Rapidata/image-to-video-physics-human-feedback-benchmark

Rapidata Physics Benchmark

Live leaderboard: benchmark.ai/world-model-physics

Size: 22,625 rows · 298,700 individual human votes in total (verified against the parquet files on 2026-11-03)

Built by Rapidata.

Do video and world models understand physics? We gave 25 video- and world models the same real-world starting frame and scene description from Physics-IQ and asked them to predict what happens next. ~283,000 human votes, collected with the Rapidata Python SDK, decided which continuation is more realistic — with the real recording competing as a hidden 26th participant.

Each row is a head-to-head matchup between two participants on the same scenario, scored on two questions (blind and with-reference, explained below), with every individual vote preserved. No automated metrics, only humans.

If you get value from this dataset and would like to see more in the future, please consider liking it ❤️

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

🏆 Live Leaderboard

Explore the full interactive leaderboard — switch between the two leaderboards, filter by model and scenario, and watch individual head-to-head matchups on Rapidata.

👆 Click to open and interact with it on rapidata.ai

At a glance

Head-to-head comparisons (rows)	135,765 — every pair of participants on every scenario (325 pairs × 65 scenarios)
Human votes	798,810 (blind 141,351 · with-reference 141,347)
Participants	31 — 25 generative models + the real recording (Ground Truth)
Scenarios	65 Physics-IQ scenes (solid mechanics, fluid dynamics, optics, thermodynamics, magnetism)
Videos per model	260 — 4 independent generations per scenario, ~5 s each
Video pairs voted on	~52,000, each shown to ≥ 5 annotators (boosted pairs 10)
Annotators	217,606 sessions from 141 countries, every vote carries demographics

The two leaderboards

Both leaderboards ask the same question about the same pairs of videos. The only difference is what else the annotator gets to see:

Leaderboard	Question shown to annotators	Reference recording shown?	Measures
blind	"Which video is more realistic?"	No	perceived realism on its own — does the clip look like real footage?
with-reference	"Which video is more realistic?"	Yes — the real continuation plays alongside the two candidates	realism and physical correctness — did the model predict what actually happened?

What the models were given

Every scenario follows the Physics-IQ protocol. A real physical event was filmed, and one frame of that recording — the switch frame — was picked by hand. The switch frame shows the objects and the event as it is happening: the kettlebell already touches the cup, the ball has just left the pipe, the match is entering the water. That is enough context for a model to know what is about to happen and to set the motion in motion, but the outcome itself is not visible yet.

Each model received exactly two inputs per scenario:

Image input — the switch frame

Text input — the scene description

A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.

The description says what is in the scene and what is being done to it, never how it ends. From these two inputs every model generated the next ~5 seconds. The real recording of those 5 seconds is the Ground Truth participant on the leaderboards and the reference clip on the with-reference leaderboard.

with-reference

The with-reference score measures realism and physical correctness. Annotators see the real recording alongside the two candidates and are asked: "Which video is more realistic?". The winner has a green border; the vote count is the human vote on this exact pair.

A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass.

Reference

real recording, shown to annotators

Dreamina Seedance 2.0

9 of 10 votes · with-reference Elo: 1268 (#5)

Vidu Q3 Pro

1 of 10 votes · with-reference Elo: 987 (#18)

Displacement: the potato sinks and the liquid rises past the green tape, exactly as in the recording. In the losing clip the potato hangs in the liquid and the level barely moves.

A lit match is being lowered into a glass of water.

Reference

real recording, shown to annotators

Dreamina Seedance 2.5

10 of 10 votes · with-reference Elo: 1247 (#7)

Wan 3.0

0 of 10 votes · with-reference Elo: 1183 (#8)

Fire meets water. One model extinguishes the match; the other pulls it back out and keeps it burning.

A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again.

Reference

real recording, shown to annotators

Wan 2.7

10 of 10 votes · with-reference Elo: 1009 (#17)

Grok Imagine Video 1.5

0 of 10 votes · with-reference Elo: 1143 (#9)

The prompt says “magnet”, but a tennis ball is not magnetic. Grok lifts it anyway — and a lower-ranked model wins the pair 10 – 0.

blind

The blind score measures perceived realism on its own. Without a reference, annotators are asked: "Which video is more realistic?".

A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side.

Gemini Omni 1.1 Flash

10 of 10 votes · blind Elo: 1350 (#2)

Wan 2.5 Preview

0 of 10 votes · blind Elo: 923 (#22)

The cup either crumples, or the kettlebell comes to rest on top of it.

