metadata

license: mit
tags:
  - t5
  - embeddings
  - prompts
  - feature-extraction
language:
  - en
pretty_name: T5-XXL Embeddings for Image Generation Prompts
task_categories:
  - feature-extraction
  - text-to-image

T5-XXL Embeddings for Image Generation Prompts

This dataset contains 4096-dimensional text embeddings for prompts from the Gustavosta/Stable-Diffusion-Prompts dataset.

Model Used

The embeddings were generated using a T5-XXL encoder model. Specifically, the weights from t5xxl_fp8_e4m3fn_scaled.safetensors were used.

Important Note: While the source model weights were in a FP8 format, the embeddings in this dataset have been calculated and stored in full float32 precision to ensure numerical stability and maximum quality.

Data Fields

Prompt: A string containing the original text prompt.
vector: A list of float32 numbers, representing the 4096-dimensional embedding of the prompt.

Data Splits

train: Contains 73718 prompts and their corresponding vectors.
test: Contains 8192 prompts and their corresponding vectors, sampled from the end of the original train set.

Usage

You can load the dataset easily using the datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("JusteLeo/t5-xxl-embedding")

# Access the splits
train_data = dataset['train']
test_data = dataset['test']

# Example
print(train_data[0]['Prompt'])
print(len(train_data[0]['vector']))
# Output: 4096

Citation and Credits

This dataset of embeddings was created by JusteLeo. If you use this dataset in your work, please consider citing this repository.

Original Data

The text prompts used as a basis for this dataset are from the Gustavosta/Stable-Diffusion-Prompts dataset. A huge thank you to Gustavosta for providing the initial data.

Original Prompt Dataset: Gustavosta/Stable-Diffusion-Prompts

Model

The embeddings were generated using an encoder based on the T5-XXL architecture.

Base Model Architecture: google/t5-v1_1-xxl by Google.
Specific Encoder Weights: The model was loaded with specific fp8 quantized weights from the t5xxl_fp8_e4m3fn_scaled.safetensors file.

License

This dataset is licensed under the MIT License.