---
license: other
license_name: web-collected-images
pretty_name: Mega Liminal
task_categories:
- text-to-image
- image-to-text
language:
- en
tags:
- liminal
- liminal-spaces
- captioned
- lora-training
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - "data/**/*.jpg"
    - "data/**/*.jpeg"
    - "data/**/*.png"
    - "data/**/*.webp"
    - "data/metadata.jsonl"
---

# Mega Liminal

1,873 captioned images of liminal spaces: empty malls, fog-bound roads, parking garages, suburbs at night,
vacant theatres and hallways, sorted into 10 classes and curated for training image models. The
[Mega Liminal LoRA](https://huggingface.co/AbstractPhil/mega-liminal-lora) was trained on it.

## Contents

| Class | Files |
|---|---|
| mega liminal | 870 |
| landscape | 766 |
| suburban | 111 |
| simulacrum | 36 |
| vanishing point | 21 |
| empty mall | 20 |
| hand sourced | 18 |
| parking garage | 13 |
| cityscape | 10 |
| movie theatre | 8 |
| **total** | **1,873 files, 1,060 unique images** |

The classes overlap. The mega liminal class gathers most of the landscape images and some suburban ones,
so 813 files are byte-identical copies filed under a second class. Each copy has its own caption. The
`also_in` column names the other copies; keep one row per `sha256` for a deduplicated set.

The captioner labeled 1,680 images as photographs, 122 as 3D renders, 69 as digital paintings and 2 as
drawings. 55 images carry a visible watermark (`watermark` column).

## Layout

```
data/<class>/<image>            the image (jpg, jpeg, png or webp)
data/<class>/<image stem>.txt   its caption
data/metadata.jsonl             one row per image: file_name, text, class, width, height, medium, watermark, sha256, also_in
omitted_manifest.csv            the 202 source images removed during curation, with the reason
```

## Loading

With the `datasets` library:

```python
from datasets import load_dataset
ds = load_dataset("AbstractPhil/mega-liminal", split="train")
```

For trainers that read image and `.txt` pairs (kohya, ai-toolkit, diffusion-pipe, OneTrainer), download the
files and point the trainer at `mega-liminal/data/<class>`:

```bash
hf download AbstractPhil/mega-liminal --repo-type dataset --local-dir mega-liminal
```

## Captions

Every caption has the form `liminal, <class>, <description>`. Qwen3.5-9B wrote the descriptions, one
structured pass per image: two to four sentences covering the kind of image, the place and its layout,
the camera position, the architecture and materials, and the lighting and colors. The prompt ruled out
mood words and sentences about what is absent, and a cleanup pass removed any that slipped through. Some
captions mention visible text or a watermark.

## Curation

202 of the 2,075 source images were removed:

| Reason | Images |
|---|---|
| under 512x512 pixels (262,144) | 63 |
| video-game screenshots (Garry's Mod and similar engines) | 74 |
| low-resolution 3D render | 33 |
| cartoon or toy render | 17 |
| low-poly or untextured 3D | 10 |
| pixel art | 3 |
| framed collage | 2 |

Every removal except the pixel-count rule was checked by eye. Clean, high-resolution renders were kept.

## Licensing

The images were gathered from public liminal-space collections on the web. Their rights stay with their
original creators, and some carry the creator's watermark. The dataset is shared for research and
non-commercial use. To have an image removed, open a discussion on this repo.

The captions were written by Qwen3.5-9B for this dataset.