Updates from latest commit, looks like good fixes
Browse fileshttps://github.com/huggingface/diffusers/commit/618260409f5c0ac6b6cbf79ed21ef51ba57db1c7
- pipeline.py +609 -21
pipeline.py
CHANGED
@@ -16,6 +16,7 @@
|
|
16 |
|
17 |
import ast
|
18 |
import gc
|
|
|
19 |
import math
|
20 |
import warnings
|
21 |
from collections.abc import Iterable
|
@@ -23,16 +24,29 @@ from typing import Any, Callable, Dict, List, Optional, Union
|
|
23 |
|
24 |
import torch
|
25 |
import torch.nn.functional as F
|
|
|
26 |
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
|
27 |
|
|
|
|
|
|
|
28 |
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
29 |
from diffusers.models.attention import Attention, GatedSelfAttentionDense
|
30 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
31 |
-
from diffusers.
|
|
|
32 |
from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
|
33 |
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
|
34 |
from diffusers.schedulers import KarrasDiffusionSchedulers
|
35 |
-
from diffusers.utils import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
|
38 |
EXAMPLE_DOC_STRING = """
|
@@ -44,6 +58,7 @@ EXAMPLE_DOC_STRING = """
|
|
44 |
>>> pipe = DiffusionPipeline.from_pretrained(
|
45 |
... "longlian/lmd_plus",
|
46 |
... custom_pipeline="llm_grounded_diffusion",
|
|
|
47 |
... variant="fp16", torch_dtype=torch.float16
|
48 |
... )
|
49 |
>>> pipe.enable_model_cpu_offload()
|
@@ -96,7 +111,12 @@ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
|
96 |
|
97 |
# All keys in Stable Diffusion models: [('down', 0, 0, 0), ('down', 0, 1, 0), ('down', 1, 0, 0), ('down', 1, 1, 0), ('down', 2, 0, 0), ('down', 2, 1, 0), ('mid', 0, 0, 0), ('up', 1, 0, 0), ('up', 1, 1, 0), ('up', 1, 2, 0), ('up', 2, 0, 0), ('up', 2, 1, 0), ('up', 2, 2, 0), ('up', 3, 0, 0), ('up', 3, 1, 0), ('up', 3, 2, 0)]
|
98 |
# Note that the first up block is `UpBlock2D` rather than `CrossAttnUpBlock2D` and does not have attention. The last index is always 0 in our case since we have one `BasicTransformerBlock` in each `Transformer2DModel`.
|
99 |
-
DEFAULT_GUIDANCE_ATTN_KEYS = [
|
|
|
|
|
|
|
|
|
|
|
100 |
|
101 |
|
102 |
def convert_attn_keys(key):
|
@@ -126,7 +146,15 @@ def scale_proportion(obj_box, H, W):
|
|
126 |
|
127 |
# Adapted from the parent class `AttnProcessor2_0`
|
128 |
class AttnProcessorWithHook(AttnProcessor2_0):
|
129 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
super().__init__()
|
131 |
self.attn_processor_key = attn_processor_key
|
132 |
self.hidden_size = hidden_size
|
@@ -187,7 +215,13 @@ class AttnProcessorWithHook(AttnProcessor2_0):
|
|
187 |
|
188 |
if self.hook is not None and self.enabled:
|
189 |
# Call the hook with query, key, value, and attention maps
|
190 |
-
self.hook(
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
if self.fast_attn:
|
193 |
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
@@ -203,7 +237,12 @@ class AttnProcessorWithHook(AttnProcessor2_0):
|
|
203 |
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
204 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
205 |
hidden_states = F.scaled_dot_product_attention(
|
206 |
-
query,
|
|
|
|
|
|
|
|
|
|
|
207 |
)
|
208 |
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
209 |
hidden_states = hidden_states.to(query.dtype)
|
@@ -227,7 +266,9 @@ class AttnProcessorWithHook(AttnProcessor2_0):
|
|
227 |
return hidden_states
|
228 |
|
229 |
|
230 |
-
class LLMGroundedDiffusionPipeline(
|
|
|
|
|
231 |
r"""
|
232 |
Pipeline for layout-grounded text-to-image generation using LLM-grounded Diffusion (LMD+): https://arxiv.org/pdf/2305.13655.pdf.
|
233 |
|
@@ -258,6 +299,11 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
258 |
Whether a safety checker is needed for this pipeline.