A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug.

Veo 3.1 Fast

10 of 10 votes · blind Elo: 1039 (#16)

Wan 2.6

0 of 10 votes · blind Elo: 983 (#20)

Ceramic on concrete: shards, or a mug that quietly vanishes.

When reality loses

The real recording is a participant like any other. Blind, it wins just 51 % of its votes against the five top models (61 % once the reference is shown) — and sometimes loses a pair outright. Against the five world models it still wins 77 % blind.

A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks.

MiniMax H3

8 of 10 votes · blind Elo: 1298 (#4)

Ground Truth

2 of 10 votes · blind Elo: 1407 (#1) · real recording

Blind, 8 of 10 annotators preferred MiniMax H3's clip over the actual footage. The real liquid being drawn up into the jar looked “wrong” to them.

A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball.

Ground Truth

10 of 10 votes · blind Elo: 1407 (#1) · real recording

Dream-X-World

0 of 10 votes · blind Elo: 580 (#27)

A world model against reality. The ball morphs instead of dropping — 10 – 0 for the recording.

Rankings

Elo is the Bradley–Terry score fitted on all pairwise votes, ± is the 95 % confidence interval. The overall ranking aggregates both leaderboards with equal weight.

Overall ranking

#	Model	Lab	Elo	±
1	Ground Truth	real recording	1736 (#1)	26
2	Gemini Omni 1.1 Flash	Google	1297 (#3)	27
3	MiniMax H3	MiniMax	1292 (#4)	27
4	Gemini Omni Flash	Google	1217 (#8)	27
5	Dreamina Seedance 2.0	ByteDance	1279 (#5)	26
6	FLUX 3 Video	Black Forest Labs	1255 (#7)	26
7	Dreamina Seedance 2.5	ByteDance	1259 (#6)	26
8	Wan 3.0	Alibaba Cloud	1091 (#12)	26
9	Grok Imagine Video 1.5	xAI	1132 (#9)	26
10	Grok Imagine Video	xAI	1081 (#13)	26
11	Kling 3.0 Pro	Kuaishou	1117 (#10)	25
12	HappyHorse 1.0	Alibaba	1106 (#11)	26
13	Veo 3.1	Google	1056 (#15)	26
14	Kling 2.6 Pro	Kuaishou	1080 (#14)	26
15	Veo 3.1 Fast	Google	1020 (#17)	26
16	Wan 2.7	Alibaba Cloud	1012 (#18)	25
17	Wan 2.6	Alibaba Cloud	1022 (#16)	26
18	Vidu Q3 Pro	Shengshu	974 (#20)	26
19	Pixverse V5.6	AIsphere	934 (#22)	25
20	Wan 2.5 Preview	Alibaba Cloud	954 (#21)	26
21	Seedance V1.5 Pro	ByteDance	706 (#25)	27
22	Cosmos 2.5	NVIDIA	694 (#26)	22
23	Dream-X-World	Amap	495 (#28)	23
24	AlayaWorld	Alaya Lab	535 (#27)	25
25	SANA-WM	NVIDIA	392 (#29)	27
26	Yume 1.5	Shanghai AI Laboratory	233 (#30)	27

Per-leaderboard Elo rankings (click to expand)

with-reference

#	Model	Lab	Elo	±
1	Ground Truth	real recording	1476	38
2	Gemini Omni 1.1 Flash	Google	1366	39
3	MiniMax H3	MiniMax	1339	38
4	Gemini Omni Flash	Google	1286	38
5	Dreamina Seedance 2.0	ByteDance	1268	37
6	FLUX 3 Video	Black Forest Labs	1265	37
7	Dreamina Seedance 2.5	ByteDance	1247	37
8	Wan 3.0	Alibaba Cloud	1183	37
9	Grok Imagine Video 1.5	xAI	1143	37
10	Kling 3.0 Pro	Kuaishou	1123	36
11	Grok Imagine Video	xAI	1118	37
12	HappyHorse 1.0	Alibaba	1085	37
13	Veo 3.1	Google	1077	36
14	Veo 3.1 Fast	Google	1068	36
15	Kling 2.6 Pro	Kuaishou	1034	36
16	Wan 2.6	Alibaba Cloud	1023	36
17	Wan 2.7	Alibaba Cloud	1009	36
18	Vidu Q3 Pro	Shengshu	987	36
19	Pixverse V5.6	AIsphere	942	36
20	Wan 2.5 Preview	Alibaba Cloud	918	37
21	Cosmos 2.5	NVIDIA	726	32
22	Seedance V1.5 Pro	ByteDance	697	39
23	Dream-X-World	Amap	535	34
24	AlayaWorld	Alaya Lab	497	34
25	SANA-WM	NVIDIA	336	38
26	Yume 1.5	Shanghai AI Laboratory	251	39

