Upload comfy/cldm/cldm.py with huggingface_hub
Browse files- comfy/cldm/cldm.py +433 -0
comfy/cldm/cldm.py
ADDED
|
@@ -0,0 +1,433 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#taken from: https://github.com/lllyasviel/ControlNet
|
| 2 |
+
#and modified
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
import torch.nn as nn
|
| 6 |
+
|
| 7 |
+
from ..ldm.modules.diffusionmodules.util import (
|
| 8 |
+
timestep_embedding,
|
| 9 |
+
)
|
| 10 |
+
|
| 11 |
+
from ..ldm.modules.attention import SpatialTransformer
|
| 12 |
+
from ..ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample
|
| 13 |
+
from ..ldm.util import exists
|
| 14 |
+
from .control_types import UNION_CONTROLNET_TYPES
|
| 15 |
+
from collections import OrderedDict
|
| 16 |
+
import comfy.ops
|
| 17 |
+
from comfy.ldm.modules.attention import optimized_attention
|
| 18 |
+
|
| 19 |
+
class OptimizedAttention(nn.Module):
|
| 20 |
+
def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None):
|
| 21 |
+
super().__init__()
|
| 22 |
+
self.heads = nhead
|
| 23 |
+
self.c = c
|
| 24 |
+
|
| 25 |
+
self.in_proj = operations.Linear(c, c * 3, bias=True, dtype=dtype, device=device)
|
| 26 |
+
self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
|
| 27 |
+
|
| 28 |
+
def forward(self, x):
|
| 29 |
+
x = self.in_proj(x)
|
| 30 |
+
q, k, v = x.split(self.c, dim=2)
|
| 31 |
+
out = optimized_attention(q, k, v, self.heads)
|
| 32 |
+
return self.out_proj(out)
|
| 33 |
+
|
| 34 |
+
class QuickGELU(nn.Module):
|
| 35 |
+
def forward(self, x: torch.Tensor):
|
| 36 |
+
return x * torch.sigmoid(1.702 * x)
|
| 37 |
+
|
| 38 |
+
class ResBlockUnionControlnet(nn.Module):
|
| 39 |
+
def __init__(self, dim, nhead, dtype=None, device=None, operations=None):
|
| 40 |
+
super().__init__()
|
| 41 |
+
self.attn = OptimizedAttention(dim, nhead, dtype=dtype, device=device, operations=operations)
|
| 42 |
+
self.ln_1 = operations.LayerNorm(dim, dtype=dtype, device=device)
|
| 43 |
+
self.mlp = nn.Sequential(
|
| 44 |
+
OrderedDict([("c_fc", operations.Linear(dim, dim * 4, dtype=dtype, device=device)), ("gelu", QuickGELU()),
|
| 45 |
+
("c_proj", operations.Linear(dim * 4, dim, dtype=dtype, device=device))]))
|
| 46 |
+
self.ln_2 = operations.LayerNorm(dim, dtype=dtype, device=device)
|
| 47 |
+
|
| 48 |
+
def attention(self, x: torch.Tensor):
|
| 49 |
+
return self.attn(x)
|
| 50 |
+
|
| 51 |
+
def forward(self, x: torch.Tensor):
|
| 52 |
+
x = x + self.attention(self.ln_1(x))
|
| 53 |
+
x = x + self.mlp(self.ln_2(x))
|
| 54 |
+
return x
|
| 55 |
+
|
| 56 |
+
class ControlledUnetModel(UNetModel):
|
| 57 |
+
#implemented in the ldm unet
|
| 58 |
+
pass
|
| 59 |
+
|
| 60 |
+
class ControlNet(nn.Module):
|
| 61 |
+
def __init__(
|
| 62 |
+
self,
|
| 63 |
+
image_size,
|
| 64 |
+
in_channels,
|
| 65 |
+
model_channels,
|
| 66 |
+
hint_channels,
|
| 67 |
+
num_res_blocks,
|
| 68 |
+
dropout=0,
|
| 69 |
+
channel_mult=(1, 2, 4, 8),
|
| 70 |
+
conv_resample=True,
|
| 71 |
+
dims=2,
|
| 72 |
+
num_classes=None,
|
| 73 |
+
use_checkpoint=False,
|
| 74 |
+
dtype=torch.float32,
|
| 75 |
+
num_heads=-1,
|
| 76 |
+
num_head_channels=-1,
