and SANA-WM (Efficient-Large-Model/SANA-WM_bidirectional, 720p, 16 fps, static camera). Every model received the recording's switch frame as input image together with the scene prompt, and made one generation per scenario / angle combination, without selecting among several. Prompt rewriting and audio were set to off, and the seed was fixed per recording as 1000 × id + 1 wherever the model accepts one. The two Cosmos models used NVIDIA's default negative prompt and kept its Cosmos Guardrail safety filter on. Seedance 2.0 and 2.5 start most clips from the input frame zoomed in by about 6%, which cannot be turned off. Every clip was cut to exactly five seconds at the model's native frame rate, video only, so annotators judge exactly what the Physics-IQ protocol describes.

Human annotation protocol

Every match pairs the continuations of two models for the same recording, so camera perspectives and scenes are never mixed. Each pair is judged under both questions.

Leaderboard	Question shown to annotators	Real recording shown?	Prompt shown?
blind	Which video is more realistic?	No	No
referenced	Which one looks more like first video?	Yes: a looping clip next to the two candidates, playing at the same time	No

Blind measures perceived physical plausibility: does the continuation pass as real footage? Referenced shows the real recording (as released in Physics-IQ Verified, where regions with recording artefacts are held still) and measures faithfulness to what actually happened, the property the automated Physics-IQ score targets. In both questions annotators never see the scene description the models were given. Every match collects five or six responses (six for the 26 models on the board at launch) and is decided by majority vote. A tie counts as half a win for each model.

Generation failures leave a few gaps. Safety filters refused eleven Veo generations (Veo 3.1 has 192 of 198 clips, Veo 3.1 Fast 193), and three world models cover a subset (Cosmos 2.5 177, Cosmos 3 Nano 186, SANA-WM 197). Missing clips simply do not enter those models' matches.

Ranking

Match outcomes are fitted with a Bradley–Terry model. Scores are reported as 1000 + 400 ln π, so the average model scores 1000, with 95% confidence intervals (see Methodology). The overall ranking weights Blind and Referenced equally. Rank among the 26 models in brackets; the scores refresh on every sync.

Model	Served via	Overall	Blind	Referenced
FLUX 3 Max	fal	1363 (#1)	1318 (#2)	1413 (#1)
Gemini Omni 1.1 Flash	fal	1272 (#2)	1352 (#1)	1199 (#6)
Dreamina Seedance 2.5	fal	1266 (#3)	1225 (#6)	1312 (#2)
MiniMax H3	fal	1246 (#4)	1262 (#4)	1233 (#5)
Dreamina Seedance 2.0	fal	1239 (#5)	1227 (#5)	1254 (#4)
FLUX 3 Video	fal	1231 (#6)	1179 (#9)	1288 (#3)
Gemini Omni Flash	fal	1187 (#7)	1275 (#3)	1105 (#9)
Grok Imagine Video 1.5	fal	1160 (#8)	1217 (#7)	1107 (#8)
HappyHorse 1.0	fal	1133 (#9)	1084 (#11)	1185 (#7)
Wan 3.0	fal	1125 (#10)	1209 (#8)	1045 (#13)
Kling 3.0 Pro	fal	1055 (#11)	1079 (#12)	1032 (#14)
Kling 2.6 Pro	fal	1049 (#12)	1006 (#16)	1094 (#10)
Grok Imagine Video	fal	1047 (#13)	1126 (#10)	971 (#17)
Wan 2.7	fal	1029 (#14)	968 (#18)	1092 (#11)
Wan 2.6	fal	1023 (#15)	984 (#17)	1064 (#12)
Veo 3.1 Fast	fal	1009 (#16)	1027 (#14)	992 (#16)
Vidu Q3 Pro	fal	992 (#17)	1028 (#13)	958 (#18)
Veo 3.1	fal	985 (#18)	1020 (#15)	951 (#19)
Wan 2.5 Preview	fal	963 (#19)	925 (#19)	1000 (#15)
Pixverse V5.6	fal	915 (#20)	907 (#20)	923 (#20)
Cosmos 3 Nano	world model	799 (#21)	793 (#21)	805 (#21)
Cosmos 2.5	world model	706 (#22)	706 (#22)	706 (#23)
Seedance V1.5 Pro	fal	642 (#23)	705 (#23)	574 (#24)
AlayaWorld	world model	612 (#24)	446 (#25)	752 (#22)
Dream-X-World	world model	496 (#25)	560 (#24)	423 (#26)
SANA-WM	world model	455 (#26)	376 (#26)	522 (#25)

Dataset structure

One row per head-to-head comparison of two samples generated from the same prompt.

