Create controlnet_sd3.py
Browse files- controlnet_sd3.py +552 -0
controlnet_sd3.py
ADDED
@@ -0,0 +1,552 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
|
16 |
+
from dataclasses import dataclass
|
17 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
18 |
+
|
19 |
+
import torch
|
20 |
+
import torch.nn as nn
|
21 |
+
|
22 |
+
import diffusers
|
23 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
24 |
+
from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
|
25 |
+
from diffusers.models.attention import JointTransformerBlock
|
26 |
+
from diffusers.models.attention_processor import Attention, AttentionProcessor
|
27 |
+
from diffusers.models.modeling_utils import ModelMixin
|
28 |
+
from diffusers.utils import (
|
29 |
+
USE_PEFT_BACKEND,
|
30 |
+
is_torch_version,
|
31 |
+
logging,
|
32 |
+
scale_lora_layers,
|
33 |
+
unscale_lora_layers,
|
34 |
+
)
|
35 |
+
from diffusers.models.controlnet import BaseOutput, zero_module
|
36 |
+
from diffusers.models.embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
|
37 |
+
from diffusers.models.transformers.transformer_2d import Transformer2DModelOutput
|
38 |
+
from torch.nn import functional as F
|
39 |
+
|
40 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
41 |
+
from packaging import version
|
42 |
+
|
43 |
+
class ControlNetConditioningEmbedding(nn.Module):
|
44 |
+
"""
|
45 |
+
Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
|
46 |
+
[11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
|
47 |
+
training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
|
48 |
+
convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
|
49 |
+
(activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
|
50 |
+
model) to encode image-space conditions ... into feature maps ..."
|
51 |
+
"""
|
52 |
+
|
53 |
+
def __init__(
|
54 |
+
self,
|
55 |
+
conditioning_embedding_channels: int,
|
56 |
+
conditioning_channels: int = 3,
|
57 |
+
block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
|
58 |
+
):
|
59 |
+
super().__init__()
|
60 |
+
|
61 |
+
self.conv_in = nn.Conv2d(
|
62 |
+
conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
|
63 |
+
)
|
64 |
+
|
65 |
+
self.blocks = nn.ModuleList([])
|
66 |
+
|
67 |
+
for i in range(len(block_out_channels) - 1):
|
68 |
+
channel_in = block_out_channels[i]
|
69 |
+
channel_out = block_out_channels[i + 1]
|
70 |
+
self.blocks.append(
|
71 |
+
nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1)
|
72 |
+
)
|
73 |
+
self.blocks.append(
|
74 |
+
nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2)
|
75 |
+
)
|
76 |
+
|
77 |
+
self.conv_out = zero_module(
|
78 |
+
nn.Conv2d(
|
79 |
+
block_out_channels[-1],
|
80 |
+
conditioning_embedding_channels,
|
81 |
+
kernel_size=3,
|
82 |
+
padding=1,
|
83 |
+
)
|
84 |
+
)
|
85 |
+
|
86 |
+
def forward(self, conditioning):
|
87 |
+
embedding = self.conv_in(conditioning)
|
88 |
+
embedding = F.silu(embedding)
|
89 |
+
|
90 |
+
for block in self.blocks:
|
91 |
+
embedding = block(embedding)
|
92 |
+
embedding = F.silu(embedding)
|
93 |
+
|
94 |
+
embedding = self.conv_out(embedding)
|
95 |
+
|
96 |
+
return embedding
|
97 |
+
|
98 |
+
|
99 |
+
@dataclass
|
100 |
+
class SD3ControlNetOutput(BaseOutput):
|
101 |
+
controlnet_block_samples: Tuple[torch.Tensor]
|
102 |
+
|
103 |
+
|
104 |
+
class SD3ControlNetModel(
|
105 |
+
ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin
|
106 |
+
):
|
107 |
+
_supports_gradient_checkpointing = True
|
108 |
+
|
109 |
+
@register_to_config
|
110 |
+
def __init__(
|
111 |
+
self,
|
112 |
+
sample_size: int = 128,
|
113 |
+
patch_size: int = 2,
|
114 |
+
in_channels: int = 16,
|
115 |
+
num_layers: int = 18,