|
259 |
"""
|
260 |
|
|
|
|
|
|
|
|
|
|
|
261 |
objects_text = "Objects: "
|
262 |
bg_prompt_text = "Background prompt: "
|
263 |
bg_prompt_text_no_trailing_space = bg_prompt_text.rstrip()
|
@@ -276,18 +322,88 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
276 |
image_encoder: CLIPVisionModelWithProjection = None,
|
277 |
requires_safety_checker: bool = True,
|
278 |
):
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
285 |
safety_checker=safety_checker,
|
286 |
feature_extractor=feature_extractor,
|
287 |
image_encoder=image_encoder,
|
288 |
-
requires_safety_checker=requires_safety_checker,
|
289 |
)
|
|
|
|
|
|
|
290 |
|
|
|
291 |
self.register_attn_hooks(unet)
|
292 |
self._saved_attn = None
|
293 |
|
@@ -474,7 +590,14 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
474 |
|
475 |
return token_map
|
476 |
|
477 |
-
def get_phrase_indices(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
478 |
for obj in phrases:
|
479 |
# Suffix the prompt with object name for attention guidance if object is not in the prompt, using "|" to separate the prompt and the suffix
|
480 |
if obj not in prompt:
|
@@ -495,7 +618,14 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
495 |
phrase_token_map_str = " ".join(phrase_token_map)
|
496 |
|
497 |
if verbose:
|
498 |
-
logger.info(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
499 |
|
500 |
# Count the number of token before substr
|
501 |
# The substring comes with a trailing space that needs to be removed by minus one in the index.
|
@@ -562,7 +692,15 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
562 |
|
563 |
return loss
|
564 |
|
565 |
-
def compute_ca_loss(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
566 |
"""
|
567 |
The `saved_attn` is supposed to be passed to `save_attn_to_dict` in `cross_attention_kwargs` prior to computing ths loss.
|
568 |
`AttnProcessor` will put attention maps into the `save_attn_to_dict`.
|
@@ -615,6 +753,7 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
615 |
latents: Optional[torch.FloatTensor] = None,
|
616 |
prompt_embeds: Optional[torch.FloatTensor] = None,
|
617 |
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
|
|
618 |
output_type: Optional[str] = "pil",
|
619 |
return_dict: bool = True,
|
620 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
@@ -672,6 +811,7 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
672 |
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
673 |
Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
|
674 |
not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
|
|
|
675 |
output_type (`str`, *optional*, defaults to `"pil"`):
|
676 |
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
|
677 |
return_dict (`bool`, *optional*, defaults to `True`):
|
@@ -734,9 +874,10 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
734 |
phrase_indices = []
|
735 |
prompt_parsed = []
|
736 |
for prompt_item in prompt:
|
737 |
-
|
738 |
-
|
739 |
-
|
|
|
740 |
phrase_indices.append(phrase_indices_parsed_item)
|
741 |
prompt_parsed.append(prompt_parsed_item)
|
742 |
prompt = prompt_parsed
|
@@ -769,6 +910,11 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
769 |
if do_classifier_free_guidance:
|
770 |
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
|
771 |
|
|
|
|
|
|
|
|
|
|
|
772 |
# 4. Prepare timesteps
|
773 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
774 |
timesteps = self.scheduler.timesteps
|
@@ -811,7 +957,10 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
811 |
if n_objs:
|
812 |
cond_boxes[:n_objs] = torch.tensor(boxes)
|
813 |
text_embeddings = torch.zeros(
|
814 |
-
max_objs,
|
|
|
|
|
|
|
815 |
)
|
816 |
if n_objs:
|
817 |
text_embeddings[:n_objs] = _text_embeddings
|
@@ -843,6 +992,9 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
843 |
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
844 |
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
845 |
|
|
|
|
|
|
|
846 |
loss_attn = torch.tensor(10000.0)
|
847 |
|
848 |
# 7. Denoising loop
|
@@ -879,6 +1031,7 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
879 |
t,
|
880 |
encoder_hidden_states=prompt_embeds,
|
881 |
cross_attention_kwargs=cross_attention_kwargs,
|
|
|
882 |
).sample
|
883 |
|
884 |
# perform guidance
|
@@ -1023,3 +1176,438 @@ class LLMGroundedDiffusionPipeline(StableDiffusionPipeline):
|
|
1023 |
self.enable_attn_hook(enabled=False)
|
1024 |
|
1025 |
return latents, loss
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
import ast
|
18 |
import gc
|
19 |
+
import inspect
|
20 |
import math
|
21 |
import warnings
|
22 |
from collections.abc import Iterable
|
|
|
24 |
|
25 |
import torch
|
26 |
import torch.nn.functional as F
|
27 |
+
from packaging import version
|
28 |
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
|
29 |
|
30 |
+
from diffusers.configuration_utils import FrozenDict
|
31 |
+
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
|
32 |
+
from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
|
33 |
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
34 |
from diffusers.models.attention import Attention, GatedSelfAttentionDense
|
35 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
36 |
+
from diffusers.models.lora import adjust_lora_scale_text_encoder
|
37 |
+
from diffusers.pipelines import DiffusionPipeline
|
38 |
from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
|
39 |
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
|
40 |
from diffusers.schedulers import KarrasDiffusionSchedulers
|
41 |
+
from diffusers.utils import (
|
42 |
+
USE_PEFT_BACKEND,
|
43 |
+
deprecate,
|
44 |
+
logging,
|
45 |
+
replace_example_docstring,
|
46 |
+
scale_lora_layers,
|
47 |
+
unscale_lora_layers,
|
48 |
+
)
|
49 |
+
from diffusers.utils.torch_utils import randn_tensor
|
50 |
|
51 |
|
52 |
EXAMPLE_DOC_STRING = """
|
|
|
58 |
>>> pipe = DiffusionPipeline.from_pretrained(
|
59 |
... "longlian/lmd_plus",
|
60 |
... custom_pipeline="llm_grounded_diffusion",
|
61 |
+
... custom_revision="main",
|
62 |
... variant="fp16", torch_dtype=torch.float16
|
63 |
... )
|
64 |
>>> pipe.enable_model_cpu_offload()
|
|
|
111 |
|
112 |
# All keys in Stable Diffusion models: [('down', 0, 0, 0), ('down', 0, 1, 0), ('down', 1, 0, 0), ('down', 1, 1, 0), ('down', 2, 0, 0), ('down', 2, 1, 0), ('mid', 0, 0, 0), ('up', 1, 0, 0), ('up', 1, 1, 0), ('up', 1, 2, 0), ('up', 2, 0, 0), ('up', 2, 1, 0), ('up', 2, 2, 0), ('up', 3, 0, 0), ('up', 3, 1, 0), ('up', 3, 2, 0)]
|
113 |
# Note that the first up block is `UpBlock2D` rather than `CrossAttnUpBlock2D` and does not have attention. The last index is always 0 in our case since we have one `BasicTransformerBlock` in each `Transformer2DModel`.
