Diffusers documentation
StableCascadeUNet
StableCascadeUNet
A UNet model from the Stable Cascade pipeline.
StableCascadeUNet
class diffusers.models.StableCascadeUNet
< source >( in_channels: int = 16out_channels: int = 16timestep_ratio_embedding_dim: int = 64patch_size: int = 1conditioning_dim: int = 2048block_out_channels: tuple = (2048, 2048)num_attention_heads: tuple = (32, 32)down_num_layers_per_block: tuple = (8, 24)up_num_layers_per_block: tuple = (24, 8)down_blocks_repeat_mappers: tuple[int] | None = (1, 1)up_blocks_repeat_mappers: tuple[int] | None = (1, 1)block_types_per_layer: tuple = (('SDCascadeResBlock', 'SDCascadeTimestepBlock', 'SDCascadeAttnBlock'), ('SDCascadeResBlock', 'SDCascadeTimestepBlock', 'SDCascadeAttnBlock'))clip_text_in_channels: int | None = Noneclip_text_pooled_in_channels = 1280clip_image_in_channels: int | None = Noneclip_seq = 4effnet_in_channels: int | None = Nonepixel_mapper_in_channels: int | None = Nonekernel_size = 3dropout: float | tuple[float] = (0.1, 0.1)self_attn: bool | tuple[bool] = Truetimestep_conditioning_type: tuple = ('sca', 'crp')switch_level: tuple[bool] | None = None )
forward
< source >( sampletimestep_ratioclip_text_pooledclip_text = Noneclip_img = Noneeffnet = Nonepixels = Nonesca = Nonecrp = Nonereturn_dict = True )
Parameters
- sample (
torch.Tensor) — The noisy input sample. - timestep_ratio (
torch.Tensor) — Timestep ratio used to compute the timestep embedding. - clip_text_pooled (
torch.Tensor) — Pooled CLIP text embeddings. - clip_text (
torch.Tensor, optional) — Sequence-level CLIP text embeddings. - clip_img (
torch.Tensor, optional) — CLIP image embeddings. - effnet (
torch.Tensor, optional) — EfficientNet feature map used as additional conditioning. - pixels (
torch.Tensor, optional) — Pixel-level conditioning tensor. IfNone, a tensor of zeros is used. - sca (
torch.Tensor, optional) — Optionalscaconditioning value used to build the timestep embedding. - crp (
torch.Tensor, optional) — Optionalcrpconditioning value used to build the timestep embedding. - return_dict (
bool, optional, defaults toTrue) — Whether or not to return aStableCascadeUNetOutputinstead of a plain tuple.