• DALL-E 2 implements pre-training mitigations to reduce risks associated with powerful image generation models, including filtering out violent and sexual content.
  • The filtering process can amplify biases in training data, so techniques are used to mitigate this effect and prevent the model from being more biased than the unfiltered version.