|
| 77 |
+
num_heads_upsample=-1,
|
| 78 |
+
use_scale_shift_norm=False,
|
| 79 |
+
resblock_updown=False,
|
| 80 |
+
use_new_attention_order=False,
|
| 81 |
+
use_spatial_transformer=False, # custom transformer support
|
| 82 |
+
transformer_depth=1, # custom transformer support
|
| 83 |
+
context_dim=None, # custom transformer support
|
| 84 |
+
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
|
| 85 |
+
legacy=True,
|
| 86 |
+
disable_self_attentions=None,
|
| 87 |
+
num_attention_blocks=None,
|
| 88 |
+
disable_middle_self_attn=False,
|
| 89 |
+
use_linear_in_transformer=False,
|
| 90 |
+
adm_in_channels=None,
|
| 91 |
+
transformer_depth_middle=None,
|
| 92 |
+
transformer_depth_output=None,
|
| 93 |
+
attn_precision=None,
|
| 94 |
+
union_controlnet_num_control_type=None,
|
| 95 |
+
device=None,
|
| 96 |
+
operations=comfy.ops.disable_weight_init,
|
| 97 |
+
**kwargs,
|
| 98 |
+
):
|
| 99 |
+
super().__init__()
|
| 100 |
+
assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
|
| 101 |
+
if use_spatial_transformer:
|
| 102 |
+
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
|
| 103 |
+
|
| 104 |
+
if context_dim is not None:
|
| 105 |
+
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
|
| 106 |
+
# from omegaconf.listconfig import ListConfig
|
| 107 |
+
# if type(context_dim) == ListConfig:
|
| 108 |
+
# context_dim = list(context_dim)
|
| 109 |
+
|
| 110 |
+
if num_heads_upsample == -1:
|
| 111 |
+
num_heads_upsample = num_heads
|
| 112 |
+
|
| 113 |
+
if num_heads == -1:
|
| 114 |
+
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
|
| 115 |
+
|
| 116 |
+
if num_head_channels == -1:
|
| 117 |
+
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
|
| 118 |
+
|
| 119 |
+
self.dims = dims
|
| 120 |
+
self.image_size = image_size
|
| 121 |
+
self.in_channels = in_channels
|
| 122 |
+
self.model_channels = model_channels
|
| 123 |
+
|
| 124 |
+
if isinstance(num_res_blocks, int):
|
| 125 |
+
self.num_res_blocks = len(channel_mult) * [num_res_blocks]
|
| 126 |
+
else:
|
| 127 |
+
if len(num_res_blocks) != len(channel_mult):
|
| 128 |
+
raise ValueError("provide num_res_blocks either as an int (globally constant) or "
|
| 129 |
+
"as a list/tuple (per-level) with the same length as channel_mult")
|
| 130 |
+
self.num_res_blocks = num_res_blocks
|
| 131 |
+
|
| 132 |
+
if disable_self_attentions is not None:
|
| 133 |
+
# should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
|
| 134 |
+
assert len(disable_self_attentions) == len(channel_mult)
|
| 135 |
+
if num_attention_blocks is not None:
|
| 136 |
+
assert len(num_attention_blocks) == len(self.num_res_blocks)
|
| 137 |
+
assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
|
| 138 |
+
|
| 139 |
+
transformer_depth = transformer_depth[:]
|
| 140 |
+
|
| 141 |
+
self.dropout = dropout
|
| 142 |
+
self.channel_mult = channel_mult
|
| 143 |
+
self.conv_resample = conv_resample
|
| 144 |
+
self.num_classes = num_classes
|
| 145 |
+
self.use_checkpoint = use_checkpoint
|
| 146 |
+
self.dtype = dtype
|
| 147 |
+
self.num_heads = num_heads
|
| 148 |
+
self.num_head_channels = num_head_channels
