Rapidata/relative-camera-movement

Rapidata Relative Camera Movement Benchmark

Built by Rapidata.

This dataset contains 900,318 human responses, collected with the Rapidata Python SDK, comparing how well 15 image-to-video models and world models move the camera relative to what is in the scene. Each row is a head-to-head comparison between two models' clips generated from the same still and the same instruction, judged by human annotators.

Our Camera Movement Benchmark asks for scene-agnostic moves — "tilt thirty degrees down", "truck left". This one asks for moves that only make sense in this picture: "slide the camera under the wooden fence, then lift it up to the red barn roof", "move the camera behind the dog, then switch it to the dog's point of view". To pass, a model has to understand the scene: which object is meant, where "behind" it is, what it would see.

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 — filter by model, movement type, number of motions and scene, and inspect 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)	22,691
Human votes	900,318 (22–44 per pair, 39.7 on average)
Annotator countries	125
Models compared	15 (12 commercial image-to-video models · 3 research world models)
Input stills	50
Prompts (hand-written, one still each)	220
Unique generated clips	3,270

The leaderboard

Leaderboard	Question shown to annotators	Prompt shown?	Measures
Alignment	"Which video matches the description better?"	Yes — the instruction is shown above the two clips	whether the camera went where the instruction said, relative to the objects it names

Examples

Each example shows the input still and the two generated clips with the share of the (userScore-weighted) votes each one received. The green border marks the winner.

Camera movement

Move the camera in close to the dials, then turn it to the tall tower on the right.

Input still

Wan 3.0

Alignment: 95% of votes

FLUX 3 Video

Alignment: 5% of votes

Perspective switching

Switch the camera to the cyclist's point of view.

Input still

Veo 3.1

Alignment: 90% of votes

MiniMax H3

Alignment: 10% of votes

Both: move, then switch perspective

Move the camera behind the tractor, then move it into the driver's seat.

Input still

MiniMax H3 Max

Alignment: 92% of votes

Vidu Q3 Pro

Alignment: 8% of votes

Overall ranking (ELO on the Alignment leaderboard)

#	Model	Lab	Type	ELO
1	Gemini Omni Flash 1.1	Google	commercial image-to-video	1390.7
2	Wan 3.0	Alibaba Cloud	commercial image-to-video	1383.4
3	Happy Horse 1.1	Alibaba	commercial image-to-video	1290.8
4	MiniMax H3 Max	MiniMax	commercial image-to-video	1271.6
5	Grok Imagine Video 1.5	xAI	commercial image-to-video	1226.0
6	MiniMax H3	MiniMax	commercial image-to-video	1152.9
7	Veo 3.1	Google	commercial image-to-video	1114.3
8	Dreamina Seedance 2.5	ByteDance	commercial image-to-video	1076.3
9	FLUX 3 Video	Black Forest Labs	commercial image-to-video	1058.2
10	Kling 3.0 Pro	Kuaishou	commercial image-to-video	1000.0
11	Vidu Q3 Pro	Shengshu AI	commercial image-to-video	977.9
12	Cosmos 3 Nano	NVIDIA	world model	717.5
13	Pixverse V5.6	AIsphere	commercial image-to-video	703.5
14	Cosmos Predict 2.5	NVIDIA	world model	452.6
15	Yume 1.5	Shanghai AI Laboratory	world model	184.3

The top two are effectively tied.

Win share by movement type and number of motions

Mean userScore-weighted win share per model across all of its matchups (50% = even with the field). The 25 prompts tagged with both movement types count in both columns.

Model	Camera movement	Perspective switching	1 motion	2 motions	Overall
Gemini Omni Flash 1.1	64%	66%	65%	64%	65%
Wan 3.0	64%	68%	64%	65%	65%
Happy Horse 1.1	61%	59%	61%	61%	61%
MiniMax H3 Max	61%	59%	57%	62%	60%
Grok Imagine Video 1.5	59%	55%	55%	60%	58%
MiniMax H3	56%	57%	53%	57%	55%
Veo 3.1	54%	58%	55%	54%	54%
Dreamina Seedance 2.5	52%	56%	55%	52%	53%
FLUX 3 Video	51%	57%	51%	53%	52%
Kling 3.0 Pro	49%	55%	53%	49%	50%
Vidu Q3 Pro	49%	49%	51%	48%	49%
Cosmos 3 Nano	40%	29%	37%	40%	39%
Pixverse V5.6	38%	37%	42%	36%	38%
Cosmos Predict 2.5	30%	21%	28%	29%	29%
Yume 1.5	21%	23%	23%	21%	22%

• Perspective switching separates the field most. Wan 3.0 wins 68% of its matchups there, while the three research world models drop to 21–29%.
• Kling 3.0 Pro and FLUX 3 Video are better at switching perspective than at moving (55% and 57%, against 49% and 51% on camera movement).
• MiniMax H3 Max gets stronger as the instruction gets longer: 57% on one-motion prompts, 62% on two.

