I know a long standing quirk in user experience across all implementations of Stable Diffusion is that reducing attention weights for a token is not as effective as increasing them (ie 0 weight can ...