|
114 |
+
DEFAULT_GUIDANCE_ATTN_KEYS = [
|
115 |
+
("mid", 0, 0, 0),
|
116 |
+
("up", 1, 0, 0),
|
117 |
+
("up", 1, 1, 0),
|
118 |
+
("up", 1, 2, 0),
|
119 |
+
]
|
120 |
|
121 |
|
122 |
def convert_attn_keys(key):
|
|
|
146 |
|
147 |
# Adapted from the parent class `AttnProcessor2_0`
|
148 |
class AttnProcessorWithHook(AttnProcessor2_0):
|
149 |
+
def __init__(
|
150 |
+
self,
|
151 |
+
attn_processor_key,
|
152 |
+
hidden_size,
|
153 |
+
cross_attention_dim,
|
154 |
+
hook=None,
|
155 |
+
fast_attn=True,
|
156 |
+
enabled=True,
|
157 |
+
):
|
158 |
super().__init__()
|
159 |
self.attn_processor_key = attn_processor_key
|
160 |
self.hidden_size = hidden_size
|
|
|
215 |
|
216 |
if self.hook is not None and self.enabled:
|
217 |
# Call the hook with query, key, value, and attention maps
|
218 |
+
self.hook(
|
219 |
+
self.attn_processor_key,
|
220 |
+
query_batch_dim,
|
221 |
+
key_batch_dim,
|
222 |
+
value_batch_dim,
|
223 |
+
attention_probs,
|
224 |
+
)
|
225 |
|
226 |
if self.fast_attn:
|
227 |
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
|
|
237 |
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
238 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
239 |
hidden_states = F.scaled_dot_product_attention(
|
240 |
+
query,
|
241 |
+
key,
|
242 |
+
value,
|
243 |
+
attn_mask=attention_mask,
|
244 |
+
dropout_p=0.0,
|
245 |
+
is_causal=False,
|
246 |
)
|
247 |
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
248 |
hidden_states = hidden_states.to(query.dtype)
|
|
|
266 |
return hidden_states
|
267 |
|
268 |
|
269 |
+
class LLMGroundedDiffusionPipeline(
|
270 |
+
DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
|
271 |
+
):
|
272 |
r"""
|
273 |
Pipeline for layout-grounded text-to-image generation using LLM-grounded Diffusion (LMD+): https://arxiv.org/pdf/2305.13655.pdf.
|
274 |
|
|
|
299 |
Whether a safety checker is needed for this pipeline.
|
300 |
"""
|
301 |
|
302 |
+
model_cpu_offload_seq = "text_encoder->unet->vae"
|
303 |
+
_optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
|
304 |
+
_exclude_from_cpu_offload = ["safety_checker"]
|
305 |
+
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
306 |
+
|
307 |
objects_text = "Objects: "
|
308 |
bg_prompt_text = "Background prompt: "
|
309 |
bg_prompt_text_no_trailing_space = bg_prompt_text.rstrip()
|
|
|
322 |
image_encoder: CLIPVisionModelWithProjection = None,
|
323 |
requires_safety_checker: bool = True,
|
324 |
):