blind

#	Model	Lab	Elo	±
1	MiniMax H3	MiniMax	1298 (#4)	39
2	Ground Truth	real recording	1407 (#1)	35
3	Gemini Omni 1.1 Flash	Google	1350 (#2)	38
4	Gemini Omni Flash	Google	1286 (#5)	38
5	Dreamina Seedance 2.0	ByteDance	1279 (#6)	37
6	FLUX 3 Video	Black Forest Labs	1249 (#7)	37
7	Dreamina Seedance 2.5	ByteDance	1199 (#9)	36
8	Wan 3.0	Alibaba Cloud	1209 (#8)	37
9	Grok Imagine Video 1.5	xAI	1154 (#10)	37
10	Grok Imagine Video	xAI	1148 (#11)	37
11	Kling 3.0 Pro	Kuaishou	1121 (#12)	36
12	HappyHorse 1.0	Alibaba	1070 (#14)	36
13	Veo 3.1	Google	1086 (#13)	37
14	Kling 2.6 Pro	Kuaishou	1066 (#15)	37
15	Veo 3.1 Fast	Google	1039 (#16)	36
16	Vidu Q3 Pro	Shengshu	1013 (#17)	36
17	Wan 2.7	Alibaba Cloud	990 (#19)	36
18	Pixverse V5.6	AIsphere	963 (#21)	36
19	Wan 2.6	Alibaba Cloud	983 (#20)	36
20	Wan 2.5 Preview	Alibaba Cloud	923 (#22)	37
21	Seedance V1.5 Pro	ByteDance	756 (#25)	38
22	Cosmos 2.5	NVIDIA	699 (#26)	31
23	Dream-X-World	Amap	580 (#27)	32
24	AlayaWorld	Alaya Lab	465 (#28)	35
25	Yume 1.5	Shanghai AI Laboratory	321 (#30)	37
26	SANA-WM	NVIDIA	349 (#29)	37

Elo vs. list price per generated video second (log scale). Live version on the benchmark page.

How the dataset was built

1. Scenarios (65 from Physics-IQ)

Physics-IQ (Motamed et al., 2025) filmed 66 real-world physical events — collisions, falling and rolling objects, fluids, fire, shadows and reflections, magnets — each from three perspectives and in two takes, at 3840×2160 and 30 fps. We use 65 scenarios, one take and one perspective each. For every scenario the model input is:

• Image: the Physics-IQ switch frame — the last frame before the outcome becomes visible.
• Text: the human-written Physics-IQ scene description. It describes the set-up without giving away what happens next, and ends with "Static shot with no camera movement."

The prompt column of this dataset is that description; the scenario id (0002 … 0197) is the Physics-IQ scene id and is what links the four generations of one model, and all models, to the same event.

2. Generation (25 models, 4 outputs each)

Every model generated 4 videos per scenario (~5 s, image-to-video with the text prompt) to average out sampling noise; a model competes with all four. The real 5-second continuation was added as the participant Ground Truth (one clip per scenario) so the leaderboards have an anchor. A handful of scenarios failed for some models, so a few participants have 253–259 rather than 260 clips.

The 25 models, grouped by lab:

Google — Gemini Omni 1.1 Flash, Gemini Omni Flash, Veo 3.1, Veo 3.1 Fast · MiniMax — MiniMax H3 · ByteDance — Dreamina Seedance 2.0, Dreamina Seedance 2.5, Seedance V1.5 Pro · Black Forest Labs — FLUX 3 Video · Alibaba Cloud — Wan 3.0, Wan 2.7, Wan 2.6, Wan 2.5 Preview · Alibaba — HappyHorse 1.0 · xAI — Grok Imagine Video, Grok Imagine Video 1.5 · Kuaishou — Kling 3.0 Pro, Kling 2.6 Pro · Shengshu — Vidu Q3 Pro · AIsphere — Pixverse V5.6 · NVIDIA — Cosmos 2.5, SANA-WM · Amap — Dream-X-World · Alaya Lab — AlayaWorld · Shanghai AI Laboratory — Yume 1.5

Cosmos 2.5, SANA-WM, Dream-X-World, AlayaWorld and Yume 1.5 are world models; the rest are video generation models.

