Spaces:
Running
on
Zero
Running
on
Zero
Leimingkun
commited on
Commit
•
6fe0b16
1
Parent(s):
da92c10
stylestudio
Browse files- app.py +4 -3
- ip_adapter/attention_processor.py +18 -627
- ip_adapter/ip_adapter.py +11 -487
app.py
CHANGED
@@ -85,6 +85,7 @@ def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
|
85 |
if randomize_seed:
|
86 |
seed = random.randint(0, MAX_SEED)
|
87 |
return seed
|
|
|
88 |
@spaces.GPU
|
89 |
def create_image(
|
90 |
style_image_pil,
|
@@ -95,7 +96,7 @@ def create_image(
|
|
95 |
crossModalAdaIN,
|
96 |
use_SAttn,
|
97 |
seed,
|
98 |
-
|
99 |
):
|
100 |
|
101 |
style_image = style_image_pil
|
@@ -109,7 +110,7 @@ def create_image(
|
|
109 |
with torch.no_grad():
|
110 |
images = csgo.generate(pil_style_image=style_image,
|
111 |
prompt=prompt,
|
112 |
-
negative_prompt=
|
113 |
height=1024,
|
114 |
width=1024,
|
115 |
guidance_scale=guidance_scale,
|
@@ -231,7 +232,7 @@ with block:
|
|
231 |
inputs=[style_image_pil, target, prompt, guidance_scale, seed, end_fusion],
|
232 |
fn=run_for_examples,
|
233 |
outputs=[generated_image],
|
234 |
-
cache_examples=
|
235 |
)
|
236 |
|
237 |
gr.Markdown(article)
|
|
|
85 |
if randomize_seed:
|
86 |
seed = random.randint(0, MAX_SEED)
|
87 |
return seed
|
88 |
+
|
89 |
@spaces.GPU
|
90 |
def create_image(
|
91 |
style_image_pil,
|
|
|
96 |
crossModalAdaIN,
|
97 |
use_SAttn,
|
98 |
seed,
|
99 |
+
neg_prompt="text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
|
100 |
):
|
101 |
|
102 |
style_image = style_image_pil
|
|
|
110 |
with torch.no_grad():
|
111 |
images = csgo.generate(pil_style_image=style_image,
|
112 |
prompt=prompt,
|
113 |
+
negative_prompt=neg_prompt,
|
114 |
height=1024,
|
115 |
width=1024,
|
116 |
guidance_scale=guidance_scale,
|
|
|
232 |
inputs=[style_image_pil, target, prompt, guidance_scale, seed, end_fusion],
|
233 |
fn=run_for_examples,
|
234 |
outputs=[generated_image],
|
235 |
+
cache_examples=False,
|
236 |
)
|
237 |
|
238 |
gr.Markdown(article)
|
ip_adapter/attention_processor.py
CHANGED
@@ -757,441 +757,6 @@ class CNAttnProcessor2_0:
|
|
757 |
|
758 |
return hidden_states
|
759 |
|
760 |
-
class IP_FuAd_AttnProcessor2_0(torch.nn.Module):
|
761 |
-
r"""
|
762 |
-
Attention processor for IP-Adapater for PyTorch 2.0.
|
763 |
-
Args:
|
764 |
-
hidden_size (`int`):
|
765 |
-
The hidden size of the attention layer.
|
766 |
-
cross_attention_dim (`int`):
|
767 |
-
The number of channels in the `encoder_hidden_states`.
|
768 |
-
scale (`float`, defaults to 1.0):
|
769 |
-
the weight scale of image prompt.
|
770 |
-
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
|
771 |
-
The context length of the image features.
|
772 |
-
"""
|
773 |
-
|
774 |
-
def __init__(self, hidden_size, cross_attention_dim=None, content_scale=1.0,style_scale=1.0, num_content_tokens=4,num_style_tokens=4,
|
775 |
-
skip=False,content=False, style=False, fuAttn=False, fuIPAttn=False, adainIP=False,
|
776 |
-
fuScale=0, end_fusion=0, attn_name=None):
|
777 |
-
super().__init__()
|
778 |
-
|
779 |
-
if not hasattr(F, "scaled_dot_product_attention"):
|
780 |
-
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
781 |
-
|
782 |
-
self.hidden_size = hidden_size
|
783 |
-
self.cross_attention_dim = cross_attention_dim
|
784 |
-
self.content_scale = content_scale
|
785 |
-
self.style_scale = style_scale
|
786 |
-
self.num_style_tokens = num_style_tokens
|
787 |
-
self.skip = skip
|
788 |
-
|
789 |
-
self.content = content
|
790 |
-
self.style = style
|
791 |
-
|
792 |
-
self.fuAttn = fuAttn
|
793 |
-
self.fuIPAttn = fuIPAttn
|
794 |
-
self.adainIP = adainIP
|
795 |
-
self.fuScale = fuScale
|
796 |
-
self.denoise_step = 0
|
797 |
-
self.end_fusion = end_fusion
|
798 |
-
self.name = attn_name
|
799 |
-
|
800 |
-
if self.content or self.style:
|
801 |
-
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
802 |
-
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
803 |
-
self.to_k_ip_content =None
|
804 |
-
self.to_v_ip_content =None
|
805 |
-
|
806 |
-
# def set_content_ipa(self,content_scale=1.0):
|
807 |
-
|
808 |
-
# self.to_k_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
|
809 |
-
# self.to_v_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
|
810 |
-
# self.content_scale=content_scale
|
811 |
-
# self.content =True
|
812 |
-
|
813 |
-
def reset_denoise_step(self):
|
814 |
-
if self.denoise_step == 50:
|
815 |
-
self.denoise_step = 0
|
816 |
-
# if "up_blocks.0.attentions.1.transformer_blocks.0.attn2" in self.name:
|
817 |
-
# print("attn2 reset successful")
|
818 |
-
|
819 |
-
def __call__(
|
820 |
-
self,
|
821 |
-
attn,
|
822 |
-
hidden_states,
|
823 |
-
encoder_hidden_states=None,
|
824 |
-
attention_mask=None,
|
825 |
-
temb=None,
|
826 |
-
):
|
827 |
-
self.denoise_step += 1
|
828 |
-
residual = hidden_states
|
829 |
-
|
830 |
-
if attn.spatial_norm is not None:
|
831 |
-
hidden_states = attn.spatial_norm(hidden_states, temb)
|
832 |
-
|
833 |
-
input_ndim = hidden_states.ndim
|
834 |
-
|
835 |
-
if input_ndim == 4:
|
836 |
-
batch_size, channel, height, width = hidden_states.shape
|
837 |
-
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
838 |
-
|
839 |
-
batch_size, sequence_length, _ = (
|
840 |
-
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
841 |
-
)
|
842 |
-
|
843 |
-
if attention_mask is not None:
|
844 |
-
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
845 |
-
# scaled_dot_product_attention expects attention_mask shape to be
|
846 |
-
# (batch, heads, source_length, target_length)
|
847 |
-
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
848 |
-
|
849 |
-
if attn.group_norm is not None:
|
850 |
-
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
851 |
-
|
852 |
-
query = attn.to_q(hidden_states)
|
853 |
-
|
854 |
-
if encoder_hidden_states is None:
|
855 |
-
encoder_hidden_states = hidden_states
|
856 |
-
else:
|
857 |
-
# get encoder_hidden_states, ip_hidden_states
|
858 |
-
end_pos = encoder_hidden_states.shape[1] -self.num_style_tokens
|
859 |
-
encoder_hidden_states, ip_style_hidden_states = (
|
860 |
-
encoder_hidden_states[:, :end_pos, :],
|
861 |
-
encoder_hidden_states[:, end_pos:, :],
|
862 |
-
)
|
863 |
-
if attn.norm_cross:
|
864 |
-
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
865 |
-
|
866 |
-
key = attn.to_k(encoder_hidden_states)
|
867 |
-
value = attn.to_v(encoder_hidden_states)
|
868 |
-
|
869 |
-
inner_dim = key.shape[-1]
|
870 |
-
head_dim = inner_dim // attn.heads
|
871 |
-
|
872 |
-
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
873 |
-
|
874 |
-
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
875 |
-
|
876 |
-
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
877 |
-
|
878 |
-
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
879 |
-
# TODO: add support for attn.scale when we move to Torch 2.1
|
880 |
-
# # modified the attnMap of the Stylization Image
|
881 |
-
|
882 |
-
if self.fuAttn and self.denoise_step <= self.end_fusion:
|
883 |
-
assert query.shape[0] == 4
|
884 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
885 |
-
text_attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
|
886 |
-
text_attn_probs[1] = self.fuScale*text_attn_probs[1] + (1-self.fuScale)*text_attn_probs[0]
|
887 |
-
text_attn_probs[3] = self.fuScale*text_attn_probs[3] + (1-self.fuScale)*text_attn_probs[2]
|
888 |
-
hidden_states = torch.matmul(text_attn_probs, value)
|
889 |
-
else:
|
890 |
-
hidden_states = F.scaled_dot_product_attention(
|
891 |
-
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
892 |
-
)
|
893 |
-
|
894 |
-
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
895 |
-
hidden_states = hidden_states.to(query.dtype)
|
896 |
-
|
897 |
-
raw_hidden_states = hidden_states
|
898 |
-
|
899 |
-
if not self.skip and self.style is True:
|
900 |
-
|
901 |
-
# for ip-style-adapter
|
902 |
-
ip_style_key = self.to_k_ip(ip_style_hidden_states)
|
903 |
-
ip_style_value = self.to_v_ip(ip_style_hidden_states)