|
116 |
+
attention_head_dim: int = 64,
|
117 |
+
num_attention_heads: int = 18,
|
118 |
+
joint_attention_dim: int = 4096,
|
119 |
+
caption_projection_dim: int = 1152,
|
120 |
+
pooled_projection_dim: int = 2048,
|
121 |
+
out_channels: int = 16,
|
122 |
+
pos_embed_max_size: int = 96,
|
123 |
+
conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (
|
124 |
+
16,
|
125 |
+
32,
|
126 |
+
96,
|
127 |
+
256,
|
128 |
+
),
|
129 |
+
conditioning_channels: int = 3,
|
130 |
+
):
|
131 |
+
"""
|
132 |
+
conditioning_channels: condition image pixel space channels
|
133 |
+
conditioning_embedding_out_channels: intermediate channels
|
134 |
+
|
135 |
+
"""
|
136 |
+
super().__init__()
|
137 |
+
default_out_channels = in_channels
|
138 |
+
self.out_channels = (
|
139 |
+
out_channels if out_channels is not None else default_out_channels
|
140 |
+
)
|
141 |
+
self.inner_dim = num_attention_heads * attention_head_dim
|
142 |
+
|
143 |
+
self.pos_embed = PatchEmbed(
|
144 |
+
height=sample_size,
|
145 |
+
width=sample_size,
|
146 |
+
patch_size=patch_size,
|
147 |
+
in_channels=in_channels,
|
148 |
+
embed_dim=self.inner_dim,
|
149 |
+
pos_embed_max_size=pos_embed_max_size, # hard-code for now.
|
150 |
+
)
|
151 |
+
self.time_text_embed = CombinedTimestepTextProjEmbeddings(
|
152 |
+
embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
|
153 |
+
)
|
154 |
+
self.context_embedder = nn.Linear(joint_attention_dim, caption_projection_dim)
|
155 |
+
|
156 |
+
# control net conditioning embedding
|
157 |
+
# self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
|
158 |
+
# conditioning_embedding_channels=default_out_channels,
|
159 |
+
# block_out_channels=conditioning_embedding_out_channels,
|
160 |
+
# conditioning_channels=conditioning_channels,
|
161 |
+
# )
|
162 |
+
|
163 |
+
# `attention_head_dim` is doubled to account for the mixing.
|
164 |
+
# It needs to crafted when we get the actual checkpoints.
|
165 |
+
self.transformer_blocks = nn.ModuleList(
|
166 |
+
[
|
167 |
+
JointTransformerBlock(
|
168 |
+
dim=self.inner_dim,
|
169 |
+
num_attention_heads=num_attention_heads,
|
170 |
+
attention_head_dim=attention_head_dim if version.parse(diffusers.__version__) >= version.parse('0.30.0.dev0') else self.inner_dim,
|
171 |
+
context_pre_only=False,
|
172 |
+
)
|
173 |
+
for _ in range(num_layers)
|
174 |
+
]
|
175 |
+
)
|
176 |
+
|
177 |
+
# controlnet_blocks
|
178 |
+
self.controlnet_blocks = nn.ModuleList([])
|
179 |
+
for _ in range(len(self.transformer_blocks)):
|
180 |
+
controlnet_block = zero_module(nn.Linear(self.inner_dim, self.inner_dim))
|
181 |
+
self.controlnet_blocks.append(controlnet_block)
|
182 |
+
|
183 |
+
# control condition embedding
|
184 |
+
pos_embed_cond = PatchEmbed(
|
185 |
+
height=sample_size,
|
186 |
+
width=sample_size,
|
187 |
+
patch_size=patch_size,
|
188 |
+
in_channels=in_channels + 1,
|
189 |
+
embed_dim=self.inner_dim,
|
190 |
+
pos_embed_type=None,
|
191 |
+
)
|
192 |
+
# pos_embed_cond = nn.Linear(in_channels + 1, self.inner_dim)
|
193 |
+
self.pos_embed_cond = zero_module(pos_embed_cond)
|
194 |
+
|
195 |
+
self.gradient_checkpointing = False
|
196 |
+
|
197 |
+
# Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
|
198 |
+
def enable_forward_chunking(
|
199 |
+
self, chunk_size: Optional[int] = None, dim: int = 0
|
200 |
+
) -> None:
|
201 |
+
"""
|
202 |
+
Sets the attention processor to use [feed forward
|
203 |
+
chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
|
204 |
+
|
205 |
+
Parameters:
|
206 |
+
chunk_size (`int`, *optional*):
|
207 |
+
The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
|
208 |
+
over each tensor of dim=`dim`.
