---
license: apache-2.0
tags:
- image-editing
- image-variation
- identity-preservation
- diffusion
- image-generation
- aesthetics
- art
task_categories:
- image-to-image
- text-to-image
size_categories:
- 1K<n<10K
pretty_name: Aesthetic Variations
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: original_image
    dtype:
      image:
        decode: false
  - name: variant_image
    dtype:
      image:
        decode: false
  - name: semantic_category_changes
    sequence: string
  - name: base_prompt
    dtype: string
  - name: variation_prompt
    dtype: string
  - name: variation_set
    dtype: float32
  - name: prompt_topic
    dtype: string
  splits:
  - name: train
    num_bytes: 8451541428.066
    num_examples: 2854
  download_size: 2227202641
  dataset_size: 8451541428.066
---

# 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)**.

<table>
  <tr>
    <td><img src="2_original.jpg" width="240"/></td>
    <td><img src="2_variant.jpg" width="240"/></td>
    <td style="border-left:2px solid #ccc;"></td>
    <td><img src="6_original.jpg" width="240"/></td>
    <td><img src="6_variant.jpg" width="240"/></td>
  </tr>

  <tr>
    <td><img src="7_original.jpg" width="240"/></td>
    <td><img src="7_variant.jpg" width="240"/></td>
    <td style="border-left:2px solid #ccc;"></td>
    <td><img src="10_original.jpg" width="240"/></td>
    <td><img src="10_variant.jpg" width="240"/></td>
  </tr>

  <tr>
    <td><img src="12_original.jpg" width="240"/></td>
    <td><img src="12_variant.jpg" width="240"/></td>
    <td style="border-left:2px solid #ccc;"></td>
    <td><img src="13_original.jpg" width="240"/></td>
    <td><img src="13_variant.jpg" width="240"/></td>
  </tr>

  <tr>
    <td><img src="14_original.jpeg" width="240"/></td>
    <td><img src="14_variant.jpg" width="240"/></td>
    <td style="border-left:2px solid #ccc;"></td>
    <td><img src="a1_original.jpg" width="240"/></td>
    <td><img src="a1_variant.jpg" width="240"/></td>
  </tr>

  <tr>
    <td><img src="a2_original.jpg" width="240"/></td>
    <td><img src="a2_variant.jpg" width="240"/></td>
    <td style="border-left:2px solid #ccc;"></td>
    <td><img src="a3_original.jpeg" width="240"/></td>
    <td><img src="a3_variant.jpg" width="240"/></td>
  </tr>
</table>

---
## 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.

```python
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:

```bibtex
@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}
}