|
904 |
-
|
905 |
-
ip_style_key = ip_style_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
906 |
-
ip_style_value = ip_style_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
907 |
-
|
908 |
-
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
909 |
-
# TODO: add support for attn.scale when we move to Torch 2.1
|
910 |
-
if self.fuIPAttn and self.denoise_step <= self.end_fusion:
|
911 |
-
assert query.shape[0] == 4
|
912 |
-
if "down" in self.name:
|
913 |
-
print("wrong! coding")
|
914 |
-
exit()
|
915 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
916 |
-
ip_attn_probs = torch.matmul(query, ip_style_key.transpose(-2, -1)) * scale_factor
|
917 |
-
ip_attn_probs = F.softmax(ip_attn_probs, dim=-1)
|
918 |
-
ip_attn_probs[1] = self.fuScale*ip_attn_probs[1] + (1-self.fuScale)*ip_attn_probs[0]
|
919 |
-
ip_attn_probs[3] = self.fuScale*ip_attn_probs[3] + (1-self.fuScale)*ip_attn_probs[2]
|
920 |
-
ip_style_hidden_states = torch.matmul(ip_attn_probs, ip_style_value)
|
921 |
-
else:
|
922 |
-
ip_style_hidden_states = F.scaled_dot_product_attention(
|
923 |
-
query, ip_style_key, ip_style_value, attn_mask=None, dropout_p=0.0, is_causal=False
|
924 |
-
)
|
925 |
-
|
926 |
-
ip_style_hidden_states = ip_style_hidden_states.transpose(1, 2).reshape(batch_size, -1,
|
927 |
-
attn.heads * head_dim)
|
928 |
-
ip_style_hidden_states = ip_style_hidden_states.to(query.dtype)
|
929 |
-
|
930 |
-
if not self.adainIP:
|
931 |
-
hidden_states = hidden_states + self.style_scale * ip_style_hidden_states
|
932 |
-
else:
|
933 |
-
# print("adain")
|
934 |
-
def adain(content, style):
|
935 |
-
content_mean = content.mean(dim=1, keepdim=True)
|
936 |
-
content_std = content.std(dim=1, keepdim=True)
|
937 |
-
style_mean = style.mean(dim=1, keepdim=True)
|
938 |
-
style_std = style.std(dim=1, keepdim=True)
|
939 |
-
normalized_content = (content - content_mean) / content_std
|
940 |
-
stylized_content = normalized_content * style_std + style_mean
|
941 |
-
return stylized_content
|
942 |
-
hidden_states = adain(content=hidden_states, style=ip_style_hidden_states)
|
943 |
-
|
944 |
-
if hidden_states.shape[0] == 4:
|
945 |
-
hidden_states[0] = raw_hidden_states[0]
|
946 |
-
hidden_states[2] = raw_hidden_states[2]
|
947 |
-
# hidden_states = raw_hidden_states
|
948 |
-
|
949 |
-
# linear proj
|
950 |
-
hidden_states = attn.to_out[0](hidden_states)
|
951 |
-
# dropout
|
952 |
-
hidden_states = attn.to_out[1](hidden_states)
|
953 |
-
|
954 |
-
if input_ndim == 4:
|
955 |
-
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
956 |
-
|
957 |
-
if attn.residual_connection:
|
958 |
-
hidden_states = hidden_states + residual
|
959 |
-
|
960 |
-
hidden_states = hidden_states / attn.rescale_output_factor
|
961 |
-
|
962 |
-
self.reset_denoise_step()
|
963 |
-
return hidden_states
|
964 |
-
|
965 |
-
class IP_FuAd_AttnProcessor2_0_exp(torch.nn.Module):
|
966 |
-
r"""
|
967 |
-
Attention processor for IP-Adapater for PyTorch 2.0.
|
968 |
-
Args:
|
969 |
-
hidden_size (`int`):
|
970 |
-
The hidden size of the attention layer.
|
971 |
-
cross_attention_dim (`int`):
|
972 |
-
The number of channels in the `encoder_hidden_states`.
|
973 |
-
scale (`float`, defaults to 1.0):
|
974 |
-
the weight scale of image prompt.
|
975 |
-
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
|
976 |
-
The context length of the image features.
|
977 |
-
"""
|
978 |
-
|
979 |
-
def __init__(self, hidden_size, cross_attention_dim=None, content_scale=1.0,style_scale=1.0, num_content_tokens=4,num_style_tokens=4,
|
980 |
-
skip=False,content=False, style=False, fuAttn=False, fuIPAttn=False, adainIP=False,
|
981 |
-
fuScale=0, end_fusion=0, attn_name=None, save_attn_map=False):
|
982 |
-
super().__init__()
|
983 |
-
|
984 |
-
if not hasattr(F, "scaled_dot_product_attention"):
|
985 |
-
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
986 |
-
|
987 |
-
self.hidden_size = hidden_size
|
988 |
-
self.cross_attention_dim = cross_attention_dim
|
989 |
-
self.content_scale = content_scale
|
990 |
-
self.style_scale = style_scale
|
991 |
-
self.num_style_tokens = num_style_tokens
|
992 |
-
self.skip = skip
|
993 |
-
|
994 |
-
self.content = content
|
995 |
-
self.style = style
|
996 |
-
|
997 |
-
self.fuAttn = fuAttn
|
998 |
-
self.fuIPAttn = fuIPAttn
|
999 |
-
self.adainIP = adainIP
|
1000 |
-
self.fuScale = fuScale
|
1001 |
-
self.denoise_step = 0
|
1002 |
-
self.end_fusion = end_fusion
|
1003 |
-
self.name = attn_name
|
1004 |
-
|
1005 |
-
self.save_attn_map = save_attn_map
|
1006 |
-
|
1007 |
-
if self.content or self.style:
|
1008 |
-
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
1009 |
-
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
1010 |
-
self.to_k_ip_content =None
|
1011 |
-
self.to_v_ip_content =None
|
1012 |
-
|
1013 |
-
# def set_content_ipa(self,content_scale=1.0):
|
1014 |
-
|
1015 |
-
# self.to_k_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
|
1016 |
-
# self.to_v_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
|
1017 |
-
# self.content_scale=content_scale
|
1018 |
-
# self.content =True
|
1019 |
-
def reset_denoise_step(self):
|
1020 |
-
if self.denoise_step == 50:
|
1021 |
-
self.denoise_step = 0
|
1022 |
-
# if "up_blocks.0.attentions.1.transformer_blocks.0.attn2" in self.name:
|
1023 |
-
# print("attn2 reset successful")
|
1024 |
-
|
1025 |
-
def __call__(
|
1026 |
-
self,
|
1027 |
-
attn,
|
1028 |
-
hidden_states,
|
1029 |
-
encoder_hidden_states=None,
|
1030 |
-
attention_mask=None,
|
1031 |
-
temb=None,
|
1032 |
-
):
|
1033 |
-
self.denoise_step += 1
|
1034 |
-
residual = hidden_states
|
1035 |
-
|
1036 |
-
if attn.spatial_norm is not None:
|
1037 |
-
hidden_states = attn.spatial_norm(hidden_states, temb)
|
1038 |
-
|
1039 |
-
input_ndim = hidden_states.ndim
|
1040 |
-
|
1041 |
-
if input_ndim == 4:
|
1042 |
-
batch_size, channel, height, width = hidden_states.shape
|
1043 |
-
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
1044 |
-
|
1045 |
-
batch_size, sequence_length, _ = (
|
1046 |
-
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
1047 |
-
)
|
1048 |
-
|
1049 |
-
if attention_mask is not None:
|
1050 |
-
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
1051 |
-
# scaled_dot_product_attention expects attention_mask shape to be
|
1052 |
-
# (batch, heads, source_length, target_length)
|
1053 |
-
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
1054 |
-
|
1055 |
-
if attn.group_norm is not None:
|
1056 |
-
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
1057 |
-
|
1058 |
-
query = attn.to_q(hidden_states)
|
1059 |
-
|
1060 |
-
if encoder_hidden_states is None:
|
1061 |
-
encoder_hidden_states = hidden_states
|
1062 |
-
else:
|
1063 |
-
# get encoder_hidden_states, ip_hidden_states
|
1064 |
-
end_pos = encoder_hidden_states.shape[1] - self.num_content_tokens-self.num_style_tokens
|
1065 |
-
encoder_hidden_states, ip_style_hidden_states = (
|
1066 |
-
encoder_hidden_states[:, :end_pos, :],
|
1067 |
-
encoder_hidden_states[:, end_pos:, :],
|
1068 |
-
)
|
1069 |
-
if attn.norm_cross:
|
1070 |
-
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
1071 |
-
|
1072 |
-
key = attn.to_k(encoder_hidden_states)
|
1073 |
-
value = attn.to_v(encoder_hidden_states)
|
1074 |
-
|
1075 |
-
## attention map
|
1076 |
-
if self.save_attn_map:
|
1077 |
-
attention_probs = attn.get_attention_scores(attn.head_to_batch_dim(query), attn.head_to_batch_dim(value), attention_mask)
|
1078 |
-
if attention_probs is not None:
|
1079 |
-
if not hasattr(attn, "attn_map"):
|
1080 |
-
setattr(attn, "attn_map", {})
|
1081 |
-
setattr(attn, "inference_step", 0)
|
1082 |
-
else:
|
1083 |
-
attn.inference_step += 1
|
1084 |
-
|
1085 |
-
# # maybe we need to save all the timestep
|
1086 |
-
# if attn.inference_step in self.attn_map_save_steps:
|
1087 |
-
attn.attn_map[attn.inference_step] = attention_probs.clone().cpu().detach()
|
1088 |
-