|
209 |
+
dim (`int`, *optional*, defaults to `0`):
|
210 |
+
The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
|
211 |
+
or dim=1 (sequence length).
|
212 |
+
"""
|
213 |
+
if dim not in [0, 1]:
|
214 |
+
raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
|
215 |
+
|
216 |
+
# By default chunk size is 1
|
217 |
+
chunk_size = chunk_size or 1
|
218 |
+
|
219 |
+
def fn_recursive_feed_forward(
|
220 |
+
module: torch.nn.Module, chunk_size: int, dim: int
|
221 |
+
):
|
222 |
+
if hasattr(module, "set_chunk_feed_forward"):
|
223 |
+
module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
|
224 |
+
|
225 |
+
for child in module.children():
|
226 |
+
fn_recursive_feed_forward(child, chunk_size, dim)
|
227 |
+
|
228 |
+
for module in self.children():
|
229 |
+
fn_recursive_feed_forward(module, chunk_size, dim)
|
230 |
+
|
231 |
+
@property
|
232 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
|
233 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
234 |
+
r"""
|
235 |
+
Returns:
|
236 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
237 |
+
indexed by its weight name.
|
238 |
+
"""
|
239 |
+
# set recursively
|
240 |
+
processors = {}
|
241 |
+
|
242 |
+
def fn_recursive_add_processors(
|
243 |
+
name: str,
|
244 |
+
module: torch.nn.Module,
|
245 |
+
processors: Dict[str, AttentionProcessor],
|
246 |
+
):
|
247 |
+
if hasattr(module, "get_processor"):
|
248 |
+
processors[f"{name}.processor"] = module.get_processor(
|
249 |
+
return_deprecated_lora=True
|
250 |
+
)
|
251 |
+
|
252 |
+
for sub_name, child in module.named_children():
|
253 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
254 |
+
|
255 |
+
return processors
|
256 |
+
|
257 |
+
for name, module in self.named_children():
|
258 |
+
fn_recursive_add_processors(name, module, processors)
|
259 |
+
|
260 |
+
return processors
|
261 |
+
|
262 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
|
263 |
+
def set_attn_processor(
|
264 |
+
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
|
265 |
+
):
|
266 |
+
r"""
|
267 |
+
Sets the attention processor to use to compute attention.
|
268 |
+
|
269 |
+
Parameters:
|
270 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
271 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
272 |
+
for **all** `Attention` layers.
|
273 |
+
|
274 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
275 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
276 |
+
|
277 |
+
"""
|
278 |
+
count = len(self.attn_processors.keys())
|
279 |
+
|
280 |
+
if isinstance(processor, dict) and len(processor) != count:
|
281 |
+
raise ValueError(
|
282 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
283 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
284 |
+
)
|
285 |
+
|
286 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
287 |
+
if hasattr(module, "set_processor"):
|
288 |
+
if not isinstance(processor, dict):
|
289 |
+
module.set_processor(processor)
|
290 |
+
else:
|
291 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
292 |
+
|
293 |
+
for sub_name, child in module.named_children():
|
294 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
295 |
+
|
296 |
+
for name, module in self.named_children():
|
297 |
+
fn_recursive_attn_processor(name, module, processor)
|
298 |
+
|
299 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
|
300 |
+
def fuse_qkv_projections(self):
|
301 |
+
"""
|
302 |
+
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
|
303 |
+
are fused. For cross-attention modules, key and value projection matrices are fused.
