README.md

---
license: cc0-1.0
language:
- en
tags:
- art
- painting
- hammershoi
- danish
- lora-training
- flux
- diffusion
- text-to-image
task_categories:
- text-to-image
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.jsonl
---

# Dataset Card for Hammershoi Training Dataset

A curated dataset of **244 image-caption pairs** based on the works of Danish painter **Vilhelm Hammershøi** (1864–1916), created for fine-tuning text-to-image diffusion models. Used to train the [Hammershoi Flux1 LoRA](https://huggingface.co/jejunepixels/Hammershoi-Flux1-LoRA).

## Dataset Details

### Dataset Description

A structured image-caption dataset of cropped and framed reproductions from Hammershøi's paintings, annotated with detailed captions designed for LoRA training. Each image is paired with a structured caption describing category, framing, lighting, subject, and style tags. The dataset covers portraits, interiors, landscapes, and exterior architectural details from Hammershøi's body of work.

- **Curated by:** [jejunepixels](https://huggingface.co/jejunepixels)
- **Language(s):** English
- **License:** CC0 1.0

### Dataset Sources

- **Model trained on this dataset:** [jejunepixels/Hammershoi-Flux1-LoRA](https://huggingface.co/jejunepixels/Hammershoi-Flux1-LoRA)
- **Demo:** [jejunepixels/Hammershoi-Flux1-Space](https://huggingface.co/spaces/jejunepixels/Hammershoi-Flux1-Space)

## Uses

### Direct Use

This dataset is intended for fine-tuning text-to-image diffusion models (such as FLUX.1-dev) using LoRA or similar techniques. It can be used to train models that generate images in the style of Vilhelm Hammershøi — characterised by muted palettes, quiet domestic interiors, and introspective figures.

### Out-of-Scope Use

This dataset is not suitable for training face recognition systems or models intended to generate photorealistic images of real people. The images are paintings, not photographs.

## Dataset Structure

Each record in `metadata.jsonl` contains:

| Field | Description |
|---|---|
| `file_name` | Image filename (e.g. `portrait001.png`) |
| `caption` | Full training caption starting with trigger word `hmrsh` |
| `category` | `portrait painting`, `interior painting`, or `landscape painting` |
| `lighting` | Lighting condition used in caption |
| `composition` | Framing used in caption |
| `width` | Image width in pixels |
| `height` | Image height in pixels |
| `source` | Source institution |
| `source_url` | Original image URL |

**Caption format:**
```
hmrsh, [category], [framing], [lighting], [subject description], oil on canvas, danish symbolism, muted palette
```

**Resolution buckets used in training:**

| Resolution | Count | Orientation |
|---|---|---|
| 896×1152 | 59 | Portrait |
| 1024×1024 | 56 | Square |
| 768×1344 | 25 | Tall portrait |
| 1344×768 | 5 | Wide/cinematic |

## Dataset Creation

### Curation Rationale

This dataset was created to train a style LoRA capable of generating images in the distinctive aesthetic of Vilhelm Hammershøi. The structured caption format — with explicit fields for category, framing, lighting, and subject — was designed to give users fine-grained control over generated outputs at inference time.

### Source Data

#### Data Collection and Processing

Images were sourced from open-access museum collections as high-resolution reproductions of Hammershøi's original paintings. Each painting was cropped into multiple image tiles at different framings (wide, medium, close up, extreme close up) to maximise the variety of compositional examples. Captions were written manually for each crop, following a consistent structured format.

#### Who are the source data producers?

The original artworks were created by **Vilhelm Hammershøi** (1864–1916). Digital reproductions were provided by open-access museum collections:

| Institution | URL |
|---|---|
| SMK Open — Statens Museum for Kunst | [open.smk.dk](https://open.smk.dk) |
| Wikimedia Commons | [commons.wikimedia.org](https://commons.wikimedia.org) |
| Nationalmuseum (Sweden) | [nationalmuseum.se](https://www.nationalmuseum.se) |
| The Metropolitan Museum of Art | [metmuseum.org](https://www.metmuseum.org) |

### Annotations

#### Annotation process

Captions were written manually for each image crop following a fixed schema: `hmrsh, [category], [framing], [lighting], [subject description], oil on canvas, danish symbolism, muted palette`. Category, framing, and lighting values were chosen from a controlled vocabulary to ensure consistency across the dataset.

#### Who are the annotators?

Captions were written by [jejunepixels](https://huggingface.co/jejunepixels).

#### Personal and Sensitive Information

This dataset contains no personal or sensitive information. All subjects depicted are figures in 19th-century paintings.

## Bias, Risks, and Limitations

The dataset reflects the subject matter of Hammershøi's existing body of work, which predominantly depicts white European women in domestic interior settings. Models trained on this dataset will reflect this bias and may not generalise well to diverse subjects or settings outside this aesthetic. The dataset covers a single artist's style and is not intended as a general-purpose art dataset.

### Recommendations

Users training on this dataset should be aware that outputs will reflect Hammershøi's characteristic aesthetic: muted greys, quiet interiors, and solitary figures. For broader stylistic range, supplement with additional training data.

## Dataset Card Authors

[jejunepixels](https://huggingface.co/jejunepixels)

## Dataset Card Contact

Contact via the [Hugging Face community tab](https://huggingface.co/datasets/jejunepixels/Hammershoi-Training-Dataset/discussions).