3. Human evaluation

The clips were uploaded to a Rapidata MRI benchmark as one participant per model. Two leaderboards were run over the same pool of matchups:

• Every matchup pairs two clips of the same scenario from two different participants and is shown to at least 5 annotators; the Rapidata matchmaker spends extra votes on pairs and models where the ranking is still uncertain (many pairs have 10).
• Annotators watch both clips side by side and answer "Which video is more realistic?". On the with-reference leaderboard the real continuation plays above the two candidates; on blind it does not.
• Each vote is weighted by the annotator's userScore, Rapidata's ongoing quality estimate for that person, and carries their country, language, gender and age bucket.
• Rankings are Bradley–Terry / Elo fits over all weighted votes; the overall score averages the two leaderboards.

Dataset structure

One row per scenario × pair of participants (65 × 325 = 21,125 rows). Each participant has four clips per scenario, so a row aggregates the votes over all clip pairs of that model pair; video1 / video2 point to one representative clip each. The two weighted_results_* columns of a leaderboard are the userScore-weighted vote sums for each side, not probabilities — divide by their sum for a win share. A pair that was only sampled on one leaderboard has null in the other leaderboard's columns.

Column	Type	Description
prompt	string	the Physics-IQ scene description both clips were generated from
video1	string	public URL of the first clip (https://assets.rapidata.ai/….mp4)
video2	string	public URL of the second clip
model1	string	participant that produced video1 (e.g. Ground Truth, Wan 3.0)
model2	string	participant that produced video2
weighted_results_video1_blind	float32	weighted votes for video1 on the blind leaderboard
weighted_results_video2_blind	float32	weighted votes for video2 on the blind leaderboard
detailed_results_blind	string	every blind vote as a JSON array — votedFor (A = video1, B = video2), userScore, country, language, gender, age
weighted_results_video1_with_reference	float32	weighted votes for video1 on the with-reference leaderboard
weighted_results_video2_with_reference	float32	weighted votes for video2 on the with-reference leaderboard
detailed_results_with_reference	string	every with-reference vote, same shape as above

from datasets import load_dataset

ds = load_dataset("Rapidata/physics", split="train")
row = ds[0]
print(row["prompt"])
print(row["model1"], "vs", row["model2"])
a, b = row["weighted_results_video1_with_reference"], row["weighted_results_video2_with_reference"]
if a is not None:
    print(f"with-reference win share for {row['model1']}: {a / (a + b):.0%}")
print(row["video1"], row["video2"])  # stream or download the mp4s

Add your model

The benchmark is open: generating the 260 clips and submitting them takes one notebook run. The inputs (inputs.csv with scene id, prompt and switch-frame URL) and the submission notebook live in RapidataAI/scripts-odyssey/physics. New participants are matched against the whole field on both leaderboards and appear on the live benchmark.

Licensing & attribution

• Scenarios, prompts and ground-truth recordings come from Physics-IQ by Motamed, Culp, Swersky, Jaini and Geirhos (Google DeepMind / INSAIT), introduced in Do generative video models understand physical principles? (arXiv:2501.09038, 2025), and remain under the terms of that benchmark.
• Generated videos are outputs of third-party models and are governed by the terms of use of each respective model provider.
• Human annotations were collected via Rapidata and are released under CC-BY-4.0.

Citation

If you use this dataset, please cite it:

@misc{teikari2026physics,
  title     = {{image-to-video-physics-human-feedback-benchmark}},
  author    = {Teikari, Elias and Kennedy, James and Paravicini, Jorge and Sundararajan, Aniruddha and Tollis, Manuel and Vikat, Karl and Corkill, Jason},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/10986},
  url       = {https://huggingface.co/datasets/Rapidata/image-to-video-physics-human-feedback-benchmark}
}

Teikari, E., Kennedy, J., Paravicini, J., Sundararajan, A., Tollis, M., Vikat, K., & Corkill, J. (2026). image-to-video-physics-human-feedback-benchmark [Data set]. Hugging Face. https://doi.org/10.57967/hf/10986

About Rapidata

Rapidata's technology makes collecting human feedback at scale faster and more accessible than ever before. Visit rapidata.ai to learn more about how we're revolutionizing human feedback collection for AI development.

Explore our latest model rankings on our website.