|
| 149 |
+
self.num_heads_upsample = num_heads_upsample
|
| 150 |
+
self.predict_codebook_ids = n_embed is not None
|
| 151 |
+
|
| 152 |
+
time_embed_dim = model_channels * 4
|
| 153 |
+
self.time_embed = nn.Sequential(
|
| 154 |
+
operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
|
| 155 |
+
nn.SiLU(),
|
| 156 |
+
operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
|
| 157 |
+
)
|
| 158 |
+
|
| 159 |
+
if self.num_classes is not None:
|
| 160 |
+
if isinstance(self.num_classes, int):
|
| 161 |
+
self.label_emb = nn.Embedding(num_classes, time_embed_dim)
|
| 162 |
+
elif self.num_classes == "continuous":
|
| 163 |
+
self.label_emb = nn.Linear(1, time_embed_dim)
|
| 164 |
+
elif self.num_classes == "sequential":
|
| 165 |
+
assert adm_in_channels is not None
|
| 166 |
+
self.label_emb = nn.Sequential(
|
| 167 |
+
nn.Sequential(
|
| 168 |
+
operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
|
| 169 |
+
nn.SiLU(),
|
| 170 |
+
operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
|
| 171 |
+
)
|
| 172 |
+
)
|
| 173 |
+
else:
|
| 174 |
+
raise ValueError()
|
| 175 |
+
|
| 176 |
+
self.input_blocks = nn.ModuleList(
|
| 177 |
+
[
|
| 178 |
+
TimestepEmbedSequential(
|
| 179 |
+
operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
|
| 180 |
+
)
|
| 181 |
+
]
|
| 182 |
+
)
|
| 183 |
+
self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels, operations=operations, dtype=self.dtype, device=device)])
|
| 184 |
+
|
| 185 |
+
self.input_hint_block = TimestepEmbedSequential(
|
| 186 |
+
operations.conv_nd(dims, hint_channels, 16, 3, padding=1, dtype=self.dtype, device=device),
|
| 187 |
+
nn.SiLU(),
|
| 188 |
+
operations.conv_nd(dims, 16, 16, 3, padding=1, dtype=self.dtype, device=device),
|
| 189 |
+
nn.SiLU(),
|
| 190 |
+
operations.conv_nd(dims, 16, 32, 3, padding=1, stride=2, dtype=self.dtype, device=device),
|
| 191 |
+
nn.SiLU(),
|
| 192 |
+
operations.conv_nd(dims, 32, 32, 3, padding=1, dtype=self.dtype, device=device),
|
| 193 |
+
nn.SiLU(),
|
| 194 |
+
operations.conv_nd(dims, 32, 96, 3, padding=1, stride=2, dtype=self.dtype, device=device),
|
| 195 |
+
nn.SiLU(),
|
| 196 |
+
operations.conv_nd(dims, 96, 96, 3, padding=1, dtype=self.dtype, device=device),
|
| 197 |
+
nn.SiLU(),
|
| 198 |
+
operations.conv_nd(dims, 96, 256, 3, padding=1, stride=2, dtype=self.dtype, device=device),
|
| 199 |
+
nn.SiLU(),
|
| 200 |
+
operations.conv_nd(dims, 256, model_channels, 3, padding=1, dtype=self.dtype, device=device)
|
| 201 |
+
)
|
| 202 |
+
|
| 203 |
+
self._feature_size = model_channels
|
| 204 |
+
input_block_chans = [model_channels]
|
| 205 |
+
ch = model_channels
|
| 206 |
+
ds = 1
|
| 207 |
+
for level, mult in enumerate(channel_mult):
|
| 208 |
+
for nr in range(self.num_res_blocks[level]):
|
| 209 |
+
layers = [
|
| 210 |
+
ResBlock(
|
| 211 |
+
ch,
|
| 212 |
+
time_embed_dim,
|
| 213 |
+
dropout,
|
| 214 |
+
out_channels=mult * model_channels,
|
| 215 |
+
dims=dims,
|
| 216 |
+
use_checkpoint=use_checkpoint,
|
| 217 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 218 |
+
dtype=self.dtype,
|
| 219 |
+
device=device,
|
| 220 |
+
operations=operations,