# attn.attn_map[attn.inference_step] = attention_probs.detach()
|
1089 |
-
## end of attention map
|
1090 |
-
else:
|
1091 |
-
print(f"{attn} didn't get the attention probs")
|
1092 |
-
|
1093 |
-
inner_dim = key.shape[-1]
|
1094 |
-
head_dim = inner_dim // attn.heads
|
1095 |
-
|
1096 |
-
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1097 |
-
|
1098 |
-
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1099 |
-
|
1100 |
-
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1101 |
-
|
1102 |
-
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
1103 |
-
# TODO: add support for attn.scale when we move to Torch 2.1
|
1104 |
-
# # modified the attnMap of the Stylization Image
|
1105 |
-
|
1106 |
-
if self.fuAttn and self.denoise_step <= self.end_fusion:
|
1107 |
-
assert query.shape[0] == 4
|
1108 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
1109 |
-
text_attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
|
1110 |
-
text_attn_probs[1] = self.fuScale*text_attn_probs[1] + (1-self.fuScale)*text_attn_probs[0]
|
1111 |
-
text_attn_probs[3] = self.fuScale*text_attn_probs[3] + (1-self.fuScale)*text_attn_probs[2]
|
1112 |
-
hidden_states = torch.matmul(text_attn_probs, value)
|
1113 |
-
else:
|
1114 |
-
hidden_states = F.scaled_dot_product_attention(
|
1115 |
-
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
1116 |
-
)
|
1117 |
-
|
1118 |
-
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
1119 |
-
hidden_states = hidden_states.to(query.dtype)
|
1120 |
-
|
1121 |
-
raw_hidden_states = hidden_states
|
1122 |
-
|
1123 |
-
if not self.skip and self.style is True:
|
1124 |
-
|
1125 |
-
# for ip-style-adapter
|
1126 |
-
ip_style_key = self.to_k_ip(ip_style_hidden_states)
|
1127 |
-
ip_style_value = self.to_v_ip(ip_style_hidden_states)
|
1128 |
-
|
1129 |
-
ip_style_key = ip_style_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1130 |
-
ip_style_value = ip_style_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1131 |
-
|
1132 |
-
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
1133 |
-
# TODO: add support for attn.scale when we move to Torch 2.1
|
1134 |
-
if self.fuIPAttn and self.denoise_step <= self.end_fusion:
|
1135 |
-
assert query.shape[0] == 4
|
1136 |
-
if "down" in self.name:
|
1137 |
-
print("wrong! coding")
|
1138 |
-
exit()
|
1139 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
1140 |
-
ip_attn_probs = torch.matmul(query, ip_style_key.transpose(-2, -1)) * scale_factor
|
1141 |
-
ip_attn_probs = F.softmax(ip_attn_probs, dim=-1)
|
1142 |
-
ip_attn_probs[1] = self.fuScale*ip_attn_probs[1] + (1-self.fuScale)*ip_attn_probs[0]
|
1143 |
-
ip_attn_probs[3] = self.fuScale*ip_attn_probs[3] + (1-self.fuScale)*ip_attn_probs[2]
|
1144 |
-
ip_style_hidden_states = torch.matmul(ip_attn_probs, ip_style_value)
|
1145 |
-
else:
|
1146 |
-
ip_style_hidden_states = F.scaled_dot_product_attention(
|
1147 |
-
query, ip_style_key, ip_style_value, attn_mask=None, dropout_p=0.0, is_causal=False
|
1148 |
-
)
|
1149 |
-
|
1150 |
-
ip_style_hidden_states = ip_style_hidden_states.transpose(1, 2).reshape(batch_size, -1,
|
1151 |
-
attn.heads * head_dim)
|
1152 |
-
ip_style_hidden_states = ip_style_hidden_states.to(query.dtype)
|
1153 |
-
|
1154 |
-
# if self.adainIP and self.denoise_step >= self.start_adain:
|
1155 |
-
if self.adainIP:
|
1156 |
-
# print("adain")
|
1157 |
-
# if self.denoise_step == 1 and "up_blocks.1.attentions.2.transformer_blocks.1" in self.name:
|
1158 |
-
# print("adain")
|
1159 |
-
def adain(content, style):
|
1160 |
-
content_mean = content.mean(dim=1, keepdim=True)
|
1161 |
-
content_std = content.std(dim=1, keepdim=True)
|
1162 |
-
print("exp code")
|
1163 |
-
pdb.set_trace()
|
1164 |
-
style_mean = style.mean(dim=1, keepdim=True)
|
1165 |
-
style_std = style.std(dim=1, keepdim=True)
|
1166 |
-
normalized_content = (content - content_mean) / content_std
|
1167 |
-
stylized_content = normalized_content * style_std + style_mean
|
1168 |
-
return stylized_content
|
1169 |
-
pdb.set_trace()
|
1170 |
-
hidden_states = adain(content=hidden_states, style=ip_style_hidden_states)
|
1171 |
-
else:
|
1172 |
-
hidden_states = hidden_states + self.style_scale * ip_style_hidden_states
|
1173 |
-
|
1174 |
-
if hidden_states.shape[0] == 4:
|
1175 |
-
hidden_states[0] = raw_hidden_states[0]
|
1176 |
-
hidden_states[2] = raw_hidden_states[2]
|
1177 |
-
# hidden_states = raw_hidden_states
|
1178 |
-
|
1179 |
-
# linear proj
|
1180 |
-
hidden_states = attn.to_out[0](hidden_states)
|
1181 |
-
# dropout
|
1182 |
-
hidden_states = attn.to_out[1](hidden_states)
|
1183 |
-
|
1184 |
-
if input_ndim == 4:
|
1185 |
-
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
1186 |
-
|
1187 |
-
if attn.residual_connection:
|
1188 |
-
hidden_states = hidden_states + residual
|
1189 |
-
|
1190 |
-
hidden_states = hidden_states / attn.rescale_output_factor
|
1191 |
-
|
1192 |
-
self.reset_denoise_step()
|
1193 |
-
return hidden_states
|
1194 |
-
|
1195 |
class AttnProcessor2_0_hijack(torch.nn.Module):
|
1196 |
r"""
|
1197 |
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
|
@@ -1204,131 +769,8 @@ class AttnProcessor2_0_hijack(torch.nn.Module):
|
|
1204 |
save_in_unet='down',
|
1205 |
atten_control=None,
|
1206 |
fuSAttn=False,
|
1207 |
-
fuScale=0,
|
1208 |
-
end_fusion=0,
|
1209 |
-
attn_name=None,
|
1210 |
-
):
|
1211 |
-
super().__init__()
|
1212 |
-
if not hasattr(F, "scaled_dot_product_attention"):
|
1213 |
-
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
1214 |
-
self.atten_control = atten_control
|
1215 |
-
self.save_in_unet = save_in_unet
|
1216 |
-
|
1217 |
-
self.fuSAttn = fuSAttn
|
1218 |
-
self.fuScale = fuScale
|
1219 |
-
self.denoise_step = 0
|
1220 |
-
self.end_fusion = end_fusion
|
1221 |
-
self.name = attn_name
|
1222 |
-
|
1223 |
-
def reset_denoise_step(self):
|
1224 |
-
if self.denoise_step == 50:
|
1225 |
-
self.denoise_step = 0
|
1226 |
-
# if "up_blocks.0.attentions.1.transformer_blocks.0.attn1" in self.name:
|
1227 |
-
# print("attn1 reset successful")
|
1228 |
-
|
1229 |
-
def __call__(
|
1230 |
-
self,
|
1231 |
-
attn,
|
1232 |
-
hidden_states,
|
1233 |
-
encoder_hidden_states=None,
|
1234 |
-
attention_mask=None,
|
1235 |
-
temb=None,
|
1236 |
-
):
|
1237 |
-
self.denoise_step += 1
|
1238 |
-
residual = hidden_states
|
1239 |
-
|
1240 |
-
if attn.spatial_norm is not None:
|
1241 |
-
hidden_states = attn.spatial_norm(hidden_states, temb)
|
1242 |
-
|
1243 |
-
input_ndim = hidden_states.ndim
|
1244 |
-
|
1245 |
-
if input_ndim == 4:
|
1246 |
-
batch_size, channel, height, width = hidden_states.shape
|
1247 |
-
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
1248 |
-
|
1249 |
-
batch_size, sequence_length, _ = (
|
1250 |
-
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
1251 |
-
)
|
1252 |
-
|
1253 |
-
if attention_mask is not None:
|
1254 |
-
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
1255 |
-
# scaled_dot_product_attention expects attention_mask shape to be
|
1256 |
-
# (batch, heads, source_length, target_length)
|
1257 |
-
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
1258 |
-
|
1259 |
-
if attn.group_norm is not None:
|
1260 |
-
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
1261 |
-
|
1262 |
-
query = attn.to_q(hidden_states)
|
1263 |
-
|
1264 |
-
if encoder_hidden_states is None:
|
1265 |
-
encoder_hidden_states = hidden_states
|
1266 |
-
elif attn.norm_cross:
|
1267 |
-
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
1268 |
-
|
1269 |
-
key = attn.to_k(encoder_hidden_states)
|
1270 |
-
value = attn.to_v(encoder_hidden_states)
|
1271 |
-
|
1272 |
-
inner_dim = key.shape[-1]
|
1273 |
-
head_dim = inner_dim // attn.heads
|
1274 |
-
|
1275 |
-
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1276 |
-
|
1277 |
-
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1278 |
-
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
1279 |
-
|
1280 |
-
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