|
304 |
+
|
305 |
+
<Tip warning={true}>
|
306 |
+
|
307 |
+
This API is 🧪 experimental.
|
308 |
+
|
309 |
+
</Tip>
|
310 |
+
"""
|
311 |
+
self.original_attn_processors = None
|
312 |
+
|
313 |
+
for _, attn_processor in self.attn_processors.items():
|
314 |
+
if "Added" in str(attn_processor.__class__.__name__):
|
315 |
+
raise ValueError(
|
316 |
+
"`fuse_qkv_projections()` is not supported for models having added KV projections."
|
317 |
+
)
|
318 |
+
|
319 |
+
self.original_attn_processors = self.attn_processors
|
320 |
+
|
321 |
+
for module in self.modules():
|
322 |
+
if isinstance(module, Attention):
|
323 |
+
module.fuse_projections(fuse=True)
|
324 |
+
|
325 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
|
326 |
+
def unfuse_qkv_projections(self):
|
327 |
+
"""Disables the fused QKV projection if enabled.
|
328 |
+
|
329 |
+
<Tip warning={true}>
|
330 |
+
|
331 |
+
This API is 🧪 experimental.
|
332 |
+
|
333 |
+
</Tip>
|
334 |
+
|
335 |
+
"""
|
336 |
+
if self.original_attn_processors is not None:
|
337 |
+
self.set_attn_processor(self.original_attn_processors)
|
338 |
+
|
339 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
340 |
+
if hasattr(module, "gradient_checkpointing"):
|
341 |
+
module.gradient_checkpointing = value
|
342 |
+
|
343 |
+
@classmethod
|
344 |
+
def from_transformer(
|
345 |
+
cls, transformer, num_layers=None, load_weights_from_transformer=True
|
346 |
+
):
|
347 |
+
config = transformer.config
|
348 |
+
config["num_layers"] = num_layers or config.num_layers
|
349 |
+
controlnet = cls(**config)
|
350 |
+
|
351 |
+
if load_weights_from_transformer:
|
352 |
+
controlnet.pos_embed.load_state_dict(
|
353 |
+
transformer.pos_embed.state_dict(), strict=False
|
354 |
+
)
|
355 |
+
controlnet.time_text_embed.load_state_dict(
|
356 |
+
transformer.time_text_embed.state_dict(), strict=False
|
357 |
+
)
|
358 |
+
controlnet.context_embedder.load_state_dict(
|
359 |
+
transformer.context_embedder.state_dict(), strict=False
|
360 |
+
)
|
361 |
+
controlnet.transformer_blocks.load_state_dict(
|
362 |
+
transformer.transformer_blocks.state_dict(), strict=False
|
363 |
+
)
|
364 |
+
|
365 |
+
return controlnet
|
366 |
+
|
367 |
+
def forward(
|
368 |
+
self,
|
369 |
+
hidden_states: torch.FloatTensor,
|
370 |
+
controlnet_cond: torch.Tensor,
|
371 |
+
conditioning_scale: float = 1.0,
|
372 |
+
encoder_hidden_states: torch.FloatTensor = None,
|
373 |
+
pooled_projections: torch.FloatTensor = None,
|
374 |
+
timestep: torch.LongTensor = None,
|
375 |
+
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
376 |
+
return_dict: bool = True,
|
377 |
+
) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
|
378 |
+
"""
|
379 |
+
The [`SD3Transformer2DModel`] forward method.
|
380 |
+
|
381 |
+
Args:
|
382 |
+
hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
|
383 |
+
Input `hidden_states`.
