model:
  target: rdm.models.diffusion.ddpm.RDM
  params:
    linear_start: 0.0015
    linear_end: 0.0195
    num_timesteps_cond: 1
    log_every_t: 200
    timesteps: 1000
    first_stage_key: image
    cond_stage_key: class_label
    class_cond: true
    image_size: 1
    channels: 768
    cond_stage_trainable: true
    conditioning_key: crossattn
    parameterization: x0
    use_mean_pooling: true
    unet_config:
      target: rdm.modules.diffusionmodules.latentmlp.SimpleMLP
      params:
        in_channels: 768
        time_embed_dim: 256
        model_channels: 1536
        bottleneck_channels: 1536
        out_channels: 768
        num_res_blocks: 12
        use_context: true
        context_channels: 512
    pretrained_enc_config:
      params:
        pretrained_enc_arch: dinov3_vitb16
        # Set to the absolute path of the downloaded DINOv3 ViT-B/16 checkpoint
        # (e.g., dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth). If left null,
        # torch.hub will attempt to download using the dinov3 repo helpers.
        pretrained_enc_path: /proj/mmfm/kimhi/EnrichCondImageGenerationUsingRCG/TrainRDM/dinov3_vit_b_16_checkpoint.pth
        proj_dim: 768
        pretrained_enc_withproj: false
        dinov3_check_hash: false
        # set to false to allow loading checkpoints that don't exactly match the repo architecture
        dinov3_load_strict: false
    cond_stage_config:
      target: rdm.modules.encoders.modules.ClassEmbedder
      params:
        embed_dim: 512
        n_classes: 1000
        key: class_label

# Made with Bob