|
| 221 |
+
)
|
| 222 |
+
]
|
| 223 |
+
ch = mult * model_channels
|
| 224 |
+
num_transformers = transformer_depth.pop(0)
|
| 225 |
+
if num_transformers > 0:
|
| 226 |
+
if num_head_channels == -1:
|
| 227 |
+
dim_head = ch // num_heads
|
| 228 |
+
else:
|
| 229 |
+
num_heads = ch // num_head_channels
|
| 230 |
+
dim_head = num_head_channels
|
| 231 |
+
if legacy:
|
| 232 |
+
#num_heads = 1
|
| 233 |
+
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
| 234 |
+
if exists(disable_self_attentions):
|
| 235 |
+
disabled_sa = disable_self_attentions[level]
|
| 236 |
+
else:
|
| 237 |
+
disabled_sa = False
|
| 238 |
+
|
| 239 |
+
if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
|
| 240 |
+
layers.append(
|
| 241 |
+
SpatialTransformer(
|
| 242 |
+
ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
|
| 243 |
+
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
|
| 244 |
+
use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
|
| 245 |
+
)
|
| 246 |
+
)
|
| 247 |
+
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
| 248 |
+
self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
|
| 249 |
+
self._feature_size += ch
|
| 250 |
+
input_block_chans.append(ch)
|
| 251 |
+
if level != len(channel_mult) - 1:
|
| 252 |
+
out_ch = ch
|
| 253 |
+
self.input_blocks.append(
|
| 254 |
+
TimestepEmbedSequential(
|
| 255 |
+
ResBlock(
|
| 256 |
+
ch,
|
| 257 |
+
time_embed_dim,
|
| 258 |
+
dropout,
|
| 259 |
+
out_channels=out_ch,
|
| 260 |
+
dims=dims,
|
| 261 |
+
use_checkpoint=use_checkpoint,
|
| 262 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 263 |
+
down=True,
|
| 264 |
+
dtype=self.dtype,
|
| 265 |
+
device=device,
|
| 266 |
+
operations=operations
|
| 267 |
+
)
|
| 268 |
+
if resblock_updown
|
| 269 |
+
else Downsample(
|
| 270 |
+
ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
|
| 271 |
+
)
|
| 272 |
+
)
|
| 273 |
+
)
|
| 274 |
+
ch = out_ch
|
| 275 |
+
input_block_chans.append(ch)
|
| 276 |
+
self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
|
| 277 |
+
ds *= 2
|
| 278 |
+
self._feature_size += ch
|
| 279 |
+
|
| 280 |
+
if num_head_channels == -1:
|
| 281 |
+
dim_head = ch // num_heads
|
| 282 |
+
else:
|
| 283 |
+
num_heads = ch // num_head_channels
|
| 284 |
+
dim_head = num_head_channels
|
| 285 |
+
if legacy:
|
| 286 |
+
#num_heads = 1
|
| 287 |
+
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
| 288 |
+
mid_block = [
|
| 289 |
+
ResBlock(
|
| 290 |
+
ch,
|
| 291 |
+
time_embed_dim,
|
| 292 |
+
dropout,
|
| 293 |
+
dims=dims,
|
| 294 |
+
use_checkpoint=use_checkpoint,
|
| 295 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 296 |
+
dtype=self.dtype,
|
| 297 |
+
device=device,
|
| 298 |
+
operations=operations
|
| 299 |
+
)]
|
| 300 |
+
if transformer_depth_middle >= 0:
|
| 301 |
+
mid_block += [SpatialTransformer( # always uses a self-attn
|
| 302 |
+
ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
|
| 303 |
+
disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
|
| 304 |
+
use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
|
| 305 |
+
),
|
| 306 |
+
ResBlock(
|
| 307 |
+
ch,
|
| 308 |
+
