license: cc-by-nc-sa-4.0
task_categories:

• text-to-image
  language:
• en
  tags:
• erotic
• prompt
• caption
• LoRA
• benchmark
• test
  pretty_name: Erotic Confident Woman Prompts
  size_categories:
• 10K<n<100K
  configs:
• config_name: default
  data_files:
  • split: train
    path: captions.parquet
    features:
  • name: Caption
    dtype: string
  • name: Aspect Ratio
    dtype: string
  • name: Nudity
    dtype: string
  • name: Artistic Quality
    dtype: float32
  • name: Erotic Score
    dtype: float32
  • name: Pornographic Score
    dtype: float32
  • name: Males
    dtype: int32
  • name: Females
    dtype: int32
  • name: Clothing
    dtype:
    sequence: string
  • name: Location
    dtype: string

Dataset Card for Erotic Confident Woman Prompts

Dataset Description

Dataset Summary

Large language models (LLMs) are surprisingly bad at creatively inventing new things, even though they are masters of hallucination. Asking an LLM — even an abliterated one — to produce a list of random erotic prompts therefore yields a rather boring, narrow-minded result. This is easy to overcome: give the model some inspiration and let it do what it does best — transform that inspiration into prompts in the style you need.

This dataset is a focused companion to Erotic Image Prompts. Instead of a broad, mainstream collection, it is built from a single narrow inspiration source: photographs of confident women presenting themselves. Each photo was described — never copied — by a multimodal LLM and rewritten into a long, detailed text-to-image prompt in the style of the krea-2 prompting guidelines.

The woman in every prompt is called Alice, the trigger word you can search-and-replace. Everything about the person (hair, race, skin, eyes, body, age, tattoos, …) is deliberately left out, because the trigger word is meant to carry her appearance; everything about the presentation (setting, wardrobe, fabrics, jewellery, pose, expression-as-mood, lighting, camera, palette, medium) is described in depth, per the krea-2 advice that long detailed natural-language prompts yield the best results.

To make it easy to reduce the collection to a workable subset, each prompt is accompanied by metadata and an aspect ratio.

Languages

All prompts are written in English.

Dataset Structure

Data Fields

Each entry consists of a text prompt plus the following metadata fields:

The same data is provided twice: captions.parquet is the data file used by the dataset viewer and by the datasets library (Clothing is a real list of tags); captions.csv is a plain-file twin with identical rows and columns, where Clothing is written as a tag string list, e.g. ['dress', 'heels'] ([] when no garment is mentioned). The CSV is meant for tools that read files instead of using a library; its Clothing cell is a Python list literal, so ast.literal_eval() turns it back into a list.

Note on the score columns: Artistic Quality, Erotic Score, and Pornographic Score were generated by an LLM in a zero-shot fashion and independently of every other prompt in the set. Expect heavy hallucination in these three values and never use them for training purposes. Their reliability is expected to be very poor; they are provided only as a service to very quickly reduce the number of prompts for further processing.

Naming Convention

Every person in a prompt is given a placeholder name, in order of appearance:

• Women: Alice, Bella, Clara, Diana, Emma
• Men: Adam, Bob, Charlie, David, Eric

To test a character LoRA, simply replace Alice (or Adam) with the trigger word for your character.

Data Splits

The dataset is provided as a single set, with no predefined train/validation/test split.

Dataset Creation

The prompts were created by having a multimodal LLM describe photographs of confident women presenting themselves and rewrite that description into an erotic-leaning text-to-image prompt in the required style. The photos themselves are not part of this dataset and are not redistributable; only the resulting text prompts are published, and the source URL and image file name were deliberately dropped before publishing.

Considerations for Using the Data

Discussion of Biases

Erotic is a highly subjective matter, so expect strong bias. The prompts also inherit the bias of the source photography — shots of confident women presenting themselves are skewed towards a particular styling, wardrobe and location range. To mitigate this, the sample size is large, so with some filtering you should be able to obtain a subset that matches the bias you have in mind more closely. Even then, the underlying content reflects what people generally consider "vanilla", so special interests are barely covered — or not included at all.

Known Limitations

The Artistic Quality, Erotic Score, and Pornographic Score fields are LLM-generated and unreliable (see the note above). Do not treat them as ground truth.

Clothing tags are normalised to a small controlled vocabulary: spelling and synonym variants are merged (e.g. bracelets → bracelet, garter / garters / garter straps → garter belt, crop top / tank top / lace top → top). Tags that are not garments — fabrics such as lace, props such as towel — are left as the model wrote them. Aspect Ratio is a snapped hint, not the exact ratio of the original photo.

Safety

As erotic is a sensitive topic, take care with regard to safety. Never create images depicting real persons who have not given clear consent, and only use these prompts to depict persons who, in reality and visually, meet the legally required minimum age.

Usage Restrictions

This dataset is released for research and non-commercial use only. By downloading or using it you agree to all of the following.

You must not use this dataset — neither the prompts as they are, nor any prompt derived from them (this includes prompts produced by feeding them to an LLM or any other model) — to:

• generate images that are sold or otherwise monetised;
• generate images that promote anything that is sold (advertising, product or brand promotion, sponsorship, affiliate or marketing content, …);
• create or run a virtual influencer — a fictional, computer-generated persona used to present, front or stand in for a person — in any way, whether or not that use is commercial.

If you are unsure whether an intended use falls under these restrictions, treat it as if it does and ask before using the dataset.

How to Use

You can use any tool that accepts text prompts. For convenience with ComfyUI, the Hugging Face dataset and the Basic Data Handling custom nodes can feed these prompts into the official Krea 2 Turbo workflow example with this provided workflow.

Use Cases

These prompts are ideal for testing the capabilities of a text-to-image model. They can be used as-is, or with the placeholder name(s) (Alice, Adam, …) replaced by trigger words to test character adapters — including testing character LoRAs — as long as the Safety and Usage Restrictions above are respected.

To adapt the prompt style to the convention of a specific text-to-image model or do some substantial changes like
setting an image style it is recommended to use an LLM to translate the relevant prompts of this dataset to the
requirements of the intended use case.

Statistics

13070 prompts in captions.csv / captions.parquet.

Count of Persons

HTML version (same shape as the card; rows = Males, columns = Females):

Aspect Ratio

Nudity

Artistic Quality (0.1 bins)

Erotic Score (0.1 bins)

Pornographic Score (0.1 bins)

Garments (Clothing)

Location

Additional Information

Licensing Information

This dataset is released under the CC BY-NC-SA 4.0 licence (Attribution — NonCommercial — ShareAlike). This is deliberately not the plain CC BY 4.0: the “NC” element restricts every use to non-commercial purposes, so selling images generated from these prompts — or using them to promote anything that is sold — is outside the licence, and the “SA” element means that anything you share which builds on the dataset (an edited prompt set, a derived dataset, …) must be released under the same licence. The Usage Restrictions above add one further condition: a blanket ban on virtual-influencer use, commercial or not.

Citation Information

If you use this dataset, please cite it as follows:

@misc{erotic-confident-woman-prompts,
  author       = {StableLlama},
  title        = {Erotic Confident Woman Prompts},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/stablellama/erotic-confident-woman-prompts}}
}