DiffusionDB
Polo Club of Data Science (Georgia Tech)
2 million Stable Diffusion images paired with the prompts and settings real users typed.
The first large-scale text-to-image prompt dataset, collected from the Stable Diffusion Discord. One row is one generated image with its prompt, seed, and sampler parameters. It supports study of prompt writing, human-AI interaction, and deepfake detection.
Access
- Cost
- FreeNo charge; large image archives.
- License
- Public domainCC0 1.0
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadParquet, PNG, WebP
What's in it
- One row is
- One row per generated image with its prompt and parameters
- Covers
- Global Discord users
- Years
- 2022
- Updated
- One-time release
- Size
- 2 million images in the 2M subset (a larger set exists)
Quality
Well documented and widely used for prompt research; it is a snapshot from 2022 of one model and one community.
- Contains not-safe-for-work images and offensive prompts
- Reflects only the first Stable Diffusion versions
- Prompts may include personal information
Connects to
- Stable Diffusion Prompts: Smaller prompt-only set for comparison
- LAION-5B: Real image-caption corpus used to train the model whose outputs this documents
Questions it could answer
- Which prompt words most often appear together?
- How do users change parameters across repeated tries?