1281 |
-
# TODO: add support for attn.scale when we move to Torch 2.1
|
1282 |
-
if self.fuSAttn and self.denoise_step <= self.end_fusion:
|
1283 |
-
assert query.shape[0] == 4
|
1284 |
-
if "up_blocks.1.attentions.2.transformer_blocks.1" in self.name and self.denoise_step == self.end_fusion:
|
1285 |
-
print("now: ", self.denoise_step, "end now:", self.end_fusion, "scale: ", self.fuScale)
|
1286 |
-
# pdb.set_trace()
|
1287 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
1288 |
-
attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
|
1289 |
-
attn_probs[1] = self.fuScale*attn_probs[1] + (1-self.fuScale)*attn_probs[0]
|
1290 |
-
attn_probs[3] = self.fuScale*attn_probs[3] + (1-self.fuScale)*attn_probs[2]
|
1291 |
-
hidden_states = torch.matmul(attn_probs, value)
|
1292 |
-
else:
|
1293 |
-
hidden_states = F.scaled_dot_product_attention(
|
1294 |
-
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
1295 |
-
)
|
1296 |
-
|
1297 |
-
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
1298 |
-
hidden_states = hidden_states.to(query.dtype)
|
1299 |
-
|
1300 |
-
# linear proj
|
1301 |
-
hidden_states = attn.to_out[0](hidden_states)
|
1302 |
-
# dropout
|
1303 |
-
hidden_states = attn.to_out[1](hidden_states)
|
1304 |
-
|
1305 |
-
if input_ndim == 4:
|
1306 |
-
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
1307 |
-
|
1308 |
-
if attn.residual_connection:
|
1309 |
-
hidden_states = hidden_states + residual
|
1310 |
-
|
1311 |
-
hidden_states = hidden_states / attn.rescale_output_factor
|
1312 |
-
|
1313 |
-
if self.denoise_step == 50:
|
1314 |
-
self.reset_denoise_step()
|
1315 |
-
return hidden_states
|
1316 |
-
|
1317 |
-
class AttnProcessor2_0_exp(torch.nn.Module):
|
1318 |
-
r"""
|
1319 |
-
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
|
1320 |
-
"""
|
1321 |
-
|
1322 |
-
def __init__(
|
1323 |
-
self,
|
1324 |
-
hidden_size=None,
|
1325 |
-
cross_attention_dim=None,
|
1326 |
-
save_in_unet='down',
|
1327 |
-
atten_control=None,
|
1328 |
-
fuSAttn=False,
|
1329 |
-
fuScale=0,
|
1330 |
end_fusion=0,
|
1331 |
-
|
1332 |
):
|
1333 |
super().__init__()
|
1334 |
if not hasattr(F, "scaled_dot_product_attention"):
|
@@ -1337,16 +779,10 @@ class AttnProcessor2_0_exp(torch.nn.Module):
|
|
1337 |
self.save_in_unet = save_in_unet
|
1338 |
|
1339 |
self.fuSAttn = fuSAttn
|
1340 |
-
self.fuScale = fuScale
|
1341 |
self.denoise_step = 0
|
1342 |
self.end_fusion = end_fusion
|
1343 |
-
self.
|
1344 |
|
1345 |
-
def reset_denoise_step(self):
|
1346 |
-
if self.denoise_step == 50:
|
1347 |
-
self.denoise_step = 0
|
1348 |
-
# if "up_blocks.0.attentions.1.transformer_blocks.0.attn1" in self.name:
|
1349 |
-
# print("attn1 reset successful")
|
1350 |
|
1351 |
def __call__(
|
1352 |
self,
|
@@ -1403,26 +839,10 @@ class AttnProcessor2_0_exp(torch.nn.Module):
|
|
1403 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
1404 |
if self.fuSAttn and self.denoise_step <= self.end_fusion:
|
1405 |
assert query.shape[0] == 4
|
1406 |
-
if "up_blocks.1.attentions.2.transformer_blocks.1" in self.name and self.denoise_step == self.end_fusion:
|
1407 |
-
print("now: ", self.denoise_step, "end now:", self.end_fusion, "scale: ", self.fuScale)
|
1408 |
-
# pdb.set_trace()
|
1409 |
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
1410 |
attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
|
1411 |
-
|
1412 |
-
attn_probs[
|
1413 |
-
attn_probs[3] = self.fuScale*attn_probs[3] + (1-self.fuScale)*attn_probs[2]
|
1414 |
-
print("exp code")
|
1415 |
-
pdb.set_trace()
|
1416 |
-
def adain(content, style):
|
1417 |
-
content_mean = content.mean(dim=1, keepdim=True)
|
1418 |
-
content_std = content.std(dim=1, keepdim=True)
|
1419 |
-
style_mean = style.mean(dim=1, keepdim=True)
|
1420 |
-
style_std = style.std(dim=1, keepdim=True)
|
1421 |
-
normalized_content = (content - content_mean) / content_std
|
1422 |
-
stylized_content = normalized_content * style_std + style_mean
|
1423 |
-
return stylized_content
|
1424 |
-
value[1] = adain(content=value[0], style=value[1])
|
1425 |
-
value[3] = adain(content=value[2], style=value[3])
|
1426 |
hidden_states = torch.matmul(attn_probs, value)
|
1427 |
else:
|
1428 |
hidden_states = F.scaled_dot_product_attention(
|
@@ -1445,7 +865,8 @@ class AttnProcessor2_0_exp(torch.nn.Module):
|
|
1445 |
|
1446 |
hidden_states = hidden_states / attn.rescale_output_factor
|
1447 |
|
1448 |
-
self.
|
|
|
1449 |
return hidden_states
|
1450 |
|
1451 |
class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
@@ -1463,7 +884,7 @@ class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
|
1463 |
"""
|
1464 |
|
1465 |
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,
|
1466 |
-
fuAttn=False, fuIPAttn=False, adainIP=False, end_fusion=0,
|
1467 |
super().__init__()
|
1468 |
|
1469 |
if not hasattr(F, "scaled_dot_product_attention"):
|
@@ -1478,19 +899,12 @@ class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
|
1478 |
self.fuAttn = fuAttn
|
1479 |
self.fuIPAttn = fuIPAttn
|
1480 |
self.adainIP = adainIP
|
1481 |
-
self.denoise_step =
|
1482 |
self.end_fusion = end_fusion
|
1483 |
-
self.
|
1484 |
-
self.name = attn_name
|
1485 |
|
1486 |
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
1487 |
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
1488 |
-
|
1489 |
-
def reset_denoise_step(self):
|
1490 |
-
if self.denoise_step == 50:
|
1491 |
-
self.denoise_step = 0
|
1492 |
-
# if "up_blocks.0.attentions.1.transformer_blocks.0.attn2" in self.name:
|
1493 |
-
# print("attn2 reset successful")
|
1494 |
|
1495 |
def __call__(
|
1496 |
self,
|
@@ -1552,20 +966,10 @@ class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
|
1552 |
|
1553 |
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
1554 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
1555 |
-
|
1556 |
-
|
1557 |
-
|
1558 |
-
|
1559 |
-
print("now: ", self.denoise_step, "end now:", self.end_fusion, "scale: ", self.fuScale)
|
1560 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
1561 |
-
text_attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
|
1562 |
-
text_attn_probs[1] = self.fuScale*text_attn_probs[1] + (1-self.fuScale)*text_attn_probs[0]
|
1563 |
-
text_attn_probs[3] = self.fuScale*text_attn_probs[3] + (1-self.fuScale)*text_attn_probs[2]
|
1564 |
-
hidden_states = torch.matmul(text_attn_probs, value)
|
1565 |
-
else:
|
1566 |
-
hidden_states = F.scaled_dot_product_attention(
|
1567 |
-
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
1568 |
-
)
|
1569 |
|
1570 |
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
1571 |
hidden_states = hidden_states.to(query.dtype)
|
@@ -1582,22 +986,9 @@ class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
|
1582 |
|
1583 |
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
1584 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
1585 |
-
|
1586 |
-
|
1587 |
-
|
1588 |
-
if "down" in self.name:
|
1589 |
-
print("wrong! coding")
|
1590 |
-
exit()
|
1591 |
-
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
1592 |
-
ip_attn_probs = torch.matmul(query, ip_key.transpose(-2, -1)) * scale_factor
|
1593 |
-
ip_attn_probs = F.softmax(ip_attn_probs, dim=-1)
|
1594 |
-
ip_attn_probs[1] = self.fuScale*ip_attn_probs[1] + (1-self.fuScale)*ip_attn_probs[0]
|
1595 |
-
ip_attn_probs[3] = self.fuScale*ip_attn_probs[3] + (1-self.fuScale)*ip_attn_probs[2]
|
1596 |
-
ip_hidden_states = torch.matmul(ip_attn_probs, ip_value)
|
1597 |
-
else:
|
1598 |
-
ip_hidden_states = F.scaled_dot_product_attention(
|
1599 |
-
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
|
1600 |
-
)
|
1601 |
|
1602 |
with torch.no_grad():
|
1603 |
self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
|
@@ -1639,7 +1030,7 @@ class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
|
1639 |
|
1640 |
hidden_states = hidden_states / attn.rescale_output_factor
|
1641 |
|
1642 |
-
if self.denoise_step ==
|
1643 |
-
self.