|
325 |
+
# This is copied from StableDiffusionPipeline, with hook initizations for LMD+.
|
326 |
+
super().__init__()
|
327 |
+
|
328 |
+
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
|
329 |
+
deprecation_message = (
|
330 |
+
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
|
331 |
+
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
|
332 |
+
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
|
333 |
+
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
|
334 |
+
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
|
335 |
+
" file"
|
336 |
+
)
|
337 |
+
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
|
338 |
+
new_config = dict(scheduler.config)
|
339 |
+
new_config["steps_offset"] = 1
|
340 |
+
scheduler._internal_dict = FrozenDict(new_config)
|
341 |
+
|
342 |
+
if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
|
343 |
+
deprecation_message = (
|
344 |
+
f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
|
345 |
+
" `clip_sample` should be set to False in the configuration file. Please make sure to update the"
|
346 |
+
" config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
|
347 |
+
" future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
|
348 |
+
" nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
|
349 |
+
)
|
350 |
+
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
|
351 |
+
new_config = dict(scheduler.config)
|
352 |
+
new_config["clip_sample"] = False
|
353 |
+
scheduler._internal_dict = FrozenDict(new_config)
|
354 |
+
|
355 |
+
if safety_checker is None and requires_safety_checker:
|
356 |
+
logger.warning(
|
357 |
+
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
|
358 |
+
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
|
359 |
+
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
|
360 |
+
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
|
361 |
+
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
|
362 |
+
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
|
363 |
+
)
|
364 |
+
|
365 |
+
if safety_checker is not None and feature_extractor is None:
|
366 |
+
raise ValueError(
|
367 |
+
"Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
|
368 |
+
" checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
|
369 |
+
)
|
370 |
+
|
371 |
+
is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
|
372 |
+
version.parse(unet.config._diffusers_version).base_version
|
373 |
+
) < version.parse("0.9.0.dev0")
|
374 |
+
is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
|
375 |
+
if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
|
376 |
+
deprecation_message = (
|
377 |
+
"The configuration file of the unet has set the default `sample_size` to smaller than"
|
378 |
+
" 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
|
379 |
+
" following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
|
380 |
+
" CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
|
381 |
+
" \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
|
382 |
+
" configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
|
383 |
+
" in the config might lead to incorrect results in future versions. If you have downloaded this"
|
384 |
+
" checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
|
385 |
+
" the `unet/config.json` file"
|
386 |
+
)
|
387 |
+
deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
|
388 |
+
new_config = dict(unet.config)
|
389 |
+
new_config["sample_size"] = 64
|
390 |
+
unet._internal_dict = FrozenDict(new_config)
|
391 |
+
|
392 |
+
self.register_modules(
|
393 |
+
vae=vae,
|
394 |
+
text_encoder=text_encoder,
|
395 |
+
tokenizer=tokenizer,
|
396 |
+
unet=unet,
|
397 |
+
scheduler=scheduler,
|
398 |
safety_checker=safety_checker,
|
399 |
feature_extractor=feature_extractor,
|
400 |
image_encoder=image_encoder,
|
|
|
401 |
)
|
402 |
+
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
403 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
404 |
+
self.register_to_config(requires_safety_checker=requires_safety_checker)
|
405 |
|
406 |
+
# Initialize the attention hooks for LLM-grounded Diffusion
|
407 |
self.register_attn_hooks(unet)
|
408 |
self._saved_attn = None
|
409 |
|
|
|
590 |
|
591 |
return token_map
|
592 |
|
593 |
+
def get_phrase_indices(
|
594 |
+
self,
|
595 |
+
prompt,
|
596 |
+
phrases,
|
597 |
+
token_map=None,
|
598 |
+
add_suffix_if_not_found=False,
|
599 |
+
verbose=False,
|
600 |
+
):
|
601 |
for obj in phrases:
|
602 |
# Suffix the prompt with object name for attention guidance if object is not in the prompt, using "|" to separate the prompt and the suffix
|
603 |
if obj not in prompt:
|
|
|
618 |
phrase_token_map_str = " ".join(phrase_token_map)
|
619 |
|
620 |
if verbose:
|
621 |
+
logger.info(
|
622 |
+
"Full str:",
|
623 |
+
token_map_str,
|
624 |
+
"Substr:",
|
625 |
+
phrase_token_map_str,
|
626 |
+
"Phrase:",
|
627 |
+
phrases,
|
628 |
+
)
|
629 |
|
630 |
# Count the number of token before substr
|
631 |
# The substring comes with a trailing space that needs to be removed by minus one in the index.
|
|
|
692 |
|
693 |
return loss
|
694 |
|
695 |
+
def compute_ca_loss(
|
696 |
+
self,
|
697 |
+
saved_attn,
|
698 |
+
bboxes,
|
699 |
+
phrase_indices,
|
700 |
+
guidance_attn_keys,
|
701 |
+
verbose=False,
|
702 |
+
**kwargs,
|
703 |
+
):
|
704 |
"""
|
705 |
The `saved_attn` is supposed to be passed to `save_attn_to_dict` in `cross_attention_kwargs` prior to computing ths loss.
|
706 |
`AttnProcessor` will put attention maps into the `save_attn_to_dict`.
