Quick Start

Write your prompt as a simple comma‑separated list of things you want to see in your image, then press the Run button. Prompt Squirrel will:

• highlight any unknown or misspelled tags or syntax errors,
• suggest corrected tags,
• recommend additional tags based on context, and
• list artists who produce topically similar content

You do not need to structure the prompt in any special way; just describe what you want in short phrases separated by commas.

System Overview

Prompt Squirrel uses NLP and vector‑space methods to map a free‑form prompt to the structured tag vocabulary expected by tag‑based Stable Diffusion models. Internally, we use a grammar parser, FastText embeddings, TF‑IDF and SVD for context scoring, and an approximate‑nearest‑neighbor index for artist and suggested tag retrieval. Our goal is to help users write prompts that align with the tag distributions the model was trained on.

See the Technical Details heading below for more information about how these all are used.

Prompting Guidance and Common Questions

What text to image models does this tool work for?

This instance of Prompt Squirrel works for tag-based Stable Diffusion models fine-tuned on the popular e621 dataset. The tags it returns and especially the artist names will only be recognized by models in this category, which includes popular models such as Fluffyrock and NoobAI.

Does input order matter?

No.

Should I use underscores or spaces in the input tags?

As a rule, models trained on the dataset replace underscores with spaces, so spaces are preferred

Can I use parentheses or weights as in the Stable Diffusion Automatic1111 WebUI?

Yes, but only '(' and ')' and numerical weights. These are ignored in the underlying calculations but allowed so that prompts can be copied between tools with minimal editing. An acceptable example is: ((sunset over the mountains)), (clear sky:1.5), ((eagle flying high:2.0)), river, (fish swimming in the river:1.2), (campfire, (marshmallows:2.1):1.3), stars in the sky, ((full moon:1.8)), (wolf howling:1.7)

Why are some valid tags marked as 'unknown', and why don't some artists ever get returned?

Some tags or artists are too infrequent in the dataset sample for us to make reliable predictions. Prompt Squirrel merges data from several sources, which may differ slightly in tag names or counts. Low‑frequency items or inconsistent entries may therefore not appear in results.

Why do some suggested tags not have summaries or wiki links, and why do some look truncated?

Summaries and wiki links are extracted from dataset wiki pages. Some tags do not have pages, and summaries are heuristically extracted from the page beginnings, which can introduce small errors.

Are there any special tags?

Yes. We normalized favorite counts to a range of 0–9. You may include: 'score:0' through 'score:9' These bias the output toward suggestions associated with images with higher or lower scores.

Are there any other special tricks?

Yes. Repeating a tag gives it more weight in the artist‑similarity calculation. For example: 'red fox, red fox, red fox, score:7' will bias more strongly toward artists and suggested tags associated with 'red fox' than: 'red fox, score:7'.

Why is this space tagged 'not‑for‑all‑audience'?

The dataset used by many tag‑based models contains both general‑audience and adult material. To avoid surprising users, mature tags are hidden unless the user explicitly enables them. This tool processes only text and metadata; no images from the dataset are displayed.

Technical Details

How is the artist list calculated?

Each artist is represented as a pseudo‑document containing the bag of all tags from their images. Your prompt is treated as another pseudo‑document. We compute similarity between the recognized tags in your prompt and each artist using TF‑IDF and truncated SVD, then retrieve the nearest artists using an approximate‑nearest‑neighbor index.

How do the Suggested Tags work?

Each tag is represented as a pseudo-document containing the bag of all tags it co-occurs with in the dataset. We then employ exactly the same method on them as we did with artists to suggest tags similar to your prompt.

How does the tag corrector work?

We treat each image as a document containing the set of its tags and randomly replace about 10% of the tags with aliases from the dataset's alias lists. We then train a FastText model on these documents so that tags and their aliases map to nearby vectors. This makes the system robust to spelling variations and rephrasings.

To incorporate the prompt's context, we again treat tags as pseudo-documents containing the bag of all tags they co-occur with, then compute TF‑IDF scores for the top candidate tags selected from the FastText Model, combining this similarity score with the FastText similarity. The Context Similarity Weight slider controls how much influence the TF‑IDF context score has relative to the FastText embedding similarity.

How do the sample images work?

In each gallery row, we choose an illustrative prompt and generate one image for each artist using the popular model Fluffyrock Unleashed, which was trained on this dataset. The 'No Artist' image serves as a baseline, using the same prompt without an artist name. Each subsequent row repeats this process with a different prompt. The first prompt was chosen to illustrate foreground style, the second to illustrate background style, and the third to illustrate character design. See SamplePrompts.csv in the Files section for the list of prompts used.