|
1644 |
|
1645 |
return hidden_states
|
|
|
757 |
|
758 |
return hidden_states
|
759 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
760 |
class AttnProcessor2_0_hijack(torch.nn.Module):
|
761 |
r"""
|
762 |
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
|
|
|
769 |
save_in_unet='down',
|
770 |
atten_control=None,
|
771 |
fuSAttn=False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
772 |
end_fusion=0,
|
773 |
+
num_inference_step=50,
|
774 |
):
|
775 |
super().__init__()
|
776 |
if not hasattr(F, "scaled_dot_product_attention"):
|
|
|
779 |
self.save_in_unet = save_in_unet
|
780 |
|
781 |
self.fuSAttn = fuSAttn
|
|
|
782 |
self.denoise_step = 0
|
783 |
self.end_fusion = end_fusion
|
784 |
+
self.num_inference_step=num_inference_step
|
785 |
|
|
|
|
|
|
|
|
|
|
|
786 |
|
787 |
def __call__(
|
788 |
self,
|
|
|
839 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
840 |
if self.fuSAttn and self.denoise_step <= self.end_fusion:
|
841 |
assert query.shape[0] == 4
|
|
|
|
|
|
|
842 |
scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
|
843 |
attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
|
844 |
+
attn_probs[1] = attn_probs[0]
|
845 |
+
attn_probs[3] = attn_probs[2]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
846 |
hidden_states = torch.matmul(attn_probs, value)
|
847 |
else:
|
848 |
hidden_states = F.scaled_dot_product_attention(
|
|
|
865 |
|
866 |
hidden_states = hidden_states / attn.rescale_output_factor
|
867 |
|
868 |
+
if self.denoise_step == self.num_inference_step:
|
869 |
+
self.denoise_step == 0
|
870 |
return hidden_states
|
871 |
|
872 |
class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
|
|
|
884 |
"""
|
885 |
|
886 |
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,
|
887 |
+
fuAttn=False, fuIPAttn=False, adainIP=False, end_fusion=0, num_inference_step=50):
|
888 |
super().__init__()
|
889 |
|
890 |
if not hasattr(F, "scaled_dot_product_attention"):
|
|
|
899 |
self.fuAttn = fuAttn
|
900 |
self.fuIPAttn = fuIPAttn
|
901 |
self.adainIP = adainIP
|
902 |
+
self.denoise_step = 0
|
903 |
self.end_fusion = end_fusion
|
904 |
+
self.num_inference_step = num_inference_step
|
|
|
905 |
|
906 |
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
907 |
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
908 |
|
909 |
def __call__(
|
910 |
self,
|
|
|
966 |
|
967 |
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
968 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
969 |
+
|
970 |
+
hidden_states = F.scaled_dot_product_attention(
|
971 |
+
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
972 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
973 |
|
974 |
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
975 |
hidden_states = hidden_states.to(query.dtype)
|
|
|
986 |
|
987 |
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
988 |
# TODO: add support for attn.scale when we move to Torch 2.1
|
989 |
+
ip_hidden_states = F.scaled_dot_product_attention(
|
990 |
+
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
|
991 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
992 |
|
993 |
with torch.no_grad():
|
994 |
self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
|
|
|
1030 |
|
1031 |
hidden_states = hidden_states / attn.rescale_output_factor
|
1032 |
|
1033 |
+
if self.denoise_step == self.num_inference_step:
|
1034 |
+
self.denoise_step == 0
|
1035 |
|
1036 |
return hidden_states
|
ip_adapter/ip_adapter.py
CHANGED
@@ -22,8 +22,6 @@ if is_torch2_available():
|
|
22 |
IPAttnProcessor2_0 as IPAttnProcessor,
|
23 |
)
|
24 |
from .attention_processor import IP_CS_AttnProcessor2_0 as IP_CS_AttnProcessor
|
25 |
-
from .attention_processor import IP_FuAd_AttnProcessor2_0 as IP_FuAd_AttnProcessor
|
26 |
-
from .attention_processor import IP_FuAd_AttnProcessor2_0_exp as IP_FuAd_AttnProcessor_exp
|
27 |
from .attention_processor import AttnProcessor2_0_exp as AttnProcessor_exp
|
28 |
from .attention_processor import AttnProcessor2_0_hijack as AttnProcessor_hijack
|
29 |
from .attention_processor import IPAttnProcessor2_0_cross_modal as IPAttnProcessor_cross_modal
|
@@ -949,7 +947,7 @@ class StyleStudio_Adapter(CSGO):
|
|
949 |
if block_name in name:
|
950 |
selected = True
|
951 |
# print(name)
|
952 |
-
attn_procs[name] =
|
953 |
hidden_size=hidden_size,
|
954 |
cross_attention_dim=cross_attention_dim,
|
955 |
style_scale=1.0,
|
@@ -963,7 +961,7 @@ class StyleStudio_Adapter(CSGO):
|
|
963 |
attn_name=name,
|
964 |
)
|
965 |
if selected is False:
|
966 |
-
attn_procs[name] =
|
967 |
hidden_size=hidden_size,
|
968 |
cross_attention_dim=cross_attention_dim,
|
969 |
num_style_tokens=self.num_style_tokens,
|
@@ -1011,7 +1009,7 @@ class StyleStudio_Adapter(CSGO):
|
|
1011 |
|
1012 |
def set_scale(self, style_scale):
|
1013 |
for attn_processor in self.pipe.unet.attn_processors.values():
|
1014 |
-
if isinstance(attn_processor,
|
1015 |
if attn_processor.style is True:
|
1016 |
attn_processor.style_scale = style_scale
|
1017 |
# print('style_scale:',style_scale)
|
@@ -1100,9 +1098,14 @@ class StyleStudio_Adapter(CSGO):
|
|
1100 |
if isinstance(attn_processor, AttnProcessor_hijack):
|
1101 |
attn_processor.fuSAttn = use_SAttn
|
1102 |
|
|
|
|
|
|
|
|
|
|
|
1103 |
def set_adain(self, use_CMA):
|
1104 |
for attn_processor in self.pipe.unet.attn_processors.values():
|
1105 |
-
if isinstance(attn_processor,
|
1106 |
attn_processor.adainIP = use_CMA
|
1107 |
|
1108 |
def generate(
|
@@ -1125,6 +1128,7 @@ class StyleStudio_Adapter(CSGO):
|
|
1125 |
self.set_endFusion(end_T = end_fusion)
|
1126 |
self.set_adain(use_CMA=cross_modal_adain)
|
1127 |
self.set_SAttn(use_SAttn=use_SAttn)
|
|
|
1128 |
|
1129 |
# self.set_scale(style_scale=style_scale)
|
1130 |
num_prompts = 1 if isinstance(pil_style_image, Image.Image) else len(pil_style_image)
|
@@ -1188,93 +1192,6 @@ class StyleStudio_Adapter(CSGO):
|
|
1188 |
).images
|
1189 |
return images
|
1190 |
|
1191 |
-
# StyleStudio_Adapter experiment code
|
1192 |
-
class StyleStudio_Adapter_exp(StyleStudio_Adapter):
|
1193 |
-
def set_ip_adapter(self):
|
1194 |
-
unet = self.pipe.unet
|
1195 |
-
attn_procs = {}
|
1196 |
-
for name in unet.attn_processors.keys():
|
1197 |
-
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
|
1198 |
-
if name.startswith("mid_block"):
|
1199 |
-
hidden_size = unet.config.block_out_channels[-1]
|
1200 |
-
elif name.startswith("up_blocks"):
|
1201 |
-
block_id = int(name[len("up_blocks.")])
|
1202 |
-
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
|
1203 |
-
elif name.startswith("down_blocks"):
|
1204 |
-
block_id = int(name[len("down_blocks.")])
