---
license: other
task_categories:
  - image-to-text
  - text-to-image
tags:
  - art
  - chinese-painting
  - museum
  - caption
  - ocr
  - parquet
size_categories:
  - 10K<n<100K
---

# Chinese Painting Collection

91,093 images of Chinese paintings with VLM captions (a `captions` list of one
to three texts per image, Chinese and English, short and long), calligraphy OCR
transcriptions, view-type classification, and foreign-object detection with a
final crop box per image. The `image` column is already cropped to the artwork;
`image_original` keeps the untouched photograph.

## Sources

| Component | Source | Images | Image license |
|---|---|---|---|
| `npm_tw_c0`–`npm_tw_c3` | [National Palace Museum (Taipei) Open Data](https://theme.npm.edu.tw/opendata/) | 86,321 | Taiwan Open Government Data License v1 (attribution required) |
| `met_china` | [The Metropolitan Museum of Art Open Access](https://www.metmuseum.org/art/collection) | 1,186 | CC0 |
| `fsg` | [Smithsonian Freer Gallery of Art / National Museum of Asian Art](https://asia.si.edu/) | 1,782 | CC0 |
| `princeton` | [Princeton University Art Museum](https://artmuseum.princeton.edu/) | 1,631 | Check source terms; intended for research use |
| `aic_china` | [Art Institute of Chicago](https://www.artic.edu/collection) | 173 | CC0 |

The four NPM components are quarterly shards of the NPM Open Data release
(`c0`–`c3`), deduplicated and filtered to usable views. The museum components
were harvested through each institution's public API and filtered to
China-related Chinese-painting objects.

## Production pipeline

1. **Harvest** — images and tombstone metadata (title, artist, date, object
   number) fetched from the source APIs / open-data packages.
2. **Rule-based cleaning** — removal of non-artwork views: rolled scrolls,
   installation shots, covers, damaged or unreadable images, and near-duplicates.
   Only `full`, `mounted`, and `detail` views are kept.
3. **VLM labeling** (gemini-3.5-flash-lite) — for every kept image:
   - `caption_zh` / `caption_en`: free-form descriptions of the scene,
     composition, style, and medium, instructed to avoid boilerplate openings.
   - `ocr_text`: transcription of visible Chinese calligraphy (inscriptions,
     colophons) and a description of seal impressions where present.
   - `view_type`: `full` (whole artwork), `mounted` (artwork with mounting
     visible), or `detail` (close-up crop).
   - `culture`: `chinese` (88,586), `japanese` (2,253), `korean` (82),
     `western` (10), `other` (20), `uncertain` (142). Non-Chinese items are a
     small minority kept for completeness; filter on this field if unwanted.
4. **Foreign-object detection** — museum photography of NPM flatworks often
   includes color charts, rulers, accession labels, or the surrounding desk.
   A detection pass lists what it sees (`artifacts`: `label`, `color_chart`,
   `ruler`, `desk`, `glare`, `other`; non-empty for 39,164 images) and returns a
   crop box for the artwork.
5. **Second cropping round** — for the images the first pass flagged, the model
   was shown the already-cropped view and asked for the artwork box *inside* it.
   The two boxes are composed into one box in the coordinates of the original
   photograph.
6. **Third cropping round** (DeepSeek V4.1 Flash, grounding-verified before
   scaling) — every row still flagged with `color_chart` was re-detected, and
   the new box was validated against the crop it produces. Boxes that changed
   invalidated the captions written on the old crop, so those rows were
   re-captioned from the new crop (see step 8). All rows with any artifact had
   their short zh/en captions (written on the uncropped or first-crop view)
   discarded.
7. **Text folded into the captions** — the transcription from step 3 was moved
   into the captions as a trailing block, because the caption is what a
   text-to-image model consumes. Catalogue-label lines (accession numbers,
   album titles) were dropped in the process, since the crop removes that strip
   from the training view. `ocr_text` is kept as the raw field.
8. **Long captions and re-captions** (DeepSeek V4.1 Flash, reasoning disabled)
   — 59,850 images received a much longer caption written on the *cropped* view
   (what a trainer sees), and the 38,741 rows whose crop changed in round 6 were
   re-captioned the same way. 616 candidates were discarded for repeating a
   sentence; length alone was never a reason to discard.
9. **Caption aggregation** — all surviving captions of a row (short zh, short
   en, long) are shipped as one `captions: list[str]`; the separate
   `caption_zh` / `caption_en` / `caption_long` columns no longer exist. Rows
   carry 1 caption (39,099), 2 (31,182), or 3 (20,747); 65 rows have none left
   after the round-6 discards and are kept for their metadata. Caption lengths
   in characters: median 825, p90 3,181, max 12,992.

## Data format

Parquet shards under `default/train/part-*.parquet` (374 shards, zstd), one row
per image, plus a small `metadata/metadata.parquet` with everything except the
image bytes. Both `image` and `image_original` use the
`struct<bytes: binary, path: string>` layout of the `datasets` `Image` feature,
so every image can be previewed in the dataset viewer:

```python
from datasets import load_dataset
ds = load_dataset("kaupane/chinese-painting-collection", split="train")
ds[0]["image"].size          # the cropped artwork view
ds[0]["image_original"].size # the untouched photograph
ds[0]["captions"]            # list of 1-3 captions (zh/en, short and long)
```

Columns: `image` (cropped), `image_original`, `captions` (list of strings),
`bbox` (list of four floats), `artifacts` (list of strings), `image_id`,
`shard`, `source`, `object_no`, `title`, `artist`, `category`, `culture`,
`view_type`, `ocr_text`.

`bbox` is a normalised `[x_min, y_min, x_max, y_max]` in per-mille coordinates
(0–1000) **of the original photograph**, and `image` is exactly that crop
applied (JPEG quality 95). `bbox` is present for 39,591 rows; an empty `bbox`
means the whole frame is the artwork. `previews/` shows twelve before/after
pairs so the effect of the crop can be seen without loading the data.

View-type distribution: `mounted` 43,306 · `full` 36,150 · `detail` 11,637.

## Suggested filtering

- **Training on artwork only**: use the `image` column directly — the crop is
  already applied. `image_original` + `bbox` are there if you want a different
  crop policy.
- **Cleanest subset**: drop rows whose `artifacts` is non-empty.
- **Exclude mounting borders**: keep `view_type == "full"` or `"detail"`.
- **Chinese paintings only**: keep `culture == "chinese"`.
- **Long-text training**: longer entries in `captions` are the long captions;
  short zh/en captions survive only on rows with no detected artifacts.

## Caveats

- Captions and OCR are machine-generated by lightweight VLMs; expect occasional
  errors, especially in cursive-script calligraphy transcription. Where a
  transcription is quoted in a caption, it is the model's reading, not a
  scholarly edition.
- The crop box is the model's answer to "where is the artwork". It usually
  removes the color chart, ruler and label strip, and in most cases leaves a
  thin strip of the photograph's background along one edge.
- `detail` views are close-ups of larger works and may duplicate content present
  in `full` views of the same object (same `object_no`).
- 23 rows have flagged artifacts but no reliable box; they are shipped
  uncropped. 65 rows have no captions after round-6 discards.

## Licensing

Images retain the licenses of their sources (see table above). The NPM portion
requires attribution to 國立故宮博物院 (National Palace Museum, Taipei) under
the Taiwan Open Government Data License v1. Annotations (captions, OCR,
classifications, boxes) are released under CC-BY-4.0.