Pruning does not reduce capacity, pruning removes keys from the model that are NEVER used for generation. Converting precission from fp32 to fp16 reduces the "accuracy", but most people run Stable Diffusion in fp16 mode, so it's irrelevant if the model precission is one or the other