Spaces:
Runtime error
Runtime error
Upload 5 files
Browse files- models/attn_processor.py +164 -0
- models/norm_layer.py +46 -0
- models/resampler.py +365 -0
- models/utils.py +139 -0
- pipeline.py +550 -0
models/attn_processor.py
ADDED
@@ -0,0 +1,164 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional
|
2 |
+
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch
|
5 |
+
import torch.nn.functional as F
|
6 |
+
from diffusers.models.embeddings import apply_rotary_emb
|
7 |
+
from einops import rearrange
|
8 |
+
|
9 |
+
from .norm_layer import RMSNorm
|
10 |
+
|
11 |
+
|
12 |
+
class FluxIPAttnProcessor(nn.Module):
|
13 |
+
"""Attention processor used typically in processing the SD3-like self-attention projections."""
|
14 |
+
|
15 |
+
def __init__(
|
16 |
+
self,
|
17 |
+
hidden_size=None,
|
18 |
+
ip_hidden_states_dim=None,
|
19 |
+
):
|
20 |
+
super().__init__()
|
21 |
+
self.norm_ip_q = RMSNorm(128, eps=1e-6)
|
22 |
+
self.to_k_ip = nn.Linear(ip_hidden_states_dim, hidden_size)
|
23 |
+
self.norm_ip_k = RMSNorm(128, eps=1e-6)
|
24 |
+
self.to_v_ip = nn.Linear(ip_hidden_states_dim, hidden_size)
|
25 |
+
|
26 |
+
|
27 |
+
def __call__(
|
28 |
+
self,
|
29 |
+
attn,
|
30 |
+
hidden_states: torch.FloatTensor,
|
31 |
+
encoder_hidden_states: torch.FloatTensor = None,
|
32 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
33 |
+
image_rotary_emb: Optional[torch.Tensor] = None,
|
34 |
+
emb_dict={},
|
35 |
+
subject_emb_dict={},
|
36 |
+
*args,
|
37 |
+
**kwargs,
|
38 |
+
) -> torch.FloatTensor:
|
39 |
+
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
40 |
+
|
41 |
+
# `sample` projections.
|
42 |
+
query = attn.to_q(hidden_states)
|
43 |
+
key = attn.to_k(hidden_states)
|
44 |
+
value = attn.to_v(hidden_states)
|
45 |
+
|
46 |
+
# IPadapter
|
47 |
+
ip_hidden_states = self._get_ip_hidden_states(
|
48 |
+
attn,
|
49 |
+
query if encoder_hidden_states is not None else query[:, emb_dict['length_encoder_hidden_states']:],
|
50 |
+
subject_emb_dict.get('ip_hidden_states', None)
|
51 |
+
)
|
52 |
+
|
53 |
+
inner_dim = key.shape[-1]
|
54 |
+
head_dim = inner_dim // attn.heads
|
55 |
+
|
56 |
+
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
57 |
+
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
58 |
+
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
59 |
+
|
60 |
+
if attn.norm_q is not None:
|
61 |
+
query = attn.norm_q(query)
|
62 |
+
if attn.norm_k is not None:
|
63 |
+
key = attn.norm_k(key)
|
64 |
+
|
65 |
+
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
|
66 |
+
if encoder_hidden_states is not None:
|
67 |
+
# `context` projections.
|
68 |
+
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
|
69 |
+
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
|
70 |
+
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
|
71 |
+
|
72 |
+
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
|
73 |
+
batch_size, -1, attn.heads, head_dim
|
74 |
+
).transpose(1, 2)
|
75 |
+
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
|
76 |
+
batch_size, -1, attn.heads, head_dim
|
77 |
+
).transpose(1, 2)
|
78 |
+
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
|
79 |
+
batch_size, -1, attn.heads, head_dim
|
80 |
+
).transpose(1, 2)
|
81 |
+
|
82 |
+
if attn.norm_added_q is not None:
|
83 |
+
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
|
84 |
+
if attn.norm_added_k is not None:
|
85 |
+
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
|
86 |
+
|
87 |
+
# attention
|
88 |
+
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
|
89 |
+
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
|
90 |
+
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
|
91 |
+
|
92 |
+
if image_rotary_emb is not None:
|
93 |
+
query = apply_rotary_emb(query, image_rotary_emb)
|
94 |
+
key = apply_rotary_emb(key, image_rotary_emb)
|
95 |
+
|
96 |
+
hidden_states = F.scaled_dot_product_attention(
|
97 |
+
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
98 |
+
)
|
99 |
+
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
100 |
+
hidden_states = hidden_states.to(query.dtype)
|
101 |
+
|
102 |
+
|
103 |
+
if encoder_hidden_states is not None:
|
104 |
+
encoder_hidden_states, hidden_states = (
|
105 |
+
hidden_states[:, : encoder_hidden_states.shape[1]],
|
106 |
+
hidden_states[:, encoder_hidden_states.shape[1] :],
|
107 |
+
)
|
108 |
+
|
109 |
+
if ip_hidden_states is not None:
|
110 |
+
hidden_states = hidden_states + ip_hidden_states * subject_emb_dict.get('scale', 1.0)
|
111 |
+
|
112 |
+
# linear proj
|
113 |
+
hidden_states = attn.to_out[0](hidden_states)
|
114 |
+
# dropout
|
115 |
+
hidden_states = attn.to_out[1](hidden_states)
|
116 |
+
|
117 |
+
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
|
118 |
+
|
119 |
+
return hidden_states, encoder_hidden_states
|
120 |
+
else:
|
121 |
+
|
122 |
+
if ip_hidden_states is not None:
|
123 |
+
hidden_states[:, emb_dict['length_encoder_hidden_states']:] = \
|
124 |
+
hidden_states[:, emb_dict['length_encoder_hidden_states']:] + \
|
125 |
+
ip_hidden_states * subject_emb_dict.get('scale', 1.0)
|
126 |
+
|
127 |
+
return hidden_states
|
128 |
+
|
129 |
+
|
130 |
+
def _scaled_dot_product_attention(self, query, key, value, attention_mask=None, heads=None):
|
131 |
+
query = rearrange(query, '(b h) l c -> b h l c', h=heads)
|
132 |
+
key = rearrange(key, '(b h) l c -> b h l c', h=heads)
|
133 |
+
value = rearrange(value, '(b h) l c -> b h l c', h=heads)
|
134 |
+
hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False, attn_mask=None)
|
135 |
+
hidden_states = rearrange(hidden_states, 'b h l c -> (b h) l c', h=heads)
|
136 |
+
hidden_states = hidden_states.to(query)
|
137 |
+
return hidden_states
|
138 |
+
|
139 |
+
|
140 |
+
def _get_ip_hidden_states(
|
141 |
+
self,
|
142 |
+
attn,
|
143 |
+
img_query,
|
144 |
+
ip_hidden_states,
|
145 |
+
):
|
146 |
+
if ip_hidden_states is None:
|
147 |
+
return None
|
148 |
+
|
149 |
+
if not hasattr(self, 'to_k_ip') or not hasattr(self, 'to_v_ip'):
|
150 |
+
return None
|
151 |
+
|
152 |
+
ip_query = self.norm_ip_q(rearrange(img_query, 'b l (h d) -> b h l d', h=attn.heads))
|
153 |
+
ip_query = rearrange(ip_query, 'b h l d -> (b h) l d')
|
154 |
+
ip_key = self.to_k_ip(ip_hidden_states)
|
155 |
+
ip_key = self.norm_ip_k(rearrange(ip_key, 'b l (h d) -> b h l d', h=attn.heads))
|
156 |
+
ip_key = rearrange(ip_key, 'b h l d -> (b h) l d')
|
157 |
+
ip_value = self.to_v_ip(ip_hidden_states)
|
158 |
+
ip_value = attn.head_to_batch_dim(ip_value)
|
159 |
+
ip_hidden_states = self._scaled_dot_product_attention(
|
160 |
+
ip_query.to(ip_value.dtype), ip_key.to(ip_value.dtype), ip_value, None, attn.heads)
|
161 |
+
ip_hidden_states = ip_hidden_states.to(img_query.dtype)
|
162 |
+
ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
|
163 |
+
return ip_hidden_states
|
164 |
+
|
models/norm_layer.py
ADDED
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch
|
3 |
+
|
4 |
+
class RMSNorm(nn.Module):
|
5 |
+
def __init__(self, d, p=-1., eps=1e-8, bias=False):
|
6 |
+
"""
|
7 |
+
Root Mean Square Layer Normalization
|
8 |
+
:param d: model size
|
9 |
+
:param p: partial RMSNorm, valid value [0, 1], default -1.0 (disabled)
|
10 |
+
:param eps: epsilon value, default 1e-8
|
11 |
+
:param bias: whether use bias term for RMSNorm, disabled by
|
12 |
+
default because RMSNorm doesn't enforce re-centering invariance.