|
|
|
753 |
latents: Optional[torch.FloatTensor] = None,
|
754 |
prompt_embeds: Optional[torch.FloatTensor] = None,
|
755 |
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
756 |
+
ip_adapter_image: Optional[PipelineImageInput] = None,
|
757 |
output_type: Optional[str] = "pil",
|
758 |
return_dict: bool = True,
|
759 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
|
|
811 |
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
812 |
Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
|
813 |
not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
|
814 |
+
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
|
815 |
output_type (`str`, *optional*, defaults to `"pil"`):
|
816 |
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
|
817 |
return_dict (`bool`, *optional*, defaults to `True`):
|
|
|
874 |
phrase_indices = []
|
875 |
prompt_parsed = []
|
876 |
for prompt_item in prompt:
|
877 |
+
(
|
878 |
+
phrase_indices_parsed_item,
|
879 |
+
prompt_parsed_item,
|
880 |
+
) = self.get_phrase_indices(prompt_item, add_suffix_if_not_found=True)
|
881 |
phrase_indices.append(phrase_indices_parsed_item)
|
882 |
prompt_parsed.append(prompt_parsed_item)
|
883 |
prompt = prompt_parsed
|
|
|
910 |
if do_classifier_free_guidance:
|
911 |
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
|
912 |
|
913 |
+
if ip_adapter_image is not None:
|
914 |
+
image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
|
915 |
+
if self.do_classifier_free_guidance:
|
916 |
+
image_embeds = torch.cat([negative_image_embeds, image_embeds])
|
917 |
+
|
918 |
# 4. Prepare timesteps
|
919 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
920 |
timesteps = self.scheduler.timesteps
|
|
|
957 |
if n_objs:
|
958 |
cond_boxes[:n_objs] = torch.tensor(boxes)
|
959 |
text_embeddings = torch.zeros(
|
960 |
+
max_objs,
|
961 |
+
self.unet.config.cross_attention_dim,
|
962 |
+
device=device,
|
963 |
+
dtype=self.text_encoder.dtype,
|
964 |
)
|
965 |
if n_objs:
|
966 |
text_embeddings[:n_objs] = _text_embeddings
|
|
|
992 |
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
993 |
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
994 |
|
995 |
+
# 6.1 Add image embeds for IP-Adapter
|
996 |
+
added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
|
997 |
+
|
998 |
loss_attn = torch.tensor(10000.0)
|
999 |
|
1000 |
# 7. Denoising loop
|
|
|
1031 |
t,
|
1032 |
encoder_hidden_states=prompt_embeds,
|
1033 |
cross_attention_kwargs=cross_attention_kwargs,
|
1034 |
+
added_cond_kwargs=added_cond_kwargs,
|
1035 |
).sample
|
1036 |
|
1037 |
# perform guidance
|
|
|
1176 |
self.enable_attn_hook(enabled=False)
|
1177 |
|
1178 |
return latents, loss
|
1179 |
+
|
1180 |
+
# Below are methods copied from StableDiffusionPipeline
|
1181 |
+
# The design choice of not inheriting from StableDiffusionPipeline is discussed here: https://github.com/huggingface/diffusers/pull/5993#issuecomment-1834258517
|
1182 |
+
|
1183 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
|
1184 |
+
def enable_vae_slicing(self):
|
1185 |
+
r"""
|
1186 |
+
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
1187 |
+
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
1188 |
+
"""
|
1189 |
+
self.vae.enable_slicing()
|
1190 |
+
|
1191 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
|
1192 |
+
def disable_vae_slicing(self):
|
1193 |
+
r"""
|
1194 |
+
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
|
1195 |
+
computing decoding in one step.
|
1196 |
+
"""
|
1197 |
+
self.vae.disable_slicing()
|
1198 |
+
|
1199 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
|
1200 |
+
def enable_vae_tiling(self):
|
1201 |
+
r"""
|
1202 |
+
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
1203 |
+
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
1204 |
+
processing larger images.
|
1205 |
+
"""
|
1206 |
+
self.vae.enable_tiling()
|
1207 |
+
|
1208 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
|
1209 |
+
def disable_vae_tiling(self):
|
1210 |
+
r"""
|
1211 |
+
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
|
1212 |
+
computing decoding in one step.
|
1213 |
+
"""
|
1214 |
+
self.vae.disable_tiling()
|
1215 |
+
|
1216 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
|
1217 |
+
def _encode_prompt(
|
1218 |
+
self,
|
1219 |
+
prompt,
|
1220 |
+
device,
|
1221 |
+
num_images_per_prompt,
|
1222 |
+
do_classifier_free_guidance,
|
1223 |
+
negative_prompt=None,
|
1224 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
1225 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1226 |
+
lora_scale: Optional[float] = None,
|
1227 |
+
**kwargs,
|
1228 |
+
):
|
1229 |
+
deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
|
1230 |
+
deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
|
1231 |
+
|
1232 |
+
prompt_embeds_tuple = self.encode_prompt(
|
1233 |
+
prompt=prompt,
|
1234 |
+
device=device,
|
1235 |
+
num_images_per_prompt=num_images_per_prompt,
|
1236 |
+
do_classifier_free_guidance=do_classifier_free_guidance,
|
1237 |
+
negative_prompt=negative_prompt,
|
1238 |
+
prompt_embeds=prompt_embeds,
|
1239 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
1240 |
+
lora_scale=lora_scale,
|
1241 |
+
**kwargs,
|
1242 |
+
)
|
1243 |
+
|
1244 |
+
# concatenate for backwards comp
|
1245 |
+
prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
|
1246 |
+
|
1247 |
+
return prompt_embeds
|
1248 |
+
|
1249 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
|
1250 |
+
def encode_prompt(
|
1251 |
+
self,
|
1252 |
+
prompt,
|
1253 |
+
device,
|
1254 |
+
num_images_per_prompt,
|
1255 |
+
do_classifier_free_guidance,
|
1256 |
+
negative_prompt=None,
|
1257 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
1258 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1259 |
+
lora_scale: Optional[float] = None,
|
1260 |
+
clip_skip: Optional[int] = None,
|
1261 |
+
):
|
1262 |
+
r"""
|
1263 |
+
Encodes the prompt into text encoder hidden states.
