File size: 10,923 Bytes
48c6a5b a1c16be 48c6a5b dc83b14 48c6a5b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 | ---
pretty_name: PixelExact — 1.4M Sprites with Visual Descriptions
language:
- en
size_categories:
- 1M<n<10M
task_categories:
- image-to-text
- text-to-image
- image-text-to-image
- image-text-to-image
tags:
- pixel-art
- sprites
- synthetic
- ai-generated
- image-captioning
- multimodal
- game-assets
configs:
- config_name: default
data_files:
- split: train
path: data/train-*.parquet
- split: validation
path: data/validation-*.parquet
license: cc-by-4.0
---

# PixelExact
### Tiny sprites. Serious visual supervision.
**1,409,404 sprites · Native 16×16 RGBA · New visual descriptions · Structured parts and colors**
**Pixel art compresses an entire idea into 256 pixels. Its training data should make those pixels count.**
PixelExact adds a new visual annotation layer to [Angais’s Pixel Art Atlas](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m): natural-language descriptions and structured observations of visible parts and colors, produced by a two-model annotation pipeline.
The original subject labels and captions stay available. The new `description` field gives you a second, image-conditioned account of the sprite. The `visible_parts` field breaks observations into parts and colors you can inspect, query, and use in your own pipelines.
**Same visible artwork. A much richer interface to it.**
[Inspect a row](#exhibit-a-windmill) · [Choose your text](#choose-what-your-model-reads) · [Load the data](#open-the-box) · [How it was made](#how-it-was-made)
## Exhibit A: windmill
These fields come from an actual project sample:
<table>
<tr><th align="left">The source label</th><th align="left">The visual description</th></tr>
<tr>
<td valign="top">yellow windmill</td>
<td valign="top">A pixelated windmill with yellow and black blades and an orange tower on a transparent background.</td>
</tr>
<tr><td><code>caption</code></td><td><code>description</code></td></tr>
</table>
“yello windmill” becomes “windmill with yellow and black blades and an orange tower” The description also introduces a new highlights from the image.
That difference is worth keeping. It gives you new text conditions to train on. The description is a model’s interpretation, not a human ruling on what the sprite contains.
The annotation models received the image and an unverified source subject hint. **They did not receive the original generation caption.** Their instructions favor visible details and broader names when a precise identity is unclear.
## Choose what your model reads
| Your experiment | Use | Why it is here |
|---|---|---|
| Learn from the visual description | `description` | An image-conditioned account of the sprite. |
| Reproduce the original text condition | `caption` | The source caption is preserved. |
| Work with subject labels | `subject` | The original category remains available. |
| Query a part or a color | `visible_parts` | Observations are stored as named fields. |
| Compare the same sprite across experiments | `row_id` | Original identifiers remain intact. |
**You do not have to throw away the old caption to try a new one.**
## The whole collection, accounted for
| Property | PixelExact |
|---|---:|
| Sprites | **1,409,404** |
| Training rows | **1,405,308** |
| Validation rows | **4,096** |
| Native image size | **16 × 16 RGBA** |
| Final columns | **6** |
| Image storage | **Embedded PNG bytes in Parquet** |
The images are embedded inside the Parquet files. Each row carries its sprite, its source text, and its visual annotations together.
Transparency is preserved. Images are exported as lossless PNGs at native resolution. Fully transparent pixels have their invisible RGB values set to zero during preparation; the PNG files are re-encoded, so their file bytes differ from the source files while preserving the visible artwork.
## The file you actually train on
| Column | Type | Contents |
|---|---|---|
| `row_id` | `int32` | Original source identifier. |
| `image` | `Image` | Embedded 16 × 16 RGBA PNG. |
| `subject` | `string` | Original source subject. |
| `caption` | `string` | Original source caption. |
| `description` | `string` | PixelExact’s model-generated visual description. |
| `visible_parts` | `list<struct<part: string, color: string>>` | Model observations of parts and their colors. |
## Built for the next wave of pixel-art tools
PixelExact provides a foundation for experiments in:
- **Text-to-sprite generation:** train with the visual descriptions or compare them against the original captions.
- **Sprite captioning:** study how models describe extremely small visual inputs.
- **Text–image retrieval:** explore searching sprite collections with natural-language descriptions.
- **Attribute-aware browsing:** inspect structured parts and colors alongside the image.
- **Multimodal representation learning:** investigate alignment between sprites and different sources of text supervision.
- **Dataset research:** analyze where generation labels and image-conditioned descriptions agree or diverge.
