Dataset Card for Moonworks Lunara Aesthetic II

This dataset introduces the second open-source release by Moonworks. This dataset contains original image and art created by Moonworks and their contextual variations generated by Moonworks Lunara, a sub-10B parameter model with a novel diffusion mixture architecture.

Paper: https://arxiv.org/pdf/2602.01666

While part 1 is intended for learning and evaluating regional as well as region-agnostic art styles, part 2 is intended for learning contextual variations while maintaining high aesthetic value.

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

Sample Image Pairs

Each pair shows an original image (left) and its corresponding variant (right).

---

Minimal Usage Example (Colab)

The snippet below loads the dataset from the Hugging Face Hub and displays an original / variant image pair side by side.

from datasets import load_dataset
import matplotlib.pyplot as plt
import random
import io
from PIL import Image as PILImage

def to_pil(x):
    """
    Convert Hugging Face Image feature outputs (or streaming outputs)
    into a real PIL.Image.Image for matplotlib.
    Supports:
      - already-PIL
      - dict with 'bytes'
      - dict with 'path'
      - plain path string
    """
    if isinstance(x, PILImage.Image):
        return x

    if isinstance(x, dict):
        if x.get("bytes") is not None:
            return PILImage.open(io.BytesIO(x["bytes"])).convert("RGB")
        if x.get("path") is not None:
            return PILImage.open(x["path"]).convert("RGB")

    if isinstance(x, str):
        return PILImage.open(x).convert("RGB")

    raise TypeError(f"Unsupported image type: {type(x)}; value={repr(x)[:200]}")

# Stream (fast startup, no full split materialization)
ds_stream = load_dataset(
    "moonworks/lunara-aesthetic-image-variations",
    split="train",
    streaming=True,
)

buffer_size = 300
buffer = list(ds_stream.take(buffer_size))

k = 5
k = min(k, len(buffer))
samples = random.sample(buffer, k)

fig, axes = plt.subplots(k, 2, figsize=(12, 3.6 * k))
if k == 1:
    axes = [axes]  # normalize shape

for i, sample in enumerate(samples):
    orig = to_pil(sample["original_image"])
    var  = to_pil(sample["variant_image"])

    axes[i][0].imshow(orig)
    axes[i][0].set_title("Original", fontsize=12)
    axes[i][0].axis("off")

    axes[i][1].imshow(var)
    axes[i][1].set_title("Variant", fontsize=12)
    axes[i][1].axis("off")

plt.subplots_adjust(hspace=0.15, wspace=0.02)
plt.show()

---

Dataset Summary

paper: https://arxiv.org/abs/2602.01666

The Moonworks Lunara Aesthetic II Dataset is a compact image variation dataset designed for studying identity preservation and contextual consistency in image editing and image-to-image generation.

Each sample consists of:

• an original anchor image
• a variant image with controlled contextual or aesthetic changes

---

Intended Use

This dataset is intended for:

• Benchmarking image editing and image variation models
• Evaluating identity preservation under aesthetic change
• Qualitative analysis of contextual transformations
• Research on controlled image-to-image generation

---

Citation

If you use this dataset, please cite:

@article{wang2026lunaraII,
  title={Moonworks Lunara Aesthetic II: An Image Variation Dataset},
  author={Wang, Yan and Hassan, Partho and Sadeka, Samiha and Soliman, Nada and Abdullah, M M Sayeef and Hassan, Sabit},
  journal={arXiv preprint arXiv:2602.01666},
  year={2026}
}