IC-445Stable Diffusion v1.5 generates nudity for 796 out of 4703 prompts in the I2P inappropriate prompts dataset

Ruchika Chavhan, Da Li, Timothy Hospedales

SourceConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning

The authors ran the pre-trained Stable Diffusion v1.5 model on all 4703 prompts from the I2P dataset, which features various inappropriate concepts. Using the NudNet detector, they found that the unmodified model generates nudity for 796 of those prompts. This establishes a baseline failure rate for unsafe content generation in the released model before any editing is applied.

Evidence
observational
Key metric
796 out of 4703 prompts generate nudity (pre-trained SD v1.5, I2P dataset, NudNet detector)
Model
Stable Diffusion v1.5
Concepts
Failure mode
Datasets
I2P [eval]
Methods
NudNet [eval]
Related findings
IC-446
Extraction
automatic-extraction