|
1264 |
+
|
1265 |
+
Args:
|
1266 |
+
prompt (`str` or `List[str]`, *optional*):
|
1267 |
+
prompt to be encoded
|
1268 |
+
device: (`torch.device`):
|
1269 |
+
torch device
|
1270 |
+
num_images_per_prompt (`int`):
|
1271 |
+
number of images that should be generated per prompt
|
1272 |
+
do_classifier_free_guidance (`bool`):
|
1273 |
+
whether to use classifier free guidance or not
|
1274 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
1275 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
1276 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
1277 |
+
less than `1`).
|
1278 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
1279 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
1280 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
1281 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
1282 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
1283 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
1284 |
+
argument.
|
1285 |
+
lora_scale (`float`, *optional*):
|
1286 |
+
A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
1287 |
+
clip_skip (`int`, *optional*):
|
1288 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
1289 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
1290 |
+
"""
|
1291 |
+
# set lora scale so that monkey patched LoRA
|
1292 |
+
# function of text encoder can correctly access it
|
1293 |
+
if lora_scale is not None and isinstance(self, LoraLoaderMixin):
|
1294 |
+
self._lora_scale = lora_scale
|
1295 |
+
|
1296 |
+
# dynamically adjust the LoRA scale
|
1297 |
+
if not USE_PEFT_BACKEND:
|
1298 |
+
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
1299 |
+
else:
|
1300 |
+
scale_lora_layers(self.text_encoder, lora_scale)
|
1301 |
+
|
1302 |
+
if prompt is not None and isinstance(prompt, str):
|
1303 |
+
batch_size = 1
|
1304 |
+
elif prompt is not None and isinstance(prompt, list):
|
1305 |
+
batch_size = len(prompt)
|
1306 |
+
else:
|
1307 |
+
batch_size = prompt_embeds.shape[0]
|
1308 |
+
|
1309 |
+
if prompt_embeds is None:
|
1310 |
+
# textual inversion: procecss multi-vector tokens if necessary
|
1311 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
1312 |
+
prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
|
1313 |
+
|
1314 |
+
text_inputs = self.tokenizer(
|
1315 |
+
prompt,
|
1316 |
+
padding="max_length",
|
1317 |
+
max_length=self.tokenizer.model_max_length,
|
1318 |
+
truncation=True,
|
1319 |
+
return_tensors="pt",
|
1320 |
+
)
|
1321 |
+
text_input_ids = text_inputs.input_ids
|
1322 |
+
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
1323 |
+
|
1324 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
1325 |
+
text_input_ids, untruncated_ids
|
1326 |
+
):
|
1327 |
+
removed_text = self.tokenizer.batch_decode(
|
1328 |
+
untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
|
1329 |
+
)
|
1330 |
+
logger.warning(
|
1331 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
1332 |
+
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
|
1333 |
+
)
|
1334 |
+
|
1335 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
1336 |
+
attention_mask = text_inputs.attention_mask.to(device)
|
1337 |
+
else:
|
1338 |
+
attention_mask = None
|
1339 |
+
|
1340 |
+
if clip_skip is None:
|
1341 |
+
prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
|
1342 |
+
prompt_embeds = prompt_embeds[0]
|
1343 |
+
else:
|
1344 |
+
prompt_embeds = self.text_encoder(
|
1345 |
+
text_input_ids.to(device),
|
1346 |
+
attention_mask=attention_mask,
|
1347 |
+
output_hidden_states=True,
|
1348 |
+
)
|
1349 |
+
# Access the `hidden_states` first, that contains a tuple of
|
1350 |
+
# all the hidden states from the encoder layers. Then index into
|
1351 |
+
# the tuple to access the hidden states from the desired layer.
|
1352 |
+
prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
|
1353 |
+
# We also need to apply the final LayerNorm here to not mess with the
|
1354 |
+
# representations. The `last_hidden_states` that we typically use for
|
1355 |
+
# obtaining the final prompt representations passes through the LayerNorm
|
1356 |
+
# layer.
|
1357 |
+
prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
|
1358 |
+
|
1359 |
+
if self.text_encoder is not None:
|
1360 |
+
prompt_embeds_dtype = self.text_encoder.dtype
|
1361 |
+
elif self.unet is not None:
|
1362 |
+
prompt_embeds_dtype = self.unet.dtype
|
1363 |
+
else:
|
1364 |
+
prompt_embeds_dtype = prompt_embeds.dtype
|
1365 |
+
|
1366 |
+
prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
|
1367 |
+
|
1368 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
1369 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
1370 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
1371 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
1372 |
+
|
1373 |
+
# get unconditional embeddings for classifier free guidance
|
1374 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None:
|
1375 |
+
uncond_tokens: List[str]
|
1376 |
+
if negative_prompt is None:
|
1377 |
+
uncond_tokens = [""] * batch_size
|
1378 |
+
elif prompt is not None and type(prompt) is not type(negative_prompt):
|
1379 |
+
raise TypeError(
|
1380 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
1381 |
+
f" {type(prompt)}."