Column	Type	Meaning
prompt	string	The prompt the two compared samples were generated from.
video1_preview	string	Low-resolution GIF of video1, so the dataset viewer can preview it; set on the first 100 rows only.
video2_preview	string	Low-resolution GIF of video2, so the dataset viewer can preview it; set on the first 100 rows only.
video1	string	Compared video 1: public URL of the video.
video2	string	Compared video 2: public URL of the video.
model1	string	Name of the model that produced video1.
model2	string	Name of the model that produced video2.
weighted_results_video1_blind	float32	Sum of the userScore weights of the votes for video1 on the blind leaderboard.
weighted_results_video2_blind	float32	Sum of the userScore weights of the votes for video2 on the blind leaderboard.
detailed_results_blind	string	Every vote on the blind leaderboard as a JSON array; votedFor is A for side 1 and B for side 2.
weighted_results_video1_referenced	float32	Sum of the userScore weights of the votes for video1 on the referenced leaderboard.
weighted_results_video2_referenced	float32	Sum of the userScore weights of the votes for video2 on the referenced leaderboard.
detailed_results_referenced	string	Every vote on the referenced leaderboard as a JSON array; votedFor is A for side 1 and B for side 2.

The weighted_results_* pair is a userScore-weighted vote sum per side, not a probability: the two values of a row do not sum to 1. Divide by their sum for a normalised win share.

The rows do not carry the recording identifier or camera angle. The three angles of a scenario share one prompt text, so prompt takes 65 distinct values; matchups still never mix angles.

Usage

import json
from datasets import load_dataset

ds = load_dataset("Rapidata/world-models-physics-generation-human-feedback-benchmark", split="train")

row = ds[0]
votes = json.loads(row["detailed_results_referenced"])
video1_votes = sum(v["votedFor"] == "A" for v in votes)
w1, w2 = row["weighted_results_video1_referenced"], row["weighted_results_video2_referenced"]
print(row["model1"], "vs", row["model2"], f"{video1_votes}/{len(votes)} votes for video1, weighted share {w1 / (w1 + w2):.2f}")

Data and licence

Released under CC BY 4.0. The recordings, switch frames and scene descriptions come from Physics-IQ (© DeepMind) as refined in Physics-IQ Verified, both CC BY 4.0 (Physics-IQ code: Apache 2.0); cite them as below when you use this data. The videos were generated by the listed models; their providers' terms may also apply to the clips.

Citation

@misc{rapidata2026worldmodelphysicsv2,
  title        = {World Model Physics Benchmark v2: Video and World Models Judged by Humans},
  author       = {{Rapidata}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Rapidata/world-models-physics-generation-human-feedback-benchmark}}
}

@inproceedings{motamed2026physics,
  title     = {Do generative video models understand physical principles?},
  author    = {Saman Motamed and Laura Culp and Kevin Swersky and Priyank Jaini and Robert Geirhos},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages     = {948--958},
  year      = {2026}
}

@article{radsch2026verified,
  author  = {Rädsch, Tim and Asano, Yuki M. and Kuehne, Hilde and Bauer, Stefan and Jaini, Priyank and Geirhos, Robert and Lüth, Carsten T.},
  title   = {Physics-IQ Verified},
  journal = {arXiv preprint arXiv:2606.18943},
  year    = {2026}
}

About Rapidata

Rapidata's technology makes collecting human feedback at scale faster and more accessible than ever before. Visit rapidata.ai to learn more.