|
384 |
+
controlnet_cond (`torch.Tensor`):
|
385 |
+
The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
|
386 |
+
conditioning_scale (`float`, defaults to `1.0`):
|
387 |
+
The scale factor for ControlNet outputs.
|
388 |
+
encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
|
389 |
+
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
|
390 |
+
pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
|
391 |
+
from the embeddings of input conditions.
|
392 |
+
timestep ( `torch.LongTensor`):
|
393 |
+
Used to indicate denoising step.
|
394 |
+
joint_attention_kwargs (`dict`, *optional*):
|
395 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
396 |
+
`self.processor` in
|
397 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
398 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
399 |
+
Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
|
400 |
+
tuple.
|
401 |
+
|
402 |
+
Returns:
|
403 |
+
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
|
404 |
+
`tuple` where the first element is the sample tensor.
|
405 |
+
"""
|
406 |
+
if joint_attention_kwargs is not None:
|
407 |
+
joint_attention_kwargs = joint_attention_kwargs.copy()
|
408 |
+
lora_scale = joint_attention_kwargs.pop("scale", 1.0)
|
409 |
+
else:
|
410 |
+
lora_scale = 1.0
|
411 |
+
|
412 |
+
if USE_PEFT_BACKEND:
|
413 |
+
# weight the lora layers by setting `lora_scale` for each PEFT layer
|
414 |
+
scale_lora_layers(self, lora_scale)
|
415 |
+
else:
|
416 |
+
if (
|
417 |
+
joint_attention_kwargs is not None
|
418 |
+
and joint_attention_kwargs.get("scale", None) is not None
|
419 |
+
):
|
420 |
+
logger.warning(
|
421 |
+
"Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
|
422 |
+
)
|
423 |
+
|
424 |
+
height, width = hidden_states.shape[-2:]
|
425 |
+
|
426 |
+
hidden_states = self.pos_embed(
|
427 |
+
hidden_states
|
428 |
+
) # takes care of adding positional embeddings too. b,c,H,W -> b, N, C
|
429 |
+
temb = self.time_text_embed(timestep, pooled_projections)
|
430 |
+
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
|
431 |
+
|
432 |
+
# add condition
|
433 |
+
hidden_states = hidden_states + self.pos_embed_cond(controlnet_cond)
|
434 |
+
|
435 |
+
block_res_samples = ()
|
436 |
+
|
437 |
+
for block in self.transformer_blocks:
|
438 |
+
if self.training and self.gradient_checkpointing:
|
439 |
+
|
440 |
+
def create_custom_forward(module, return_dict=None):
|
441 |
+
def custom_forward(*inputs):
|
442 |
+
if return_dict is not None:
|
443 |
+
return module(*inputs, return_dict=return_dict)
|
444 |
+
else:
|
445 |
+
return module(*inputs)
|
446 |
+
|
447 |
+
return custom_forward
|
448 |
+
|
449 |
+
ckpt_kwargs: Dict[str, Any] = (
|
450 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
451 |
+
)
|
452 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
453 |
+
create_custom_forward(block),
|
454 |
+
hidden_states,
|
455 |
+
encoder_hidden_states,
|
456 |
+
temb,
|
457 |
+
**ckpt_kwargs,
|
458 |
+
)
|
459 |
+
|
460 |
+
else:
|
461 |
+
encoder_hidden_states, hidden_states = block(
|
462 |
+
hidden_states=hidden_states,
|
463 |
+
encoder_hidden_states=encoder_hidden_states,
|
464 |
+
temb=temb,
|
465 |
+
)
|
466 |
+
|
467 |
+
block_res_samples = block_res_samples + (hidden_states,)
