diffusiondb

Why we included this project

DiffusionDB records what people actually typed into Stable Diffusion: around 1.8 million distinct prompts matched with 14 million generated images, along with the seeds, CFG scales, steps, and samplers used to produce each one. Because the prompt text and the generation parameters travel together, the dataset is useful for more than browsing pretty pictures. Anyone studying prompt engineering or building tools that predict and recommend prompts can mine this corpus instead of collecting their own, and the same data supports research into how textual guidance steers generative models. The Parquet metadata tables let you filter prompts, parameters, and NSFW flags without downloading the image files, and the 2M subset is small enough to prototype against on a single machine. Researchers working on detecting synthetic imagery will find the scale and the per-image NSFW labels useful for their experiments.