time_embed_dim,
|
| 309 |
+
dropout,
|
| 310 |
+
dims=dims,
|
| 311 |
+
use_checkpoint=use_checkpoint,
|
| 312 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 313 |
+
dtype=self.dtype,
|
| 314 |
+
device=device,
|
| 315 |
+
operations=operations
|
| 316 |
+
)]
|
| 317 |
+
self.middle_block = TimestepEmbedSequential(*mid_block)
|
| 318 |
+
self.middle_block_out = self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device)
|
| 319 |
+
self._feature_size += ch
|
| 320 |
+
|
| 321 |
+
if union_controlnet_num_control_type is not None:
|
| 322 |
+
self.num_control_type = union_controlnet_num_control_type
|
| 323 |
+
num_trans_channel = 320
|
| 324 |
+
num_trans_head = 8
|
| 325 |
+
num_trans_layer = 1
|
| 326 |
+
num_proj_channel = 320
|
| 327 |
+
# task_scale_factor = num_trans_channel ** 0.5
|
| 328 |
+
self.task_embedding = nn.Parameter(torch.empty(self.num_control_type, num_trans_channel, dtype=self.dtype, device=device))
|
| 329 |
+
|
| 330 |
+
self.transformer_layes = nn.Sequential(*[ResBlockUnionControlnet(num_trans_channel, num_trans_head, dtype=self.dtype, device=device, operations=operations) for _ in range(num_trans_layer)])
|
| 331 |
+
self.spatial_ch_projs = operations.Linear(num_trans_channel, num_proj_channel, dtype=self.dtype, device=device)
|
| 332 |
+
#-----------------------------------------------------------------------------------------------------
|
| 333 |
+
|
| 334 |
+
control_add_embed_dim = 256
|
| 335 |
+
class ControlAddEmbedding(nn.Module):
|
| 336 |
+
def __init__(self, in_dim, out_dim, num_control_type, dtype=None, device=None, operations=None):
|
| 337 |
+
super().__init__()
|
| 338 |
+
self.num_control_type = num_control_type
|
| 339 |
+
self.in_dim = in_dim
|
| 340 |
+
self.linear_1 = operations.Linear(in_dim * num_control_type, out_dim, dtype=dtype, device=device)
|
| 341 |
+
self.linear_2 = operations.Linear(out_dim, out_dim, dtype=dtype, device=device)
|
| 342 |
+
def forward(self, control_type, dtype, device):
|
| 343 |
+
c_type = torch.zeros((self.num_control_type,), device=device)
|
| 344 |
+
c_type[control_type] = 1.0
|
| 345 |
+
c_type = timestep_embedding(c_type.flatten(), self.in_dim, repeat_only=False).to(dtype).reshape((-1, self.num_control_type * self.in_dim))
|
| 346 |
+
return self.linear_2(torch.nn.functional.silu(self.linear_1(c_type)))
|
| 347 |
+
|
| 348 |
+
self.control_add_embedding = ControlAddEmbedding(control_add_embed_dim, time_embed_dim, self.num_control_type, dtype=self.dtype, device=device, operations=operations)
|
| 349 |
+
else:
|
| 350 |
+
self.task_embedding = None
|
| 351 |
+
self.control_add_embedding = None
|
| 352 |
+
|
| 353 |
+
def union_controlnet_merge(self, hint, control_type, emb, context):
|
| 354 |
+
# Equivalent to: https://github.com/xinsir6/ControlNetPlus/tree/main
|
| 355 |
+
inputs = []
|
| 356 |
+
condition_list = []
|
| 357 |
+
|
| 358 |
+
for idx in range(min(1, len(control_type))):
|
| 359 |
+
controlnet_cond = self.input_hint_block(hint[idx], emb, context)
|
| 360 |
+
feat_seq = torch.mean(controlnet_cond, dim=(2, 3))
|
| 361 |
+
if idx < len(control_type):
|
| 362 |
+