|
13 |
+
"""
|
14 |
+
super(RMSNorm, self).__init__()
|
15 |
+
|
16 |
+
self.eps = eps
|
17 |
+
self.d = d
|
18 |
+
self.p = p
|
19 |
+
self.bias = bias
|
20 |
+
|
21 |
+
self.scale = nn.Parameter(torch.ones(d))
|
22 |
+
self.register_parameter("scale", self.scale)
|
23 |
+
|
24 |
+
if self.bias:
|
25 |
+
self.offset = nn.Parameter(torch.zeros(d))
|
26 |
+
self.register_parameter("offset", self.offset)
|
27 |
+
|
28 |
+
def forward(self, x):
|
29 |
+
if self.p < 0. or self.p > 1.:
|
30 |
+
norm_x = x.norm(2, dim=-1, keepdim=True)
|
31 |
+
d_x = self.d
|
32 |
+
else:
|
33 |
+
partial_size = int(self.d * self.p)
|
34 |
+
partial_x, _ = torch.split(x, [partial_size, self.d - partial_size], dim=-1)
|
35 |
+
|
36 |
+
norm_x = partial_x.norm(2, dim=-1, keepdim=True)
|
37 |
+
d_x = partial_size
|
38 |
+
|
39 |
+
rms_x = norm_x * d_x ** (-1. / 2)
|
40 |
+
x_normed = x / (rms_x + self.eps)
|
41 |
+
|
42 |
+
if self.bias:
|
43 |
+
return self.scale * x_normed + self.offset
|
44 |
+
|
45 |
+
return self.scale * x_normed
|
46 |
+
|
models/resampler.py
ADDED
@@ -0,0 +1,365 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch
|
3 |
+
import math
|
4 |
+
|
5 |
+
from diffusers.models.transformers.transformer_2d import BasicTransformerBlock
|
6 |
+
from diffusers.models.embeddings import Timesteps, TimestepEmbedding
|
7 |
+
from timm.models.vision_transformer import Mlp
|
8 |
+
|
9 |
+
from .norm_layer import RMSNorm
|
10 |
+
|
11 |
+
|
12 |
+
# FFN
|
13 |
+
def FeedForward(dim, mult=4):
|
14 |
+
inner_dim = int(dim * mult)
|
15 |
+
return nn.Sequential(
|
16 |
+
nn.LayerNorm(dim),
|
17 |
+
nn.Linear(dim, inner_dim, bias=False),
|
18 |
+
nn.GELU(),
|
19 |
+
nn.Linear(inner_dim, dim, bias=False),
|
20 |
+
)
|
21 |
+
|
22 |
+
|
23 |
+
def reshape_tensor(x, heads):
|
24 |
+
bs, length, width = x.shape
|
25 |
+
#(bs, length, width) --> (bs, length, n_heads, dim_per_head)
|
26 |
+
x = x.view(bs, length, heads, -1)
|
27 |
+
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
|
28 |
+
x = x.transpose(1, 2)
|
29 |
+
# (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
|
30 |
+
x = x.reshape(bs, heads, length, -1)
|
31 |
+
return x
|
32 |
+
|
33 |
+
|
34 |
+
class PerceiverAttention(nn.Module):
|
35 |
+
def __init__(self, *, dim, dim_head=64, heads=8):
|
36 |
+
super().__init__()
|
37 |
+
self.scale = dim_head**-0.5
|
38 |
+
self.dim_head = dim_head
|
39 |
+
self.heads = heads
|
40 |
+
inner_dim = dim_head * heads
|
41 |
+
|
42 |
+
self.norm1 = nn.LayerNorm(dim)
|
43 |
+
self.norm2 = nn.LayerNorm(dim)
|
44 |
+
|
45 |
+
self.to_q = nn.Linear(dim, inner_dim, bias=False)
|
46 |
+
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
|
47 |
+
self.to_out = nn.Linear(inner_dim, dim, bias=False)
|
48 |
+
|
49 |
+
|
50 |
+
def forward(self, x, latents, shift=None, scale=None):
|
51 |
+
"""
|
52 |
+
Args:
|
53 |
+
x (torch.Tensor): image features
|
54 |
+
shape (b, n1, D)
|
55 |
+
