(
self,
in_channels: int = 4,
conditioning_channels: int = 5,
flip_sin_to_cos: bool = True,
freq_shift: int = 0,
down_block_types: Tuple[str, ...] = (
"DownBlock2D",
"DownBlock2D",
"DownBlock2D",
"DownBlock2D",
),
mid_block_type: Optional[str] = "UNetMidBlock2D",
up_block_types: Tuple[str, ...] = (
"UpBlock2D",
"UpBlock2D",
"UpBlock2D",
"UpBlock2D",
),
only_cross_attention: Union[bool, Tuple[bool]] = False,
block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
layers_per_block: int = 2,
downsample_padding: int = 1,
mid_block_scale_factor: float = 1,
act_fn: str = "silu",
norm_num_groups: Optional[int] = 32,
norm_eps: float = 1e-5,
cross_attention_dim: int = 1280,
transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
encoder_hid_dim: Optional[int] = None,
encoder_hid_dim_type: Optional[str] = None,
attention_head_dim: Union[int, Tuple[int, ...]] = 8,
num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
use_linear_projection: bool = False,
class_embed_type: Optional[str] = None,
addition_embed_type: Optional[str] = None,
addition_time_embed_dim: Optional[int] = None,
num_class_embeds: Optional[int] = None,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
projection_class_embeddings_input_dim: Optional[int] = None,
brushnet_conditioning_channel_order: str = "rgb",
conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
global_pool_conditions: bool = False,
addition_embed_type_num_heads: int = 64,
)
| 136 | |
| 137 | @register_to_config |
| 138 | def __init__( |
| 139 | self, |
| 140 | in_channels: int = 4, |
| 141 | conditioning_channels: int = 5, |
| 142 | flip_sin_to_cos: bool = True, |
| 143 | freq_shift: int = 0, |
| 144 | down_block_types: Tuple[str, ...] = ( |
| 145 | "DownBlock2D", |
| 146 | "DownBlock2D", |
| 147 | "DownBlock2D", |
| 148 | "DownBlock2D", |
| 149 | ), |
| 150 | mid_block_type: Optional[str] = "UNetMidBlock2D", |
| 151 | up_block_types: Tuple[str, ...] = ( |
| 152 | "UpBlock2D", |
| 153 | "UpBlock2D", |
| 154 | "UpBlock2D", |
| 155 | "UpBlock2D", |
| 156 | ), |
| 157 | only_cross_attention: Union[bool, Tuple[bool]] = False, |
| 158 | block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280), |
| 159 | layers_per_block: int = 2, |
| 160 | downsample_padding: int = 1, |
| 161 | mid_block_scale_factor: float = 1, |
| 162 | act_fn: str = "silu", |
| 163 | norm_num_groups: Optional[int] = 32, |
| 164 | norm_eps: float = 1e-5, |
| 165 | cross_attention_dim: int = 1280, |
| 166 | transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1, |
| 167 | encoder_hid_dim: Optional[int] = None, |
| 168 | encoder_hid_dim_type: Optional[str] = None, |
| 169 | attention_head_dim: Union[int, Tuple[int, ...]] = 8, |
| 170 | num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None, |
| 171 | use_linear_projection: bool = False, |
| 172 | class_embed_type: Optional[str] = None, |
| 173 | addition_embed_type: Optional[str] = None, |
| 174 | addition_time_embed_dim: Optional[int] = None, |
| 175 | num_class_embeds: Optional[int] = None, |
| 176 | upcast_attention: bool = False, |
| 177 | resnet_time_scale_shift: str = "default", |
| 178 | projection_class_embeddings_input_dim: Optional[int] = None, |
| 179 | brushnet_conditioning_channel_order: str = "rgb", |
| 180 | conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256), |
| 181 | global_pool_conditions: bool = False, |
| 182 | addition_embed_type_num_heads: int = 64, |
| 183 | ): |
| 184 | super().__init__() |
| 185 | |
| 186 | # If `num_attention_heads` is not defined (which is the case for most models) |
| 187 | # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. |
| 188 | # The reason for this behavior is to correct for incorrectly named variables that were introduced |
| 189 | # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 |
| 190 | # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking |
| 191 | # which is why we correct for the naming here. |
| 192 | num_attention_heads = num_attention_heads or attention_head_dim |
| 193 | |
| 194 | # Check inputs |
| 195 | if len(down_block_types) != len(up_block_types): |
nothing calls this directly
no test coverage detected