|
1205 |
-
hidden_size = unet.config.block_out_channels[block_id]
|
1206 |
-
if cross_attention_dim is None:
|
1207 |
-
attn_procs[name] = AttnProcessor_exp(
|
1208 |
-
fuSAttn=self.fuSAttn,
|
1209 |
-
fuScale=self.fuScale,
|
1210 |
-
end_fusion=self.end_fusion,
|
1211 |
-
attn_name=name)
|
1212 |
-
else:
|
1213 |
-
# layername_id += 1
|
1214 |
-
selected = False
|
1215 |
-
for block_name in self.style_target_blocks:
|
1216 |
-
if block_name in name:
|
1217 |
-
selected = True
|
1218 |
-
# print(name)
|
1219 |
-
# 将所有的StyleBlock中的都改为FuAdAttn
|
1220 |
-
attn_procs[name] = IP_FuAd_AttnProcessor_exp(
|
1221 |
-
hidden_size=hidden_size,
|
1222 |
-
cross_attention_dim=cross_attention_dim,
|
1223 |
-
style_scale=1.0,
|
1224 |
-
style=True,
|
1225 |
-
num_content_tokens=self.num_content_tokens,
|
1226 |
-
num_style_tokens=self.num_style_tokens,
|
1227 |
-
fuAttn=self.fuAttn,
|
1228 |
-
fuIPAttn=self.fuIPAttn,
|
1229 |
-
adainIP=self.adainIP,
|
1230 |
-
fuScale=self.fuScale,
|
1231 |
-
end_fusion=self.end_fusion,
|
1232 |
-
attn_name=name,
|
1233 |
-
save_attn_map=self.save_attn_map,
|
1234 |
-
)
|
1235 |
-
# 没有CSGO中关于Content Control的需求 因此就将这个处理Content tokens Cross Attention 删除
|
1236 |
-
# 并且这里应该是CSGO代码中 有问题的部分 不论如何这里都会被之后的重置
|
1237 |
-
# 并且在CSGO的设计里Content Block和Style Block是没有子集的
|
1238 |
-
# selected False表明不是Style Block 关键是 Skip = True
|
1239 |
-
if selected is False:
|
1240 |
-
attn_procs[name] = IP_FuAd_AttnProcessor_exp(
|
1241 |
-
hidden_size=hidden_size,
|
1242 |
-
cross_attention_dim=cross_attention_dim,
|
1243 |
-
num_content_tokens=self.num_content_tokens,
|
1244 |
-
num_style_tokens=self.num_style_tokens,
|
1245 |
-
skip=True,
|
1246 |
-
fuAttn=self.fuAttn,
|
1247 |
-
fuIPAttn=self.fuIPAttn,
|
1248 |
-
adainIP=self.adainIP,
|
1249 |
-
fuScale=self.fuScale,
|
1250 |
-
end_fusion=self.end_fusion,
|
1251 |
-
attn_name=name,
|
1252 |
-
save_attn_map=self.save_attn_map,
|
1253 |
-
)
|
1254 |
-
# attn_procs[name] = IP_FuAd_AttnProcessor_exp(
|
1255 |
-
# hidden_size=hidden_size,
|
1256 |
-
# cross_attention_dim=cross_attention_dim,
|
1257 |
-
# num_content_tokens=self.num_content_tokens,
|
1258 |
-
# num_style_tokens=self.num_style_tokens,
|
1259 |
-
# skip=True,
|
1260 |
-
# fuAttn=self.fuAttn,
|
1261 |
-
# fuIPAttn=self.fuIPAttn,
|
1262 |
-
# )
|
1263 |
-
|
1264 |
-
attn_procs[name].to(self.device, dtype=torch.float16)
|
1265 |
-
unet.set_attn_processor(attn_procs)
|
1266 |
-
if hasattr(self.pipe, "controlnet"):
|
1267 |
-
if self.controlnet_adapter is False:
|
1268 |
-
if isinstance(self.pipe.controlnet, MultiControlNetModel):
|
1269 |
-
for controlnet in self.pipe.controlnet.nets:
|
1270 |
-
controlnet.set_attn_processor(CNAttnProcessor(
|
1271 |
-
num_tokens=self.num_content_tokens + self.num_style_tokens))
|
1272 |
-
else:
|
1273 |
-
self.pipe.controlnet.set_attn_processor(CNAttnProcessor(
|
1274 |
-
num_tokens=self.num_content_tokens + self.num_style_tokens))
|
1275 |
-
# 因为我们的代码中没有controlnet需要将Style 注入 这并不是一个I2I的任务
|
1276 |
-
# 因此我们将原本CSGO中和ControlNet中注入Style的部分给删除了
|
1277 |
-
|
1278 |
class IPAdapterXL(IPAdapter):
|
1279 |
"""SDXL"""
|
1280 |
|
@@ -1361,397 +1278,4 @@ class IPAdapterXL(IPAdapter):
|
|
1361 |
**kwargs,
|
1362 |
).images
|
1363 |
|
1364 |
-
return images
|
1365 |
-
|
1366 |
-
|
1367 |
-
class IPAdapterXL_cross_modal(IPAdapterXL):
|
1368 |
-
def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4,
|
1369 |
-
target_blocks=["block"],
|
1370 |
-
fuAttn=False,
|
1371 |
-
fuSAttn=False,
|
1372 |
-
fuIPAttn=False,
|
1373 |
-
fuScale=0,
|
1374 |
-
adainIP=False,
|
1375 |
-
end_fusion=0,
|
1376 |
-
save_attn_map=False,):
|
1377 |
-
self.fuAttn = fuAttn
|
1378 |
-
self.fuSAttn = fuSAttn
|
1379 |
-
self.fuIPAttn = fuIPAttn
|
1380 |
-
self.adainIP = adainIP
|
1381 |
-
self.fuScale = fuScale
|
1382 |
-
if self.fuSAttn:
|
1383 |
-
print(f"hijack Self AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
|
1384 |
-
if self.fuAttn:
|
1385 |
-
print(f"hijack Cross AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
|
1386 |
-
if self.fuIPAttn:
|
1387 |
-
print(f"hijack IP AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
|
1388 |
-
self.end_fusion = end_fusion
|
1389 |
-
self.save_attn_map = save_attn_map
|
1390 |
-
|
1391 |
-
self.device = device
|
1392 |
-
self.image_encoder_path = image_encoder_path
|
1393 |
-
self.ip_ckpt = ip_ckpt
|
1394 |
-
self.num_tokens = num_tokens
|
1395 |
-
self.target_blocks = target_blocks
|
1396 |
-
|
1397 |
-
self.pipe = sd_pipe.to(self.device)
|
1398 |
-
self.set_ip_adapter()
|
1399 |
-
|
1400 |
-
# load image encoder
|
1401 |
-
self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
|
1402 |
-
self.device, dtype=torch.float16
|
1403 |
-
)
|
1404 |
-
self.clip_image_processor = CLIPImageProcessor()
|
1405 |
-
# image proj model
|
1406 |
-
self.image_proj_model = self.init_proj()
|
1407 |
-
|
1408 |
-
self.load_ip_adapter()
|
1409 |
-
|
1410 |
-
def init_proj(self):
|
1411 |
-
image_proj_model = ImageProjModel(
|
1412 |
-
cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
|
1413 |
-
clip_embeddings_dim=self.image_encoder.config.projection_dim,
|
1414 |
-
clip_extra_context_tokens=self.num_tokens,
|
1415 |
-
).to(self.device, dtype=torch.float16)
|
1416 |
-
return image_proj_model
|
1417 |
-
|
1418 |
-
def set_ip_adapter(self):
|
1419 |
-
unet = self.pipe.unet
|
1420 |
-
attn_procs = {}
|
1421 |
-
for name in unet.attn_processors.keys():
|
1422 |
-
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
|
1423 |
-
if name.startswith("mid_block"):
|
1424 |
-
hidden_size = unet.config.block_out_channels[-1]
|
1425 |
-
elif name.startswith("up_blocks"):
|
1426 |
-
block_id = int(name[len("up_blocks.")])
|
1427 |
-
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
|
1428 |
-
elif name.startswith("down_blocks"):
|
1429 |
-
block_id = int(name[len("down_blocks.")])