feat_seq += self.task_embedding[control_type[idx]].to(dtype=feat_seq.dtype, device=feat_seq.device)
|
| 363 |
+
|
| 364 |
+
inputs.append(feat_seq.unsqueeze(1))
|
| 365 |
+
condition_list.append(controlnet_cond)
|
| 366 |
+
|
| 367 |
+
x = torch.cat(inputs, dim=1)
|
| 368 |
+
x = self.transformer_layes(x)
|
| 369 |
+
controlnet_cond_fuser = None
|
| 370 |
+
for idx in range(len(control_type)):
|
| 371 |
+
alpha = self.spatial_ch_projs(x[:, idx])
|
| 372 |
+
alpha = alpha.unsqueeze(-1).unsqueeze(-1)
|
| 373 |
+
o = condition_list[idx] + alpha
|
| 374 |
+
if controlnet_cond_fuser is None:
|
| 375 |
+
controlnet_cond_fuser = o
|
| 376 |
+
else:
|
| 377 |
+
controlnet_cond_fuser += o
|
| 378 |
+
return controlnet_cond_fuser
|
| 379 |
+
|
| 380 |
+
def make_zero_conv(self, channels, operations=None, dtype=None, device=None):
|
| 381 |
+
return TimestepEmbedSequential(operations.conv_nd(self.dims, channels, channels, 1, padding=0, dtype=dtype, device=device))
|
| 382 |
+
|
| 383 |
+
def forward(self, x, hint, timesteps, context, y=None, **kwargs):
|
| 384 |
+
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
|
| 385 |
+
emb = self.time_embed(t_emb)
|
| 386 |
+
|
| 387 |
+
guided_hint = None
|
| 388 |
+
if self.control_add_embedding is not None: #Union Controlnet
|
| 389 |
+
control_type = kwargs.get("control_type", [])
|
| 390 |
+
|
| 391 |
+
if any([c >= self.num_control_type for c in control_type]):
|
| 392 |
+
max_type = max(control_type)
|
| 393 |
+
max_type_name = {
|
| 394 |
+
v: k for k, v in UNION_CONTROLNET_TYPES.items()
|
| 395 |
+
}[max_type]
|
| 396 |
+
raise ValueError(
|
| 397 |
+
f"Control type {max_type_name}({max_type}) is out of range for the number of control types" +
|
| 398 |
+
f"({self.num_control_type}) supported.\n" +
|
| 399 |
+
"Please consider using the ProMax ControlNet Union model.\n" +
|
| 400 |
+
"https://huggingface.co/xinsir/controlnet-union-sdxl-1.0/tree/main"
|
| 401 |
+
)
|
| 402 |
+
|
| 403 |
+
emb += self.control_add_embedding(control_type, emb.dtype, emb.device)
|
| 404 |
+
if len(control_type) > 0:
|
| 405 |
+
if len(hint.shape) < 5:
|
| 406 |
+
hint = hint.unsqueeze(dim=0)
|
| 407 |
+
guided_hint = self.union_controlnet_merge(hint, control_type, emb, context)
|
| 408 |
+
|
| 409 |
+
if guided_hint is None:
|
| 410 |
+
guided_hint = self.input_hint_block(hint, emb, context)
|
| 411 |
+
|
| 412 |
+
out_output = []
|
| 413 |
+
out_middle = []
|
| 414 |
+
|
| 415 |
+
if self.num_classes is not None:
|
| 416 |
+
assert y.shape[0] == x.shape[0]
|
| 417 |
+
emb = emb + self.label_emb(y)
|
| 418 |
+
|
| 419 |
+
h = x
|
| 420 |
+
for module, zero_conv in zip(self.input_blocks, self.zero_convs):
|
| 421 |
+
if guided_hint is not None:
|
| 422 |
+
h = module(h, emb, context)
|
| 423 |
+
h += guided_hint
|
| 424 |
+
guided_hint = None
|
| 425 |
+
else:
|
| 426 |
+
h = module(h, emb, context)
|
| 427 |
+
out_output.append(zero_conv(h, emb, context))
|
| 428 |
+
|
| 429 |
+
h = self.middle_block(h, emb, context)
|
| 430 |
+
out_middle.append(self.middle_block_out(h, emb, context))
|
| 431 |
+
|
| 432 |
+
return {"middle": out_middle, "output": out_output}
|
| 433 |
+
|