HistoDiffusion++: A generative visual foundation model for histopathology

Histopathology image analysis has been revolutionized with the introduction of deep learning. While task-specific models have shown strong performance, they often rely on large annotated datasets that are difficult to obtain in practice. Recent foundation models mitigate this dependency via self-supervised learning and have achieved promising results in discriminative analysis and text generation. However, the tasks requiring image generation capabilities remain underexplored, especially under data-scarce conditions. Hence, we present HistoDiffusion++, a multi-magnification generative visual foundation model for histopathology. We pretrain the model on Histo-7M, a large-scale dataset of 7.4 million patches from 38,373 whole-slide images spanning 33 cancer types and 12 organs/organ systems at multiple magnifications. To improve training efficiency and synthesis quality, we further incorporate a pretrained discriminative foundation model and apply representation alignment as regularization during pretraining. We evaluate HistoDiffusion++ on two tracks: synthetic augmentation and image-to-image translation. For augmentation, our model improves classification and nuclei segmentation performance beyond baseline strategies. For translation, it achieves promising results in virtual staining and artifact restoration. The code and model checkpoints are available at https://github.com/PengJin95/HistoDiffusion_Plusplus.