<p>Humanization is a critical process in designing antibodies and nanobodies for clinical trials. Developing widely recognized deep learning frameworks for this task remains valuable yet challenging. Here, inspired by the success of diffusion models, we introduce HuDiff, an adaptive diffusion approach for humanizing antibodies and nanobodies from scratch, referred to as HuDiff-Ab and HuDiff-Nb. This approach initiates humanization exclusively with complementarity-determining region sequences, eliminating the need for humanized templates. On public benchmarks, HuDiff-Ab generates humanized antibodies that more closely resemble experimentally humanized sequences than existing models. Similarly, HuDiff-Nb produces nanobodies with higher humanness scores and nativeness than alternative methods. We apply HuDiff to humanize a murine antibody targeting the SARS-CoV-2 receptor-binding domain and two alpaca-derived nanobodies, one targeting the receptor-binding domain and the other targeting the C345c domain of C3. Bio-layer interferometry shows the best-performing humanized antibody retains binding affinity comparable to the parental antibody (0.15 nM versus 0.12 nM). Both humanized nanobodies maintain binding to their respective antigens, with the best-performing one exhibiting a substantially enhanced affinity (2.52 nM versus 5.47 nM), corresponding to a 54% improvement over the parental nanobody. Neutralization assays confirm that the humanized sequences effectively neutralize the virus. These results demonstrate that HuDiff improves antibody and nanobody humanness while preserving or enhancing binding and function.</p>