|
1430 |
-
hidden_size = unet.config.block_out_channels[block_id]
|
1431 |
-
if cross_attention_dim is None:
|
1432 |
-
attn_procs[name] = AttnProcessor_hijack(
|
1433 |
-
fuSAttn=self.fuSAttn,
|
1434 |
-
fuScale=self.fuScale,
|
1435 |
-
end_fusion=self.end_fusion,
|
1436 |
-
attn_name=name) # Self Attention
|
1437 |
-
else: # Cross Attention
|
1438 |
-
selected = False
|
1439 |
-
for block_name in self.target_blocks:
|
1440 |
-
if block_name in name:
|
1441 |
-
selected = True
|
1442 |
-
break
|
1443 |
-
if selected:
|
1444 |
-
attn_procs[name] = IPAttnProcessor_cross_modal(
|
1445 |
-
hidden_size=hidden_size,
|
1446 |
-
cross_attention_dim=cross_attention_dim,
|
1447 |
-
scale=1.0,
|
1448 |
-
num_tokens=self.num_tokens,
|
1449 |
-
fuAttn=self.fuAttn,
|
1450 |
-
fuIPAttn=self.fuIPAttn,
|
1451 |
-
adainIP=self.adainIP,
|
1452 |
-
fuScale=self.fuScale,
|
1453 |
-
end_fusion=self.end_fusion,
|
1454 |
-
attn_name=name,
|
1455 |
-
).to(self.device, dtype=torch.float16)
|
1456 |
-
else:
|
1457 |
-
attn_procs[name] = IPAttnProcessor_cross_modal(
|
1458 |
-
hidden_size=hidden_size,
|
1459 |
-
cross_attention_dim=cross_attention_dim,
|
1460 |
-
scale=1.0,
|
1461 |
-
num_tokens=self.num_tokens,
|
1462 |
-
skip=True,
|
1463 |
-
fuAttn=self.fuAttn,
|
1464 |
-
fuIPAttn=self.fuIPAttn,
|
1465 |
-
adainIP=self.adainIP,
|
1466 |
-
fuScale=self.fuScale,
|
1467 |
-
end_fusion=self.end_fusion,
|
1468 |
-
attn_name=name,
|
1469 |
-
).to(self.device, dtype=torch.float16)
|
1470 |
-
unet.set_attn_processor(attn_procs)
|
1471 |
-
if hasattr(self.pipe, "controlnet"):
|
1472 |
-
if isinstance(self.pipe.controlnet, MultiControlNetModel):
|
1473 |
-
for controlnet in self.pipe.controlnet.nets:
|
1474 |
-
controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
|
1475 |
-
else:
|
1476 |
-
self.pipe.controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
|
1477 |
-
|
1478 |
-
def load_ip_adapter(self):
|
1479 |
-
if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
|
1480 |
-
state_dict = {"image_proj": {}, "ip_adapter": {}}
|
1481 |
-
with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
|
1482 |
-
for key in f.keys():
|
1483 |
-
if key.startswith("image_proj."):
|
1484 |
-
state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
|
1485 |
-
elif key.startswith("ip_adapter."):
|
1486 |
-
state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
|
1487 |
-
else:
|
1488 |
-
state_dict = torch.load(self.ip_ckpt, map_location="cpu")
|
1489 |
-
self.image_proj_model.load_state_dict(state_dict["image_proj"])
|
1490 |
-
ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
|
1491 |
-
ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
|
1492 |
-
|
1493 |
-
@torch.inference_mode()
|
1494 |
-
def get_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None):
|
1495 |
-
if pil_image is not None:
|
1496 |
-
if isinstance(pil_image, Image.Image):
|
1497 |
-
pil_image = [pil_image]
|
1498 |
-
clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
|
1499 |
-
clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
|
1500 |
-
else:
|
1501 |
-
clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
|
1502 |
-
|
1503 |
-
if content_prompt_embeds is not None:
|
1504 |
-
clip_image_embeds = clip_image_embeds - content_prompt_embeds
|
1505 |
-
|
1506 |
-
image_prompt_embeds = self.image_proj_model(clip_image_embeds)
|
1507 |
-
uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
|
1508 |
-
return image_prompt_embeds, uncond_image_prompt_embeds
|
1509 |
-
|
1510 |
-
def set_scale(self, scale):
|
1511 |
-
for attn_processor in self.pipe.unet.attn_processors.values():
|
1512 |
-
if isinstance(attn_processor, IPAttnProcessor_cross_modal):
|
1513 |
-
attn_processor.scale = scale
|
1514 |
-
|
1515 |
-
@torch.inference_mode()
|
1516 |
-
def get_neg_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None):
|
1517 |
-
if pil_image is not None:
|
1518 |
-
if isinstance(pil_image, Image.Image):
|
1519 |
-
pil_image = [pil_image]
|
1520 |
-
clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
|
1521 |
-
clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
|
1522 |
-
else:
|
1523 |
-
clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
|
1524 |
-
|
1525 |
-
if content_prompt_embeds is not None:
|
1526 |
-
clip_image_embeds = clip_image_embeds - content_prompt_embeds
|
1527 |
-
|
1528 |
-
neg_image_prompt_embeds = self.image_proj_model(clip_image_embeds)
|
1529 |
-
return neg_image_prompt_embeds
|
1530 |
-
|
1531 |
-
def generate(
|
1532 |
-
self,
|
1533 |
-
pil_image,
|
1534 |
-
neg_pil_image=None,
|
1535 |
-
prompt=None,
|
1536 |
-
negative_prompt=None,
|
1537 |
-
scale=1.0,
|
1538 |
-
num_samples=4,
|
1539 |
-
seed=None,
|
1540 |
-
num_inference_steps=30,
|
1541 |
-
neg_content_emb=None,
|
1542 |
-
neg_content_prompt=None,
|
1543 |
-
neg_content_scale=1.0,
|
1544 |
-
**kwargs,
|
1545 |
-
):
|
1546 |
-
self.set_scale(scale)
|
1547 |
-
|
1548 |
-
num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
|
1549 |
-
|
1550 |
-
if prompt is None:
|
1551 |
-
prompt = "best quality, high quality"
|
1552 |
-
if negative_prompt is None:
|
1553 |
-
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
|
1554 |
-
|
1555 |
-
if not isinstance(prompt, List):
|
1556 |
-
prompt = [prompt] * num_prompts
|
1557 |
-
if not isinstance(negative_prompt, List):
|
1558 |
-
negative_prompt = [negative_prompt] * num_prompts
|
1559 |
-
|
1560 |
-
if neg_content_emb is None:
|
1561 |
-
if neg_content_prompt is not None:
|
1562 |
-
with torch.inference_mode():
|
1563 |
-
(
|
1564 |
-
prompt_embeds_, # torch.Size([1, 77, 2048])
|
1565 |
-
negative_prompt_embeds_,
|
1566 |
-
pooled_prompt_embeds_, # torch.Size([1, 1280])
|
1567 |
-
negative_pooled_prompt_embeds_,
|
1568 |
-
) = self.pipe.encode_prompt(
|
1569 |
-
neg_content_prompt,
|
1570 |
-
num_images_per_prompt=num_samples,
|
1571 |
-
do_classifier_free_guidance=True,
|
1572 |
-
negative_prompt=negative_prompt,
|
1573 |
-
)
|
1574 |
-
pooled_prompt_embeds_ *= neg_content_scale
|
1575 |
-
else:
|
1576 |
-
pooled_prompt_embeds_ = neg_content_emb
|
1577 |
-
else:
|
1578 |
-
pooled_prompt_embeds_ = None
|
1579 |
-
|
1580 |
-
image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image, content_prompt_embeds=pooled_prompt_embeds_)
|
1581 |
-
|
1582 |
-
if neg_pil_image is not None:
|
1583 |
-
neg_image_prompt_embeds = self.get_neg_image_embeds(neg_pil_image)
|
1584 |
-
cos_sim_neg = F.cosine_similarity(image_prompt_embeds, neg_image_prompt_embeds.squeeze(0).unsqueeze(1), dim=-1)
|
1585 |
-
cos_sim_uncond = F.cosine_similarity(image_prompt_embeds, uncond_image_prompt_embeds.squeeze(0).unsqueeze(1), dim=-1)
|
1586 |
-
print(f"neg cos sim is: {cos_sim_neg.diagonal()}")
|
1587 |
-
print(f"uncond cos sim is: {cos_sim_uncond.diagonal()}")
|
1588 |
-
uncond_image_prompt_embeds = neg_image_prompt_embeds
|
1589 |
-
|
1590 |
-
bs_embed, seq_len, _ = image_prompt_embeds.shape
|
1591 |
-
image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
|
1592 |
-
image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
|
1593 |
-
uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
|
1594 |
-
uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
|
1595 |
-
|
1596 |
-
with torch.inference_mode():
|
1597 |
-
(
|
1598 |
-
prompt_embeds,
|
1599 |
-
negative_prompt_embeds,
|
1600 |
-
pooled_prompt_embeds,
|
1601 |
-
negative_pooled_prompt_embeds,
|
1602 |
-
) = self.pipe.encode_prompt(
|
1603 |
-
prompt,
|
1604 |
-
num_images_per_prompt=num_samples,
|
1605 |
-
do_classifier_free_guidance=True,
|
1606 |
-
negative_prompt=negative_prompt,
|
1607 |
-
)
|
1608 |
-
prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
|
1609 |
-
negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
|
1610 |
-
|
1611 |
-
# self.generator = get_generator(seed, self.device)
|
1612 |
-
|
1613 |
-
images = self.pipe(
|
1614 |
-
prompt_embeds=prompt_embeds,
|
1615 |
-
negative_prompt_embeds=negative_prompt_embeds,
|
1616 |
-
pooled_prompt_embeds=pooled_prompt_embeds,
|
1617 |
-
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
1618 |
-
num_inference_steps=num_inference_steps,
