Class AttnConfig

custum_3d_diffusion/custum_modules/unifield_processor.py:12–77

* CrossAttention: Attention module (inherits knowledge), LoRA module (achieves fine-tuning), IPAdapter module (achieves conceptual control). * SelfAttention: Attention module (inherits knowledge), LoRA module (achieves fine-tuning), Reference Attention module (achieves pixel-level c

@dataclass
class AttnConfig:
    """        
    * CrossAttention: Attention module (inherits knowledge), LoRA module (achieves fine-tuning), IPAdapter module (achieves conceptual control).
    * SelfAttention: Attention module (inherits knowledge), LoRA module (achieves fine-tuning), Reference Attention module (achieves pixel-level control).
    * Multiview Attention module: Multiview Attention module (achieves multi-view consistency).
    * Cross Modality Attention module: Cross Modality Attention module (achieves multi-modality consistency).
    
    For setups:
        train_xxx_lr is implemented in the U-Net architecture.
        enable_xxx_lora is implemented in the U-Net architecture.
        enable_xxx_ip is implemented in the processor and U-Net architecture.
        enable_xxx_ref_proj_in is implemented in the processor.
    """
    latent_size: int = 64
    
    train_lr: float = 0
    # for cross attention
    # 0 learning rate for not training
    train_cross_attn_lr: float = 0
    train_cross_attn_lora_lr: float = 0       
    train_cross_attn_ip_lr: float = 0      # 0 for not trained
    init_cross_attn_lora: bool = False
    enable_cross_attn_lora: bool = False
    init_cross_attn_ip: bool = False
    enable_cross_attn_ip: bool = False
    cross_attn_lora_rank: int = 64        # 0 for not enabled
    cross_attn_lora_only_kv: bool = False
    ipadapter_pretrained_name: str = "h94/IP-Adapter"
    ipadapter_subfolder_name: str = "models"
    ipadapter_weight_name: str = "ip-adapter-plus_sd15.safetensors"
    ipadapter_effect_on: str = "all"    # all, first

    # for self attention
    train_self_attn_lr: float = 0
    train_self_attn_lora_lr: float = 0
    init_self_attn_lora: bool = False
    enable_self_attn_lora: bool = False
    self_attn_lora_rank: int = 64
    self_attn_lora_only_kv: bool = False

    train_self_attn_ref_lr: float = 0
    train_ref_unet_lr: float = 0
    init_self_attn_ref: bool = False
    enable_self_attn_ref: bool = False      
    self_attn_ref_other_model_name: str = ""
    self_attn_ref_position: str = "attn1"
    self_attn_ref_pixel_wise_crosspond: bool = False    # enable pixel_wise_crosspond in refattn
    self_attn_ref_chain_pos: str = "parralle"           # before or parralle or after
    self_attn_ref_effect_on: str = "all"                # all or first, for _crosspond attn
    self_attn_ref_zero_init: bool = True
    use_simple3d_attn: bool = False

    # for multiview attention
    init_multiview_attn: bool = False
    enable_multiview_attn: bool = False
    multiview_attn_position: str = "attn1"
    multiview_chain_pose: str = "parralle"             # before or parralle or after
    num_modalities: int = 1