cryofm-v2

CryoFM2: A Generative Foundation Model for Cryo-EM Densities Overview CryoFM2** is a flow-based generative foundation model for cryo-EM density maps. It is pretrained on curated EMDB half maps to learn general priors of high-quality cryo-EM densities and can be fine-tuned for downstream tasks. The model learns a continuous mapping from a simple Gaussian distribution to the complex distribution of cryo-EM densities, enabling stable generation and flexible adaptation. CryoFM2 can also act as a **Bayesian prior**, integrating naturally with task-specific likelihoods to support applications such as anisotropy-aware refinement, non-uniform reconstruction, and controlled density modification. Model Details CryoFM2 is pretrained on curated EMDB half maps to learn general priors of high-quality cryo-EM densities. The model can be fine-tuned for various downstream tasks such as density map enhancement and post-processing. Pre-training Architecture:** Fine-tuning Architecture (for EMhancer/EMReady style post-processing):** Architecture **Architecture Type**: 3D UNet **Input Size**: 64×64×64 voxels **Input Channels**: 2 for pre-trained model, 3 for fine-tuned model **Output Channels**: 1 **Down Blocks**: DownBlock3D, DownBlock3D, AttnDownBlock3D, AttnDownBlock3D **Up Blocks**: AttnUpBlock3D, AttnUpBlock3D, UpBlock3D, UpBlock3D **Block Output Channels**: (64, 128, 256, 512) **Layers per Block**: 2 **Attention Head Dimension**: 8 **Normalization**: GroupNorm (32 groups) **Activation**: SiLU **Time Embedding**: Positional encoding Model Variants 1. **cryofm2-pretrain**: Unconditional pretrained model for general density map generation 2. **cryofm2-emhancer**: Fine-tuned model for density map enhancement (EMhancer style) 3. **cryofm2-emready**: Fine-tuned model for density map enhancement (EMReady style) Play with CryoFM2 Installation Before using CryoFM2, you need to set up the environment and install the package. Follow these steps to get started: Unconditional Generation (Explo