datapointai/image-2-video-human-preferences-medium

I2V Human Preferences (Medium)

Human preference dataset for image-to-video (I2V) generation quality. Each row contains a reference image, two generated videos (one from Pika and one from CogVideoX), and 10 human preference annotations aggregated via majority vote.

This is the medium (2,000-row) subset. See also: small (1,000 rows), large (3,000 rows).

Dataset Summary

Metric	Value
Total rows	2,000
Annotations per row	10
Total annotations	20,000
Unique prompts	2,000
Models compared	Pika vs CogVideoX
Media format	MP4 (video), JPG (reference image)

Key Findings

Metric	Value
Pika wins	945 (47.2%)
CogVideoX wins	703 (35.1%)
Ties	352 (17.6%)
Pika overall win rate	53.0%
CogVideoX overall win rate	47.0%
Mean annotator agreement	67.0%

Pika is preferred overall but the contest is fairly close. The 18% tie rate (5v5 splits across 10 annotators) indicates genuine ambiguity on many prompts.

Task Setup

Each datapoint consists of:

1. A reference image (the input to both I2V models)
2. A text prompt describing the desired video
3. Two candidate videos — one generated by Pika, one by CogVideoX

Annotators were shown the reference image and both videos side by side and asked to choose which video better represents the prompt given the reference image. Display order was randomized to prevent position bias.

Dataset Structure

Columns

Column	Type	Description
prompt_id	string	UUID linking the reference image and both candidate videos
prompt	string	Text prompt used for video generation
reference_image	Image	JPG reference image (input to both models)
pika_video	Video	MP4 video generated by Pika
cog_video	Video	MP4 video generated by CogVideoX
winner	string	Majority-vote label: pika, cog, or tie
pika_votes	int	Number of annotators who preferred Pika (out of 10)
cog_votes	int	Number of annotators who preferred CogVideoX (out of 10)
pika_win_rate	float	Fraction of annotators who preferred Pika
cog_win_rate	float	Fraction of annotators who preferred CogVideoX
agreement	float	Inter-annotator agreement (fraction that voted with majority)
detailed_results	string (JSON)	Per-annotator votes with winner and time_taken_ms

Detailed Results Structure

Each entry in the detailed_results JSON array:

{
  "winner": "pika",
  "time_taken_ms": 16411
}

Usage

from datasets import load_dataset

ds = load_dataset("datapointai/image-2-video-human-preferences-medium", split="train")
row = ds[0]
print(row["prompt"])
print(f"Winner: {row['winner']}")
print(f"Pika: {row['pika_votes']}/10, CogVideoX: {row['cog_votes']}/10")

Filter by Agreement

# High-agreement subset (≥80% annotator consensus)
high_agreement = ds.filter(lambda x: x["agreement"] >= 0.8)
print(f"{len(high_agreement)} rows with ≥80% agreement")

Train a Reward Model

from datasets import load_dataset

ds = load_dataset("datapointai/image-2-video-human-preferences-medium", split="train")

for row in ds:
    prompt = row["prompt"]
    reference = row["reference_image"]
    winner = row["winner"]
    pika_score = row["pika_win_rate"]
    cog_score = row["cog_win_rate"]
    # Use as preference pairs for DPO, reward modeling, etc.
    # winner == "pika" → pika_video is preferred over cog_video
    # winner == "cog"  → cog_video is preferred over pika_video
    # winner == "tie"  → no clear preference (consider filtering)

Explore Ties

# Analyze the most ambiguous comparisons
ties = ds.filter(lambda x: x["winner"] == "tie")
print(f"{len(ties)} ties ({len(ties)/len(ds)*100:.1f}%)")

Data Collection

Annotations were collected through Datapoint AI's consumer app SDK using forced-choice pairwise comparison. Each comparison was annotated by 10 unique annotators. Annotators were shown the reference image alongside both candidate videos and asked which video better represents the prompt. Video display positions (left/right) were randomized per session to mitigate position bias.

Annotation Quality

• Majority vote: The winner field reflects the choice of >50% of annotators (or tie at exactly 50/50)
• Agreement scores: The agreement field measures consensus — higher values indicate stronger annotator agreement
• Time tracking: time_taken_ms in detailed results allows filtering out suspiciously fast annotations

License

CC-BY-4.0

Citation

@dataset{datapointai_i2v_preferences_2026,
  title={I2V Human Preferences: Pika vs CogVideoX},
  author={Datapoint AI},
  year={2026},
  url={https://huggingface.co/datasets/datapointai/image-2-video-human-preferences-medium},
  note={20,000 pairwise human preference labels for image-to-video generation quality}
}