**Generation. Retrieval. Captioning. Inspection. The same six columns support all four.**
These are intended applications, not reported downstream benchmark results.
## Three experiments worth running
### Change the caption. Keep the image.
Hold the image IDs, split membership, architecture, and training budget fixed. Train once with `caption` and once with `description`. Compare the results on the same held-out examples.
PixelExact gives you both text conditions for the same images. The comparison is yours to measure.
### Follow the disagreements.
Find examples where the source caption and the visual description use different colors, details, or subject names. Inspect those images. Build a targeted audit around the cases that matter to your application.
The slime example is an invitation to investigate, not an accuracy score.
### Give the sprite browser better questions.
Use `description` for text retrieval and `visible_parts` for structured filtering. Try a part/color query, compare it with a subject-only search, and inspect which sprites each approach returns.
These are proposed experiments and applications. Downstream benchmark results are not included in this release.
## Open the box
The repository uses the same split layout as the source collection:
```text
PixelExact/
README.md
banner.png
export_manifest.json
data/
train-00000.parquet
train-00001.parquet
...
train-00028.parquet
validation-00000.parquet
```
Install the loader:
```bash
pip install -U datasets pillow
```
Stream directly from [onlyaady/PixelExact](https://huggingface.co/datasets/onlyaady/PixelExact) on Hugging Face:
```python
from datasets import load_dataset
sprites = load_dataset(
"onlyaady/PixelExact",
split="train",
streaming=True,
)
row = next(iter(sprites))
print(row["row_id"], row["subject"])
print("Source: ", row["caption"])
print("Description:", row["description"])
print("Parts: ", row["visible_parts"])
print("Image: ", row["image"].size, row["image"].mode)
```
For a text–image training pair:
```python
image = row["image"]
text = row["description"]
```
The `streaming=True` option iterates over the data without loading the whole collection into memory. The image bytes are already in the Parquet row.
For local training or random access, download and cache both splits:
```python
from datasets import load_dataset
sprites = load_dataset("onlyaady/PixelExact")
train = sprites["train"]
validation = sprites["validation"]
print(len(train)) # 1,405,308
print(len(validation)) # 4,096
```
## How it was made
<details>
<summary><strong>Annotation passes, coverage checks, and manual review</strong></summary>
**Two annotation passes.** [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) produced the first pass. A selected subset of 395,454 sprites received a second pass using [Gemma 4 31B](https://huggingface.co/google/gemma-4-31B).
The final annotation layer contains:
| Final annotation source | Rows |
|---|---:|
| Qwen | 1,013,950 |
| Gemma | 395,454 |
| **Total** | **1,409,404** |
All completed Gemma rows are included. The remaining annotations come from the retained Qwen rows.
**Coverage was checked.** The completed annotation runs reported full expected coverage, with no missing, error, or invalid rows. Coverage and output validity do not establish caption correctness.
**A sample was examined manually.** The reviewers reported no issues in 5000 inspected samples from the Gemma second pass. This was a spot-check; descriptions and part/color observations throughout the release remain model-generated and can contain mistakes. Source subject hints can influence them, and original source captions remain approximate labels.
**The split is inherited.** The 4,096-row validation partition comes from Pixel Art Atlas. It does not establish unseen subjects or absence of duplicate images across train and validation. Account for those properties when setting up an evaluation.
</details>
## Where the artwork comes from
[Angais](https://huggingface.co/Angais) created and released **[Pixel Art Atlas](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m)**, including the sprite collection, its original captions, and its subject labels. The source documentation attributes image generation to Krea 2 Turbo with a pixel-art LoRA.
**PixelExact contributes the visual descriptions, the structured part/color observations, and the annotation/export pipeline.** It preserves attribution to the source artwork.
Angais releases the source collection under CC0 1.0 to the extent of the rights Angais can waive. Read the source’s [rights and provenance statement](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m#rights-and-provenance) for its terms and qualifications.
## License
PixelExact’s added visual descriptions, structured annotations, and dataset curation are released under **[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)**.
You may use, share, and adapt these contributions for research and commercial purposes. Where the license applies, credit [onlyaady/PixelExact](https://huggingface.co/datasets/onlyaady/PixelExact), link to the license, and indicate any changes.
The original sprites, captions, and subject labels from [Angais/Pixel Art Atlas](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m) retain their upstream **CC0 1.0** dedication. PixelExact does not impose new attribution requirements on those original materials. Upstream terms and third-party rights remain unaffected.
---
**Special thanks to [Angais](https://huggingface.co/Angais) the creator of [Pixel Art Atlas](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m)**
|