AMP-Diffusion is a latent space diffusion model designed for antimicrobial peptide (AMP) generation, leveraging the state-of-the-art protein language model (pLM), ESM-2. The model generates novel AMPs by introducing Gaussian noise during the forward process and reversing it to reconstruct peptide embeddings. AMP-Diffusion demonstrates strong performance in generating peptides with low perplexity, high diversity, and similarity to experimentally validated AMPs. The generated peptides exhibit physicochemical properties similar to natural AMPs, including charge, hydrophobicity, aromaticity, and isoelectric point (pI). The model also shows high Jaccard similarity with training data at the 6-mer level, indicating its ability to capture complex sequence motifs. External validation using the HydrAMP classifier confirms the antimicrobial potential of the generated peptides, with a high proportion exceeding thresholds for antimicrobial activity. The model's generated peptides align closely with natural AMPs in terms of physicochemical properties, suggesting biological plausibility. AMP-Diffusion represents a promising advancement in protein design, combining the strengths of pLMs with diffusion models. Future research will focus on experimental validation, further exploration of the model's ability to generate peptides with specific properties, and broader applications in protein engineering. The model's framework offers a versatile platform for peptide and protein design, with potential for optimization through parameter tuning and additional conditioning data.AMP-Diffusion is a latent space diffusion model designed for antimicrobial peptide (AMP) generation, leveraging the state-of-the-art protein language model (pLM), ESM-2. The model generates novel AMPs by introducing Gaussian noise during the forward process and reversing it to reconstruct peptide embeddings. AMP-Diffusion demonstrates strong performance in generating peptides with low perplexity, high diversity, and similarity to experimentally validated AMPs. The generated peptides exhibit physicochemical properties similar to natural AMPs, including charge, hydrophobicity, aromaticity, and isoelectric point (pI). The model also shows high Jaccard similarity with training data at the 6-mer level, indicating its ability to capture complex sequence motifs. External validation using the HydrAMP classifier confirms the antimicrobial potential of the generated peptides, with a high proportion exceeding thresholds for antimicrobial activity. The model's generated peptides align closely with natural AMPs in terms of physicochemical properties, suggesting biological plausibility. AMP-Diffusion represents a promising advancement in protein design, combining the strengths of pLMs with diffusion models. Future research will focus on experimental validation, further exploration of the model's ability to generate peptides with specific properties, and broader applications in protein engineering. The model's framework offers a versatile platform for peptide and protein design, with potential for optimization through parameter tuning and additional conditioning data.