terminusresearch/ideogram-75k

---
license: agpl-3.0
---

# Ideogram-75k

## Dataset Details

This dataset is not authorised by, curated by, or related to Ideogram.

#### This dataset contains the `ideogram-25k` dataset contents. Do not use both!

### Dataset Description

- **Curated by:** @pseudoterminalx
- **License:** AGPLv3.

**Note**: All models created using this dataset are a derivative of it, and must have an open release under a permissible or copyleft license.

### Dataset Sources

Pulled ~75,000 images from Ideogram, a proprietary image generation service that excels at typography.

## Uses

- Fine-tuning or training text-to-image models and classifiers
- Analysis of Ideogram user bias

## Dataset Structure

- Filenames are an SHA256 hash of the image data, and can be used to verify the integrity.
- The `caption` column was obtained by asking Microsoft Florence2 (ft) to accurately describe what it sees.

## Dataset Creation

### Curation Rationale

Ideogram's users focus on typography generations, which makes it a suitable source for a lot of high quality typography data.

As a synthetic data source, its outputs are free of copyright concerns.

#### Data Collection and Processing

Used a custom Selenium application in Python that monitors the Ideogram service for posts and immediately saves them to disk.

Data is deduplicated by its SHA256 hash.

## Bias, Risks, and Limitations

As the captions all currently come from a single synthetic source, the bias of the Llava 34B captioner is present throughout this dataset.

More captions will be added.

## Citation

If there is any model built using this dataset or any further augmentations (eg. new captions) are added, this page & Terminus Research should be cited.