Comparing local large language models for alt-text generation