dailenson/One-DM

Leverages a diffusion-based architecture with a style-enhanced module that extracts high-frequency components from reference samples to capture fine-grained handwriting patterns while suppressing noise. Built on PyTorch 1.13, it integrates a pre-trained VAE-based OCR model and ResNet18 feature extractor for style encoding, supporting multilingual generation across English, Chinese, and Japanese scripts. Achieves competitive results with minimal references through distributed training via torchrun, with pretrained checkpoints available across multiple cloud platforms.

539 stars.