moonworks/lunara-art-eval

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

FLUX.2 [klein] 4B	Moonworks Lunara	Z-Image-Turbo
HiDream-I1 Fast	Moonworks Lunara	Qwen-Image
AuraFlow	Moonworks Lunara	GPT Image 1 Mini

Dataset at a glance

Property	Description
Examples	8,000 images with prompts and scores
Evaluation prompts	1,000 shared prompts
Models	8; the paper reports 1,000 images per model
Prompt design	500 prompts covering 14 artistic styles, plus 500 open-ended art prompts
Automated judge	GPT-5.6 Sol
Reported image resolution	1024 × 1024
Storage	19 Parquet shards with embedded image bytes
Public columns	image, prompt, model, aesthetic, emotional_evocation, content_integrity

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.

Field	Type	Meaning
image	Hugging Face Image	Generated image, stored as embedded binary content in Parquet. Decodes to a Pillow image when loaded with datasets.
prompt	String	Source text prompt used to generate the image.
model	String	Display name of the model that generated the image.
aesthetic	Floating-point score	Aesthetic Quality: visual composition, color, lighting, texture, and artistic expression.
emotional_evocation	Floating-point score	Emotional Resonance: the affective impression and expressive impact of the image.
content_integrity	Floating-point score	Content Integrity: the completeness and structural coherence of requested subjects, including anatomy and unintended artifacts.

Benchmark results: Table 1

Model	Aesthetic Quality ↑	Aesthetic win rate	Emotional Resonance ↑	Emotional win rate	Content Integrity ↑	Content win rate	CLIPScore ↑	LAION Aesthetic ↑
Moonworks Lunara	8.473	23.3%	8.119	17.5%	7.944	4.7%	0.792	6.721
GPT 1 Image Mini	8.457	18.9%	8.363	31.7%	8.609	28.8%	0.777	6.661
Qwen Image	8.366	15.9%	7.979	10.7%	8.345	29.6%	0.789	6.561
AuraFlow	8.313	18.9%	8.011	15.9%	6.662	1.1%	0.781	6.745
SD 3.5 Turbo	8.081	11.4%	7.949	17.4%	7.238	3.2%	0.743	6.625
HiDream-I1 Fast	8.035	5.1%	7.568	3.5%	7.794	9.2%	0.764	6.710
FLUX-Klein-4B	8.023	5.4%	7.579	3.2%	7.438	9.0%	0.773	6.706
Z-Image-Turbo	7.230	1.2%	6.721	0.7%	7.875	17.8%	0.719	6.033

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}, 
}

research@moonworks.ai.