Skip to content

Dataset · SwayStar123

SDXL Latents, SigLIP Embeddings, and Captions for CommonCatalog

SwayStar123/preprocessed_commoncatalog-cc-by

Precomputed SDXL VAE latents, SigLIP embeddings, and moondream2 captions based on the CommonCatalog CC-BY dataset.

What's inside

The set contains 10 to 100 million records in English, prepared based on the open commoncatalog-cc-by corpus. Inside are text descriptions from the moondream2 model (also placed in a separate prompts.json file), latent representations from SDXL VAE in f32 format, and SigLIP embeddings in f16 bytes.

What you can do with it

  • Accelerated training of diffusion models without wasting resources on running through VAE and text encoder
  • Fine-tuning or LoRA training for SDXL-level image generators
  • Training multimodal text-image retrieval and matching models
Languages
English
Size
10M–100M rows
Downloads this month
98.4K
Likes
18

How to load it

from datasets import load_dataset

ds = load_dataset("SwayStar123/preprocessed_commoncatalog-cc-by")