yasserben/DATUM

Leverages Stable Diffusion with DreamBooth personalization to generate photo-realistic synthetic target datasets from a single unlabeled image, enabling semantic-guided augmentation while preserving spatial context. Integrates with mmcv and PyTorch, using DAFormer/HRDA segmentation models (MiT or ResNet encoders) to adapt across synthetic-to-real benchmarks like GTA→Cityscapes. Achieves +7.1% mIoU improvement over prior one-shot UDA methods by combining personalized diffusion generation with adversarial training-based domain adaptation.

No commits in the last 6 months.