The live leaderboard also lets you slice the standings by scene.

How the dataset was built

1. Input stills (50 images)

50 photographs of scenes with clearly nameable objects, people and animals to move relative to — a skateboarder on Ocean Drive, Rodin's Thinker on a plinth, a table saw in a timber workshop, a retriever running on sand, the back seat of a Bangkok taxi, a grand piano in an empty hall, a helicopter cockpit over a city. Each still carries between 2 and 8 prompts.

2. Prompts (220, hand-written per still)

Every prompt was written for one specific still and names things that are actually in it, so there is no cross product of stills and prompts. Each prompt is tagged on two axes, which the live leaderboard can filter by:

Movement type	Prompts	What it asks for	Example
Camera movement	194	move the camera along, past, under, around or towards named scene content	"Lower the camera to the floor, then slide it slowly under the table saw."
Perspective switching	51	cut or move to a character's own point of view	"Switch the camera to the cyclist's point of view."

25 prompts carry both tags: a camera move followed by a perspective switch, e.g. "Move the camera quickly behind the dog, then switch it to the dog's point of view."

Number of motions	Prompts
1 motion	73
2 motions	147

3. Video generation (15 models)

Each still and its prompts were sent to every model as an image-to-video request: the still is the first frame, the prompt is the instruction.

Alibaba — Wan 3.0, Happy Horse 1.1 · Google — Gemini Omni Flash 1.1, Veo 3.1 · MiniMax — H3, H3 Max · xAI — Grok Imagine Video 1.5 · ByteDance — Dreamina Seedance 2.5 · Black Forest Labs — FLUX 3 Video · Kuaishou — Kling 3.0 Pro · Shengshu AI — Vidu Q3 Pro · AIsphere — Pixverse V5.6

NVIDIA — Cosmos 3 Nano, Cosmos Predict 2.5 · Shanghai AI Laboratory — Yume 1.5 (research world models)

4. Human evaluation

The clips were uploaded to a Rapidata MRI benchmark as one participant per model, and the Alignment leaderboard was run over the resulting pairwise matchups. Annotators came from a dedicated Rapidata audience rather than the open global pool. Every pair received 22 to 44 votes (39.7 on average), from annotators in 125 countries, and per-annotator detail — chosen side, country, language, gender, age bucket, occupation and the annotator's userScore — is preserved in the detailed_results_alignment column.

Dataset structure

One row per head-to-head clip pair generated from the same still and the same instruction.

Column	Type	Description
prompt	string	the camera instruction both clips were generated from
prompt_asset	image	the input still both clips were generated from, embedded (JPEG)
video1	string	public URL of the clip from model1
video2	string	public URL of the clip from model2
model1	string	name of the model that produced video1
model2	string	name of the model that produced video2
weighted_results_video1_alignment	float32	sum of the userScore weights of the votes for video1
weighted_results_video2_alignment	float32	sum of the userScore weights of the votes for video2
detailed_results_alignment	string	every vote on the pair as a JSON array (votedFor is A for video1 and B for video2, plus annotator demographics and userScore)

The two weighted_results_* values are userScore-weighted vote sums, not probabilities — they do not sum to 1. Divide by their sum for a normalised win share, which is what the example percentages above show.

The embedded stills make the full download about 5 GB. Load only the columns you need to skip them:

import json
from datasets import load_dataset

ds = load_dataset("Rapidata/relative-camera-movement", split="train")
row = ds[0]
print(row["prompt"], "|", row["model1"], "vs", row["model2"])

w1, w2 = row["weighted_results_video1_alignment"], row["weighted_results_video2_alignment"]
print(f"win share {row['model1']}: {w1 / (w1 + w2):.0%}")

votes = json.loads(row["detailed_results_alignment"])
print(len(votes), "votes, first one:", votes[0])

row["prompt_asset"].show()  # the input still, as a PIL image

Licensing & attribution

This dataset combines material under different terms:

• Input stills. Third-party photographs, included for research and evaluation purposes; rights remain with their respective owners.
• Prompts. Written by the dataset creators, released under CC-BY-4.0.
• Generated clips. Produced by third-party models. Model outputs are governed by the terms of use of each respective model provider.
• Human annotations. Collected via Rapidata, released under CC-BY-4.0.

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.

Related datasets

• Rapidata/camera-movement — scene-agnostic camera moves (dolly, truck, pan, tilt, roll, zoom) judged against a reference animation.
• Rapidata/world-model-physics — physical plausibility of world-model and video-model rollouts.