You agree to use this dataset under the CC-BY-4.0 license with attribution to Datapoint AI, and you agree not to attempt to re-identify annotators. Usage rights for the images are governed by each model provider's terms.

Text-to-image human preferences: 2M votes across 30 models

This dataset contains the complete voting record behind the Datapoint Image Bench leaderboard: 2,161,160 validated pairwise votes — exactly 10 for each of 216,116 image pairs. The votes compare 30 text-to-image models in a complete round-robin on 500 prompts, judged by annotators from over 200 countries. Every vote includes the annotator's trust score at the time the vote was cast.

Built on the Datapoint annotation platform — purpose-built infrastructure for collecting high-quality human preference data at scale.

To the best of our knowledge, this is the largest open text-to-image human-preference dataset published to date. For comparison, Pick-a-Pic v2 contains about 960,000 judgments, HPS v2 about 800,000, and ImageReward about 137,000.

Key features

Complete pair coverage. All 435 model pairings meet on every prompt that both models rendered. This is a full round-robin, not sparse arena sampling. Each pairing has about 5,000 direct votes.
Stress-test prompt set. 500 prompts across 10 categories, 50 per category. The prompts combine constraints such as exact object counts, typography, negation, ordering, and composition, which lets human judges separate models that perform similarly on simple prompts. Each prompt includes a category, difficulty, tags, and a 7-item scoring rubric.
Controlled generation. Each model generated one image per prompt with a single fixed seed. Prompts were passed verbatim, with no prompt rewriting, no best-of-N sampling, and no internal candidate selection on any model.
Blind, randomized judging. Annotators saw two unlabeled images side by side with the prompt and one question: "Which image do you prefer?" The left/right order was randomized for each pair, and model names were never shown. Each pair received 10 votes.
Trust scores on every vote. Each vote includes the annotator's quality score, a value from 0 to 1 that was recorded when the vote was cast. Every released vote passed the platform's quality enforcement.

Dataset structure

Config	Rows	Description
`pairs` (default)	216,116	One image pair with both images, labels, and the winner
`responses`	2,161,160	One individual human vote
`images`	14,952	One generated image at full resolution
`prompts`	500	One prompt with category, difficulty, tags, and rubric
`models`	30	One model with organization, API identifier, and price per image

pairs

Each row contains one comparison. The embedded image_a and image_b are display-size copies (JPEG, up to 1024 px on the long side), which is the size annotators judged at. They render in the dataset viewer. Use image_a_key and image_b_key to join the full-resolution originals in the images config. Side a is the image annotators saw on the left.

Column	Type	Description
`pair_key`	string	Unique pair identifier
`category`, `prompt_id`, `prompt`	string	Prompt and its category
`model_a`, `model_b`	string	The two models being compared
`image_a`, `image_b`	image	Display-size copies of the two images
`image_a_key`, `image_b_key`	string	Join keys into the `images` config
`votes_a`, `votes_b`	int	Vote counts used by the leaderboard
`label_a`, `label_b`	float	Preference fractions. Ties are 0.5/0.5. Use these directly for DPO or reward-model training
`trust_weighted_votes_a`, `trust_weighted_votes_b`	float	Vote counts weighted by trust score
`winner`	string	`a`, `b`, or `tie`
`num_votes`	int	Total votes for the pair

responses

Each row contains one human vote.

Column	Type	Description
`pair_key`	string	Joins to `pairs`
`chosen`	string	`a` or `b`, the side the annotator preferred
`annotator`	string	Salted hash. Stable across the dataset, not linkable to accounts
`trust_score`	float	Annotator quality score (0 to 1) at the time of the vote
`time_taken_ms`	int	Time spent on the judgment. The median is about 11 seconds
`country`	string	Annotator country (ISO 3166-1 alpha-2)

Every row counted toward the published leaderboard: the config contains exactly the 10 validated votes per pair that the Elo fit consumed.

Splits

The pairs and responses configs have train and test splits. The test split holds out 50 prompts (5 per category, selected with a fixed seed): 21,663 pairs and 216,630 responses. No prompt appears in both splits, so you can train a reward model on train and evaluate it on test without prompt leakage. The images, prompts, and models configs are reference tables without splits.