|
1619 |
-
# generator=self.generator,
|
1620 |
-
**kwargs,
|
1621 |
-
).images
|
1622 |
-
|
1623 |
-
return images
|
1624 |
-
|
1625 |
-
|
1626 |
-
class IPAdapterPlus(IPAdapter):
|
1627 |
-
"""IP-Adapter with fine-grained features"""
|
1628 |
-
|
1629 |
-
def init_proj(self):
|
1630 |
-
image_proj_model = Resampler(
|
1631 |
-
dim=self.pipe.unet.config.cross_attention_dim,
|
1632 |
-
depth=4,
|
1633 |
-
dim_head=64,
|
1634 |
-
heads=12,
|
1635 |
-
num_queries=self.num_tokens,
|
1636 |
-
embedding_dim=self.image_encoder.config.hidden_size,
|
1637 |
-
output_dim=self.pipe.unet.config.cross_attention_dim,
|
1638 |
-
ff_mult=4,
|
1639 |
-
).to(self.device, dtype=torch.float16)
|
1640 |
-
return image_proj_model
|
1641 |
-
|
1642 |
-
@torch.inference_mode()
|
1643 |
-
def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
|
1644 |
-
if isinstance(pil_image, Image.Image):
|
1645 |
-
pil_image = [pil_image]
|
1646 |
-
clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
|
1647 |
-
clip_image = clip_image.to(self.device, dtype=torch.float16)
|
1648 |
-
clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
|
1649 |
-
image_prompt_embeds = self.image_proj_model(clip_image_embeds)
|
1650 |
-
uncond_clip_image_embeds = self.image_encoder(
|
1651 |
-
torch.zeros_like(clip_image), output_hidden_states=True
|
1652 |
-
).hidden_states[-2]
|
1653 |
-
uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
|
1654 |
-
return image_prompt_embeds, uncond_image_prompt_embeds
|
1655 |
-
|
1656 |
-
|
1657 |
-
class IPAdapterFull(IPAdapterPlus):
|
1658 |
-
"""IP-Adapter with full features"""
|
1659 |
-
|
1660 |
-
def init_proj(self):
|
1661 |
-
image_proj_model = MLPProjModel(
|
1662 |
-
cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
|
1663 |
-
clip_embeddings_dim=self.image_encoder.config.hidden_size,
|
1664 |
-
).to(self.device, dtype=torch.float16)
|
1665 |
-
return image_proj_model
|
1666 |
-
|
1667 |
-
|
1668 |
-
class IPAdapterPlusXL(IPAdapter):
|
1669 |
-
"""SDXL"""
|
1670 |
-
|
1671 |
-
def init_proj(self):
|
1672 |
-
image_proj_model = Resampler(
|
1673 |
-
dim=1280,
|
1674 |
-
depth=4,
|
1675 |
-
dim_head=64,
|
1676 |
-
heads=20,
|
1677 |
-
num_queries=self.num_tokens,
|
1678 |
-
embedding_dim=self.image_encoder.config.hidden_size,
|
1679 |
-
output_dim=self.pipe.unet.config.cross_attention_dim,
|
1680 |
-
ff_mult=4,
|
1681 |
-
).to(self.device, dtype=torch.float16)
|
1682 |
-
return image_proj_model
|
1683 |
-
|
1684 |
-
@torch.inference_mode()
|
1685 |
-
def get_image_embeds(self, pil_image):
|
1686 |
-
if isinstance(pil_image, Image.Image):
|
1687 |
-
pil_image = [pil_image]
|
1688 |
-
clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
|
1689 |
-
clip_image = clip_image.to(self.device, dtype=torch.float16)
|
1690 |
-
clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
|
1691 |
-
image_prompt_embeds = self.image_proj_model(clip_image_embeds)
|
1692 |
-
uncond_clip_image_embeds = self.image_encoder(
|
1693 |
-
torch.zeros_like(clip_image), output_hidden_states=True
|
1694 |
-
).hidden_states[-2]
|
1695 |
-
uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
|
1696 |
-
return image_prompt_embeds, uncond_image_prompt_embeds
|
1697 |
-
|
1698 |
-
def generate(
|
1699 |
-
self,
|
1700 |
-
pil_image,
|
1701 |
-
prompt=None,
|
1702 |
-
negative_prompt=None,
|
1703 |
-
scale=1.0,
|
1704 |
-
num_samples=4,
|
1705 |
-
seed=None,
|
1706 |
-
num_inference_steps=30,
|
1707 |
-
**kwargs,
|
1708 |
-
):
|
1709 |
-
self.set_scale(scale)
|
1710 |
-
|
1711 |
-
num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
|
1712 |
-
|
1713 |
-
if prompt is None:
|
1714 |
-
prompt = "best quality, high quality"
|
1715 |
-
if negative_prompt is None:
|
1716 |
-
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
|
1717 |
-
|
1718 |
-
if not isinstance(prompt, List):
|
1719 |
-
prompt = [prompt] * num_prompts
|
1720 |
-
if not isinstance(negative_prompt, List):
|
1721 |
-
negative_prompt = [negative_prompt] * num_prompts
|
1722 |
-
|
1723 |
-
image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image)
|
1724 |
-
bs_embed, seq_len, _ = image_prompt_embeds.shape
|
1725 |
-
image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
|
1726 |
-
image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
|
1727 |
-
uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
|
1728 |
-
uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
|
1729 |
-
|
1730 |
-
with torch.inference_mode():
|
1731 |
-
(
|
1732 |
-
prompt_embeds,
|
1733 |
-
negative_prompt_embeds,
|
1734 |
-
pooled_prompt_embeds,
|
1735 |
-
negative_pooled_prompt_embeds,
|
1736 |
-
) = self.pipe.encode_prompt(
|
1737 |
-
prompt,
|
1738 |
-
num_images_per_prompt=num_samples,
|
1739 |
-
do_classifier_free_guidance=True,
|
1740 |
-
negative_prompt=negative_prompt,
|
1741 |
-
)
|
1742 |
-
prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
|
1743 |
-
negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
|
1744 |
-
|
1745 |
-
generator = get_generator(seed, self.device)
|
1746 |
-
|
1747 |
-
images = self.pipe(
|
1748 |
-
prompt_embeds=prompt_embeds,
|
1749 |
-
negative_prompt_embeds=negative_prompt_embeds,
|
1750 |
-
pooled_prompt_embeds=pooled_prompt_embeds,
|
1751 |
-
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
1752 |
-
num_inference_steps=num_inference_steps,
|
1753 |
-
generator=generator,
|
1754 |
-
**kwargs,
|
1755 |
-
).images
|
1756 |
-
|
1757 |
-
return images
|
|
|
22 |
IPAttnProcessor2_0 as IPAttnProcessor,
|
23 |
)
|
24 |
from .attention_processor import IP_CS_AttnProcessor2_0 as IP_CS_AttnProcessor
|
|
|
|
|
25 |
from .attention_processor import AttnProcessor2_0_exp as AttnProcessor_exp
|
26 |
from .attention_processor import AttnProcessor2_0_hijack as AttnProcessor_hijack
|
27 |
from .attention_processor import IPAttnProcessor2_0_cross_modal as IPAttnProcessor_cross_modal
|
|
|
947 |
if block_name in name:
|
948 |
selected = True
|
949 |
# print(name)
|
950 |
+
attn_procs[name] = IPAttnProcessor_cross_modal(
|
951 |
hidden_size=hidden_size,
|
952 |
cross_attention_dim=cross_attention_dim,
|
953 |
style_scale=1.0,
|
|
|
961 |
attn_name=name,
|
962 |
)
|
963 |
if selected is False:
|
964 |
+
attn_procs[name] = IPAttnProcessor_cross_modal(
|
965 |
hidden_size=hidden_size,
|
966 |
cross_attention_dim=cross_attention_dim,
|
967 |
num_style_tokens=self.num_style_tokens,
|
|
|
1009 |
|
1010 |
def set_scale(self, style_scale):
|
1011 |
for attn_processor in self.pipe.unet.attn_processors.values():
|
1012 |
+
if isinstance(attn_processor, IPAttnProcessor_cross_modal):
|
1013 |
if attn_processor.style is True:
|
1014 |
attn_processor.style_scale = style_scale
|
1015 |
# print('style_scale:',style_scale)
|
|
|
1098 |
if isinstance(attn_processor, AttnProcessor_hijack):
|
1099 |
attn_processor.fuSAttn = use_SAttn
|
1100 |
|
1101 |
+
def set_num_inference_step(self, num_T):
|
1102 |
+
for attn_processor in self.pipe.unet.attn_processors.values():
|
1103 |
+
if isinstance(attn_processor, AttnProcessor_hijack) or isinstance(attn_processor, IPAttnProcessor_cross_modal):
|
1104 |
+
attn_processor.num_inference_step = num_T
|
1105 |
+
|
1106 |
def set_adain(self, use_CMA):
|
1107 |
for attn_processor in self.pipe.unet.attn_processors.values():
|
1108 |
+
if isinstance(attn_processor, IPAttnProcessor_cross_modal):
|
1109 |
attn_processor.adainIP = use_CMA
|
1110 |
|
1111 |
def generate(
|
|
|
1128 |
self.set_endFusion(end_T = end_fusion)
|
1129 |
self.set_adain(use_CMA=cross_modal_adain)
|
1130 |
self.set_SAttn(use_SAttn=use_SAttn)
|
1131 |
+
self.set_num_inference_step(num_T=num_inference_steps)
|
1132 |
|
1133 |
# self.set_scale(style_scale=style_scale)
|
1134 |
num_prompts = 1 if isinstance(pil_style_image, Image.Image) else len(pil_style_image)
|
|
|
1192 |
).images
|
1193 |
return images
|
1194 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1195 |
class IPAdapterXL(IPAdapter):
|
1196 |
"""SDXL"""
|
1197 |
|
|
|
1278 |
**kwargs,
|
1279 |
).images
|
1280 |
|
1281 |
+
return images
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|