|
468 |
+
|
469 |
+
controlnet_block_res_samples = ()
|
470 |
+
for block_res_sample, controlnet_block in zip(
|
471 |
+
block_res_samples, self.controlnet_blocks
|
472 |
+
):
|
473 |
+
block_res_sample = controlnet_block(block_res_sample)
|
474 |
+
controlnet_block_res_samples = controlnet_block_res_samples + (
|
475 |
+
block_res_sample,
|
476 |
+
)
|
477 |
+
|
478 |
+
# 6. scaling
|
479 |
+
controlnet_block_res_samples = [
|
480 |
+
sample * conditioning_scale for sample in controlnet_block_res_samples
|
481 |
+
]
|
482 |
+
|
483 |
+
if USE_PEFT_BACKEND:
|
484 |
+
# remove `lora_scale` from each PEFT layer
|
485 |
+
unscale_lora_layers(self, lora_scale)
|
486 |
+
|
487 |
+
if not return_dict:
|
488 |
+
return (controlnet_block_res_samples,)
|
489 |
+
|
490 |
+
return SD3ControlNetOutput(
|
491 |
+
controlnet_block_samples=controlnet_block_res_samples
|
492 |
+
)
|
493 |
+
|
494 |
+
def invert_copy_paste(self, controlnet_block_samples):
|
495 |
+
controlnet_block_samples = controlnet_block_samples + controlnet_block_samples[::-1]
|
496 |
+
return controlnet_block_samples
|
497 |
+
|
498 |
+
class SD3MultiControlNetModel(ModelMixin):
|
499 |
+
r"""
|
500 |
+
`SD3ControlNetModel` wrapper class for Multi-SD3ControlNet
|
501 |
+
|
502 |
+
This module is a wrapper for multiple instances of the `SD3ControlNetModel`. The `forward()` API is designed to be
|
503 |
+
compatible with `SD3ControlNetModel`.
|
504 |
+
|
505 |
+
Args:
|
506 |
+
controlnets (`List[SD3ControlNetModel]`):
|
507 |
+
Provides additional conditioning to the unet during the denoising process. You must set multiple
|
508 |
+
`SD3ControlNetModel` as a list.
|
509 |
+
"""
|
510 |
+
|
511 |
+
def __init__(self, controlnets):
|
512 |
+
super().__init__()
|
513 |
+
self.nets = nn.ModuleList(controlnets)
|
514 |
+
|
515 |
+
def forward(
|
516 |
+
self,
|
517 |
+
hidden_states: torch.FloatTensor,
|
518 |
+
controlnet_cond: List[torch.tensor],
|
519 |
+
conditioning_scale: List[float],
|
520 |
+
pooled_projections: torch.FloatTensor,
|
521 |
+
encoder_hidden_states: torch.FloatTensor = None,
|
522 |
+
timestep: torch.LongTensor = None,
|
523 |
+
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
524 |
+
return_dict: bool = True,
|
525 |
+
) -> Union[SD3ControlNetOutput, Tuple]:
|
526 |
+
for i, (image, scale, controlnet) in enumerate(
|
527 |
+
zip(controlnet_cond, conditioning_scale, self.nets)
|
528 |
+
):
|
529 |
+
block_samples = controlnet(
|
530 |
+
hidden_states=hidden_states,
|
531 |
+
timestep=timestep,
|
532 |
+
encoder_hidden_states=encoder_hidden_states,
|
533 |
+
pooled_projections=pooled_projections,
|
534 |
+
controlnet_cond=image,
|
535 |
+
conditioning_scale=scale,
|
536 |
+
joint_attention_kwargs=joint_attention_kwargs,
|
537 |
+
return_dict=return_dict,
|
538 |
+
)
|
539 |
+
|
540 |
+
# merge samples
|
541 |
+
if i == 0:
|
542 |
+
control_block_samples = block_samples
|
543 |
+
else:
|
544 |
+
control_block_samples = [
|
545 |
+
control_block_sample + block_sample
|
546 |
+
for control_block_sample, block_sample in zip(
|
547 |
+
control_block_samples[0], block_samples[0]
|
548 |
+
)
|
549 |
+
]
|
550 |
+
control_block_samples = (tuple(control_block_samples),)
|
551 |
+
|
552 |
+
return control_block_samples
|