Usage

from datasets import load_dataset

pairs = load_dataset("datapointai/text-2-image-human-preferences-2m", "pairs", split="train")
row = pairs[0]
print(row["prompt"])
print(row["model_a"], "vs", row["model_b"], "->", row["winner"], f'({row["votes_a"]}-{row["votes_b"]})')

Use for DPO training

With 10 votes per pair, you can select training pairs by preference margin — something single-vote datasets can't offer. Filtering to decisive pairs (for example, 7–3 or stronger) removes near-tie label noise:

# Convert decisive pairs into chosen/rejected examples.
MIN_MARGIN = 0.7  # keep pairs where the winner took >= 70% of votes

def to_dpo(row):
    confidence = max(row["label_a"], row["label_b"])
    if row["winner"] == "tie" or confidence < MIN_MARGIN:
        return None
    chosen, rejected = ("a", "b") if row["winner"] == "a" else ("b", "a")
    return {
        "prompt": row["prompt"],
        "chosen": row[f"image_{chosen}"],
        "rejected": row[f"image_{rejected}"],
        "confidence": confidence,
    }

dpo_rows = [d for d in (to_dpo(r) for r in pairs) if d]

Work with individual votes

# Every vote carries the annotator's trust score, timing, and country.
responses = load_dataset("datapointai/text-2-image-human-preferences-2m", "responses", split="train")
high_trust = responses.filter(lambda r: r["trust_score"] is not None and r["trust_score"] >= 0.8)

Get full-resolution images

# Pair rows embed display-size copies. Join image_a_key / image_b_key
# against the images config for the full-resolution originals.
images = load_dataset("datapointai/text-2-image-human-preferences-2m", "images", split="train")

Leaderboard

Elo ratings from a Bradley–Terry fit on raw votes, rounded to whole points for display and anchored at FLUX.1 [schnell] = 1000. The chart shows the overall board published on 2026-08-19, computed from all 10 categories with equal weight. It matches the live leaderboard.

Image model Elo rankings — 30 models ranked by Elo score

Key findings

The podium is contested: the top 3 models sit within 10 Elo of each other, inside their confidence intervals.
No single model dominates. Three different models hold first place across the 10 category boards: Seedream 5.0 Pro leads 5 categories, GPT Image 2 (high) leads 3, and Nano Banana 2 leads 2 — yet GPT Image 2 (high) wins overall on consistency.
Frontier models are genuinely close for human judges: 18.2% of pairs ended in a tie on raw votes.

To reproduce the leaderboard, fit a Bradley–Terry model on the full responses config, grouped by pair. The published board uses raw (unweighted) votes and gives each of the 10 categories an equal share of the fit. The trust_weighted_votes_* columns support a sensitivity analysis: refit with each vote weighted by its trust_score and compare.

Comparison to related work

	This dataset	Pick-a-Pic v2	ImageReward	HPD v2
Human judgments	2,161,160	~960K	~137K	~798K
Votes per comparison	10	1–2	3–5	~3
Fixed model roster	30 models	No	No	No
Complete round-robin coverage	Yes	No	No	No
Trust score on every vote	Yes	No	No	No
Per-vote records (timing, country)	Yes	No	No	No
Published live leaderboard	Yes	No	No	No

Intended use

Use this dataset to:

Train and evaluate reward or preference models for text-to-image generation, including DPO, RLHF, and best-of-N reranking.
Study inter-annotator agreement and annotation quality at scale.
Benchmark aggregation methods such as Bradley–Terry variants and trust weighting.
Audit the published leaderboard.

The images are single-seed model outputs collected for evaluation. They are not a curated training corpus for image generation. Don't use this dataset to attempt to identify annotators.

License

Votes, prompts, and all metadata are released under CC-BY-4.0. The images are outputs of the listed third-party models and are distributed for research and evaluation. Usage rights for model outputs are governed by each provider's terms.

More Datapoint datasets

text-2-image-dpo-human-preferences-full — 416k pairwise image judgments on two evaluation dimensions
text-2-video-ranking-human-preferences — 91k ranking labels across 18 text-to-video models
image-2-video-human-preferences-large — image-to-video preference data
Full catalog: huggingface.co/datapointai

Citation

@dataset{datapoint_t2i_preferences_2m_2026,
  title   = {Text-to-Image Human Preferences 2M: the Datapoint Image Bench voting record},
  author  = {{Datapoint AI}},
  year    = {2026},
  url     = {https://huggingface.co/datasets/datapointai/text-2-image-human-preferences-2m},
  note    = {2,161,160 leaderboard-grade pairwise votes over 30 models, 500 prompts, full round-robin}
}

Built with Datapoint

This dataset was collected using Datapoint — a data labelling platform designed for high-quality human preference data at scale. Questions or feedback: sales@trydatapoint.com

Downloads last month: 193

Total file size:

34.9 GB

Papers for datapointai/text-2-image-human-preferences-2m

Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

Paper • 2306.09341 • Published Jun 15, 2023 • 1

Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation

Paper • 2305.01569 • Published May 2, 2023 • 2

ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation

Paper • 2304.05977 • Published Apr 12, 2023 • 4