Datasets:

jejunepixels
/

Hammershoi-Ideogram4-Dataset

license: other
language:
  - en
tags:
  - art
  - painting
  - hammershoi
  - lora-training
  - ideogram-4
task_categories:
  - text-to-image
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - train/*.png
          - train/metadata.jsonl
  - config_name: boxes
    data_files:
      - split: train
        path:
          - boxes/*.jpg
          - boxes/metadata.jsonl

Hammershøi Ideogram 4 Dataset

288 crops of paintings by Vilhelm Hammershøi (1864–1916), captioned for training an Ideogram 4 LoRA.

Same images as Hammershoi-Training-Dataset minus one rejected crop, with new captions in Ideogram 4's structured JSON format.

Files

train/ holds <name>.png, <name>.txt (the JSON caption, for AI-Toolkit and similar trainers) and metadata.jsonl (the same caption plus source and size, for the dataset viewer).

boxes/ holds copies of the images with their bounding boxes drawn on. Don't train on these.

See the bounding boxes

Open the boxes subset in the viewer to see every image with its boxes drawn on. Each box has a numbered tag, and the boxes column says what each number is. The without_box column lists elements that have a description and no box.

A woman at a window, with numbered bounding boxes on the figure, window and bed A bust, a chest of drawers and a chair, each with a numbered bounding box A room with a stove, a door and a wooden floor, each with a numbered bounding box

Captions

Compact JSON following the Ideogram 4 caption schema. high_level_description starts with the trigger hmrsh painting: . Bounding boxes are [y_min, x_min, y_max, x_max] on a 0–1000 scale.

Captions and boxes were drafted by Claude from the images and earlier captions, then reviewed and corrected by hand. Boxes are close estimates, not pixel-exact. Some elements have a description and no box, where the reviewer removed the box. Colour palettes are extracted from the pixels.

Sizes

Size	Images
896×1152	91
1024×1024	77
768×1344	56
1152×896	32
1344×768	30
996×1152	2

Sources

Each row of metadata.jsonl names its source. They include museum collections and auction-house images. The paintings are in the public domain; check the source of a reproduction before reusing it.