Moonworks Lunara Art Eval

Paper: Moonworks Lunara: Modeling Artistic Intelligence by Wang et al. (2026).

An image-generation evaluation dataset containing 8,000 generated images, their source prompts, model labels, and three per-image evaluation scores. It supports comparison of artistic quality, emotional expression, and content integrity across eight text-to-image models including GPT-Image-1-mini, HiDream-I1, Qwen-Image and Flux-Klein.

Dataset · Dataset Viewer · Paper · PDF

This follows the first two releases containing aesthetic images and aesthetic image variations:

Aesthetic dataset 1: https://huggingface.co/datasets/moonworks/lunara-aesthetic

Aesthetic dataset 2: https://huggingface.co/datasets/moonworks/lunara-aesthetic-image-variations

Created with Lunara

Lunara is built on a novel Diffusion Mixture Architecture and a new active-learning-inspired training paradigm that jointly optimize for creative exploration and content integrity. A selection of images generated by Moonworks Lunara is below.

Visual model comparisons

Dataset at a glance

Human evaluation, and evaluation protocol are described in Section 5.1 of the paper.

Data fields

Each row represents one generated image and its evaluation scores.

Benchmark results: Table 1

Source: Wang et al., Moonworks Lunara: Modeling Artistic Intelligence, Table 1.

Usage

Install dependencies

python -m pip install -U "datasets[vision]"

Load the dataset and inspect an image

from datasets import load_dataset

ds = load_dataset(
    "moonworks/lunara-art-eval",
    "default",
    split="train",
)

print(ds)
print(ds.column_names)

example = ds[0]
print("Model:", example["model"])
print("Prompt:", example["prompt"])
print("Aesthetic Quality:", example["aesthetic"])
print("Emotional Resonance:", example["emotional_evocation"])
print("Content Integrity:", example["content_integrity"])

# The image is decoded from bytes embedded in the Parquet shard.
example["image"].save("lunara_art_eval_example.png")

The first load downloads the Parquet files and caches them locally. Subsequent loads reuse the cache. To restrict a first inspection to the first shard, use:

from datasets import load_dataset

preview = load_dataset(
    "moonworks/lunara-art-eval",
    "default",
    data_files={"train": "data/train-00000-of-00019.parquet"},
    split="train",
)
print("Rows in first shard:", len(preview))  # 422 in this release


import math
import random
import textwrap
import matplotlib.pyplot as plt

# Build the comparison index without decoding images.
metadata = preview.select_columns(["prompt", "model"]).to_pandas()
metadata["row_id"] = range(len(preview))

models = sorted(
    metadata["model"].unique(),
    key=lambda name: (name != "Moonworks Lunara", name),
)

# Choose prompts that have an image from every model.
counts = metadata.groupby("prompt")["model"].nunique()
eligible = counts[counts == len(models)].index.tolist()

if not eligible:
    raise ValueError("No shared prompts found across all models.")

N_PROMPTS = 2
prompts = random.Random().sample(
    eligible, min(N_PROMPTS, len(eligible))
)

for prompt in prompts:
    positions = (
        metadata.loc[metadata["prompt"] == prompt]
        .drop_duplicates("model")
        .set_index("model")["row_id"]
    )

    columns = 4
    rows = math.ceil(len(models) / columns)
    fig, axes = plt.subplots(
        rows, columns,
        figsize=(16, 4.3 * rows),
        squeeze=False,
        constrained_layout=True,
    )

    for ax in axes.flat:
        ax.axis("off")

    for ax, model in zip(axes.flat, models):
        example = preview[int(positions[model])]
        ax.imshow(example["image"])
        ax.set_title(
            textwrap.fill(model, width=24),
            fontsize=12,
            fontweight="bold",
            color="#7c3aed" if model == "Moonworks Lunara" else "#333333",
        )

    fig.suptitle(
        textwrap.fill(f"Prompt: {prompt}", width=130),
        fontsize=13,
    )
    plt.show()

Select images from one model

This example uses the ds loaded above. Restricting the filter to model avoids decoding images while evaluating the condition.

lunara = ds.filter(
    lambda model: model == "Moonworks Lunara",
    input_columns=["model"],
)
print("Lunara examples:", len(lunara))
lunara[0]["image"].save("lunara_example.png")

Compute mean scores by model

The following analysis uses the metadata columns and skips image decoding. It computes means from the released per-image scores.

from collections import defaultdict
from statistics import mean

metrics = ("aesthetic", "emotional_evocation", "content_integrity")
grouped = defaultdict(lambda: {metric: [] for metric in metrics})

for row in ds.remove_columns("image"):
    for metric in metrics:
        grouped[row["model"]][metric].append(row[metric])

summary = [
    {
        "model": model,
        "n": len(values["aesthetic"]),
        **{metric: mean(values[metric]) for metric in metrics},
    }
    for model, values in grouped.items()
]
summary.sort(key=lambda row: row["aesthetic"], reverse=True)

for row in summary:
    print(
        f"{row['model']}: n={row['n']}, "
        f"aesthetic={row['aesthetic']:.3f}, "
        f"emotional_resonance={row['emotional_evocation']:.3f}, "
        f"content_integrity={row['content_integrity']:.3f}"
    )

Table 1 above records the published results. Recomputed means depend on the precision of the released scores. Reproducing CLIP, LAION, or win-rate results are not produced by this mean-score example.

Compare models on the same prompt

Prompt text links generations across models in this six-column release.

prompt_text = ds[0]["prompt"]
same_prompt = ds.filter(
    lambda prompt: prompt == prompt_text,
    input_columns=["prompt"],
)

for row in same_prompt.remove_columns("image"):
    print(
        row["model"],
        row["aesthetic"],
        row["emotional_evocation"],
        row["content_integrity"],
    )

License

Apache 2.0 The images generated by the different models carry their respective restrictions and licenses.

Citation

If you use this dataset or its published benchmark results, cite the paper:

@misc{wang2026moonworkslunaramodelingartistic,
      title={Moonworks Lunara: Modeling Artistic Intelligence}, 
      author={Yan Wang and Yanzu Wang and Maitreyee Joshi and Samiha Sadeka and Partho Hassan and Reza Jarral and Sayeef Abdullah and Sabit Hassan},
      year={2026},
      eprint={2609.22272},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.22272}, 
}

Previous paper citations:

@misc{wang2026moonworkslunaraaestheticdataset,
      title={Moonworks Lunara Aesthetic Dataset}, 
      author={Yan Wang and Sayeef Abdullah and Partho Hassan and Sabit Hassan},
      year={2026},
      eprint={2601.07941},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.07941}, 
}

@misc{wang2026moonworkslunaraaestheticii,
      title={Moonworks Lunara Aesthetic II: An Image Variation Dataset}, 
      author={Yan Wang and Partho Hassan and Samiha Sadeka and Nada Soliman and Sayeef Abdullah and Sabit Hassan},
      year={2026},
      eprint={2602.01666},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.01666}, 
}

[email protected].