|
1382 |
+
)
|
1383 |
+
elif isinstance(negative_prompt, str):
|
1384 |
+
uncond_tokens = [negative_prompt]
|
1385 |
+
elif batch_size != len(negative_prompt):
|
1386 |
+
raise ValueError(
|
1387 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
1388 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
1389 |
+
" the batch size of `prompt`."
|
1390 |
+
)
|
1391 |
+
else:
|
1392 |
+
uncond_tokens = negative_prompt
|
1393 |
+
|
1394 |
+
# textual inversion: procecss multi-vector tokens if necessary
|
1395 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
1396 |
+
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
|
1397 |
+
|
1398 |
+
max_length = prompt_embeds.shape[1]
|
1399 |
+
uncond_input = self.tokenizer(
|
1400 |
+
uncond_tokens,
|
1401 |
+
padding="max_length",
|
1402 |
+
max_length=max_length,
|
1403 |
+
truncation=True,
|
1404 |
+
return_tensors="pt",
|
1405 |
+
)
|
1406 |
+
|
1407 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
1408 |
+
attention_mask = uncond_input.attention_mask.to(device)
|
1409 |
+
else:
|
1410 |
+
attention_mask = None
|
1411 |
+
|
1412 |
+
negative_prompt_embeds = self.text_encoder(
|
1413 |
+
uncond_input.input_ids.to(device),
|
1414 |
+
attention_mask=attention_mask,
|
1415 |
+
)
|
1416 |
+
negative_prompt_embeds = negative_prompt_embeds[0]
|
1417 |
+
|
1418 |
+
if do_classifier_free_guidance:
|
1419 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
1420 |
+
seq_len = negative_prompt_embeds.shape[1]
|
1421 |
+
|
1422 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
|
1423 |
+
|
1424 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
1425 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
1426 |
+
|
1427 |
+
if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
|
1428 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
1429 |
+
unscale_lora_layers(self.text_encoder, lora_scale)
|
1430 |
+
|
1431 |
+
return prompt_embeds, negative_prompt_embeds
|
1432 |
+
|
1433 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
|
1434 |
+
def encode_image(self, image, device, num_images_per_prompt):
|
1435 |
+
dtype = next(self.image_encoder.parameters()).dtype
|
1436 |
+
|
1437 |
+
if not isinstance(image, torch.Tensor):
|
1438 |
+
image = self.feature_extractor(image, return_tensors="pt").pixel_values
|
1439 |
+
|
1440 |
+
image = image.to(device=device, dtype=dtype)
|
1441 |
+
image_embeds = self.image_encoder(image).image_embeds
|
1442 |
+
image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
|
1443 |
+
|
1444 |
+
uncond_image_embeds = torch.zeros_like(image_embeds)
|
1445 |
+
return image_embeds, uncond_image_embeds
|
1446 |
+
|
1447 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
|
1448 |
+
def run_safety_checker(self, image, device, dtype):
|
1449 |
+
if self.safety_checker is None:
|
1450 |
+
has_nsfw_concept = None
|
1451 |
+
else:
|
1452 |
+
if torch.is_tensor(image):
|
1453 |
+
feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
|
1454 |
+
else:
|
1455 |
+
feature_extractor_input = self.image_processor.numpy_to_pil(image)
|
1456 |
+
safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
|
1457 |
+
image, has_nsfw_concept = self.safety_checker(
|
1458 |
+
images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
|
1459 |
+
)
|
1460 |
+
return image, has_nsfw_concept
|
1461 |
+
|
1462 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
|
1463 |
+
def decode_latents(self, latents):
|
1464 |
+
deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
|
1465 |
+
deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
|
1466 |
+
|
1467 |
+
latents = 1 / self.vae.config.scaling_factor * latents
|
1468 |
+
image = self.vae.decode(latents, return_dict=False)[0]
|
1469 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
1470 |
+
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
|
1471 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
1472 |
+
return image
|
1473 |
+
|
1474 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
|
1475 |
+
def prepare_extra_step_kwargs(self, generator, eta):
|
1476 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
1477 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
1478 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
1479 |
+
# and should be between [0, 1]
|
1480 |
+
|
1481 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
1482 |
+
extra_step_kwargs = {}
|
1483 |
+
if accepts_eta:
|
1484 |
+
extra_step_kwargs["eta"] = eta
|
1485 |
+
|
1486 |
+
# check if the scheduler accepts generator
|
1487 |
+
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
1488 |
+
if accepts_generator:
|
1489 |
+
extra_step_kwargs["generator"] = generator
|
1490 |
+
return extra_step_kwargs
|
1491 |
+
|
1492 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
|
1493 |
+
def prepare_latents(
|
1494 |
+
self,
|
1495 |
+
batch_size,
