CVPR Daily - Thursday

13 DAILY CVPR Thursday assuming all these vectors are two- dimensional , you can say we have this direction of the unsafe concept, and we draw a circle around there, which is the area we don’t want to enter. Then we try to push the generation to avoid this area in the latent space. ” Patrick adds: “ We want to go in this safe direction and compute the error, the gradient between the unsafe point. Then we move along this gradient. ” Patrick says certain concepts proved difficult to circumvent. For instance, in the case of Asian women, the model demonstrated a stubborn bias toward generating nudity. “ We could only completely avoid nudity if we moved strongly away from the text prompt, ” he tells us. “ In the first instance, you would align this to find those hyperparameters where you stay close to the original image. ” Could adding the word ‘dressed’ to the prompt somewhat alleviate the problem? The paper compares a baseline approach known as negative prompting , commonly employed in text-to-image models, involving specifying concepts to be avoided in the prompt. However, negative prompting exhibits drawbacks Figure 3. Illustration of text-conditioned diffusion processes. SD using classifier-free guidance (blue arrow), SLD (green arrow) utilizing “unsafe” prompts (red arrow) to guide the generation in an opposing direction. Safe Latent Diffusion ...