|
1496 |
+
num_channels_latents,
|
1497 |
+
height,
|
1498 |
+
width,
|
1499 |
+
dtype,
|
1500 |
+
device,
|
1501 |
+
generator,
|
1502 |
+
latents=None,
|
1503 |
+
):
|
1504 |
+
shape = (
|
1505 |
+
batch_size,
|
1506 |
+
num_channels_latents,
|
1507 |
+
height // self.vae_scale_factor,
|
1508 |
+
width // self.vae_scale_factor,
|
1509 |
+
)
|
1510 |
+
if isinstance(generator, list) and len(generator) != batch_size:
|
1511 |
+
raise ValueError(
|
1512 |
+
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
1513 |
+
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
1514 |
+
)
|
1515 |
+
|
1516 |
+
if latents is None:
|
1517 |
+
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
1518 |
+
else:
|
1519 |
+
latents = latents.to(device)
|
1520 |
+
|
1521 |
+
# scale the initial noise by the standard deviation required by the scheduler
|
1522 |
+
latents = latents * self.scheduler.init_noise_sigma
|
1523 |
+
return latents
|
1524 |
+
|
1525 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
|
1526 |
+
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
|
1527 |
+
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
|
1528 |
+
|
1529 |
+
The suffixes after the scaling factors represent the stages where they are being applied.
|
1530 |
+
|
1531 |
+
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
|
1532 |
+
that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
|
1533 |
+
|
1534 |
+
Args:
|
1535 |
+
s1 (`float`):
|
1536 |
+
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
|
1537 |
+
mitigate "oversmoothing effect" in the enhanced denoising process.
|
1538 |
+
s2 (`float`):
|
1539 |
+
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
|
1540 |
+
mitigate "oversmoothing effect" in the enhanced denoising process.
|
1541 |
+
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
|
1542 |
+
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
|
1543 |
+
"""
|
1544 |
+
if not hasattr(self, "unet"):
|
1545 |
+
raise ValueError("The pipeline must have `unet` for using FreeU.")
|
1546 |
+
self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
|
1547 |
+
|
1548 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
|
1549 |
+
def disable_freeu(self):
|
1550 |
+
"""Disables the FreeU mechanism if enabled."""
|
1551 |
+
self.unet.disable_freeu()
|
1552 |
+
|
1553 |
+
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
|
1554 |
+
def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
|
1555 |
+
"""
|
1556 |
+
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
|
1557 |
+
|
1558 |
+
Args:
|
1559 |
+
timesteps (`torch.Tensor`):
|
1560 |
+
generate embedding vectors at these timesteps
|
1561 |
+
embedding_dim (`int`, *optional*, defaults to 512):
|
1562 |
+
dimension of the embeddings to generate
|
1563 |
+
dtype:
|
1564 |
+
data type of the generated embeddings
|
1565 |
+
|
1566 |
+
Returns:
|
1567 |
+
`torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
|
1568 |
+
"""
|
1569 |
+
assert len(w.shape) == 1
|
1570 |
+
w = w * 1000.0
|
1571 |
+
|
1572 |
+
half_dim = embedding_dim // 2
|
1573 |
+
emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
|
1574 |
+
emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
|
1575 |
+
emb = w.to(dtype)[:, None] * emb[None, :]
|
1576 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
|
1577 |
+
if embedding_dim % 2 == 1: # zero pad
|
1578 |
+
emb = torch.nn.functional.pad(emb, (0, 1))
|
1579 |
+
assert emb.shape == (w.shape[0], embedding_dim)
|
1580 |
+
return emb
|
1581 |
+
|
1582 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_scale
|
1583 |
+
@property
|
1584 |
+
def guidance_scale(self):
|
1585 |
+
return self._guidance_scale
|
1586 |
+
|
1587 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.guidance_rescale
|
1588 |
+
@property
|
1589 |
+
def guidance_rescale(self):
|
1590 |
+
return self._guidance_rescale
|
1591 |
+
|
1592 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.clip_skip
|
1593 |
+
@property
|
1594 |
+
def clip_skip(self):
|
1595 |
+
return self._clip_skip
|
1596 |
+
|
1597 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
1598 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
1599 |
+
# corresponds to doing no classifier free guidance.
|
1600 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.do_classifier_free_guidance
|
1601 |
+
@property
|
1602 |
+
def do_classifier_free_guidance(self):
|
1603 |
+
return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
|
1604 |
+
|
1605 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.cross_attention_kwargs
|
1606 |
+
@property
|
1607 |
+
def cross_attention_kwargs(self):
|
1608 |
+
return self._cross_attention_kwargs
|
1609 |
+
|
1610 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.num_timesteps
|
1611 |
+
@property
|
1612 |
+
def num_timesteps(self):
|
1613 |
+
return self._num_timesteps
|