latent (torch.Tensor): latent features
|
56 |
+
shape (b, n2, D)
|
57 |
+
"""
|
58 |
+
x = self.norm1(x)
|
59 |
+
latents = self.norm2(latents)
|
60 |
+
|
61 |
+
if shift is not None and scale is not None:
|
62 |
+
latents = latents * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
63 |
+
|
64 |
+
b, l, _ = latents.shape
|
65 |
+
|
66 |
+
q = self.to_q(latents)
|
67 |
+
kv_input = torch.cat((x, latents), dim=-2)
|
68 |
+
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
|
69 |
+
|
70 |
+
q = reshape_tensor(q, self.heads)
|
71 |
+
k = reshape_tensor(k, self.heads)
|
72 |
+
v = reshape_tensor(v, self.heads)
|
73 |
+
|
74 |
+
# attention
|
75 |
+
scale = 1 / math.sqrt(math.sqrt(self.dim_head))
|
76 |
+
weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
|
77 |
+
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
|
78 |
+
out = weight @ v
|
79 |
+
|
80 |
+
out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
|
81 |
+
|
82 |
+
return self.to_out(out)
|
83 |
+
|
84 |
+
|
85 |
+
class ReshapeExpandToken(nn.Module):
|
86 |
+
def __init__(self, expand_token, token_dim):
|
87 |
+
super().__init__()
|
88 |
+
self.expand_token = expand_token
|
89 |
+
self.token_dim = token_dim
|
90 |
+
|
91 |
+
def forward(self, x):
|
92 |
+
x = x.reshape(-1, self.expand_token, self.token_dim)
|
93 |
+
return x
|
94 |
+
|
95 |
+
|
96 |
+
class TimeResampler(nn.Module):
|
97 |
+
def __init__(
|
98 |
+
self,
|
99 |
+
dim=1024,
|
100 |
+
depth=8,
|
101 |
+
dim_head=64,
|
102 |
+
heads=16,
|
103 |
+
num_queries=8,
|
104 |
+
embedding_dim=768,
|
105 |
+
output_dim=1024,
|
106 |
+
ff_mult=4,
|
107 |
+
timestep_in_dim=320,
|
108 |
+
timestep_flip_sin_to_cos=True,
|
109 |
+
timestep_freq_shift=0,
|
110 |
+
expand_token=None,
|
111 |
+
extra_dim=None,
|
112 |
+
):
|
113 |
+
super().__init__()
|
114 |
+
|
115 |
+
self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
|
116 |
+
|
117 |
+
self.expand_token = expand_token is not None
|
118 |
+
if expand_token:
|
119 |
+
self.expand_proj = torch.nn.Sequential(
|
120 |
+
torch.nn.Linear(embedding_dim, embedding_dim * 2),
|
121 |
+
torch.nn.GELU(),
|
122 |
+
torch.nn.Linear(embedding_dim * 2, embedding_dim * expand_token),
|
123 |
+
ReshapeExpandToken(expand_token, embedding_dim),
|
124 |
+
RMSNorm(embedding_dim, eps=1e-8),
|
125 |
+
)
|
126 |
+
|
127 |
+
self.proj_in = nn.Linear(embedding_dim, dim)
|
128 |
+
|
129 |
+
self.extra_feature = extra_dim is not None
|
130 |
+
if self.extra_feature:
|
131 |
+
self.proj_in_norm = RMSNorm(dim, eps=1e-8)
|
132 |
+
self.extra_proj_in = torch.nn.Sequential(
|
133 |
+
nn.Linear(extra_dim, dim),
|
134 |
+
RMSNorm(dim, eps=1e-8),
|
135 |
+
)
|
136 |
+
|
137 |
+
self.proj_out = nn.Linear(dim, output_dim)
|
138 |
+
self.norm_out = nn.LayerNorm(output_dim)
|
139 |
+
|
140 |
+
self.layers = nn.ModuleList([])
|
141 |
+
for _ in range(depth):
|
142 |
+
self.layers.append(
|
143 |
+
nn.ModuleList(
|
144 |
+
[
|
145 |
+
# msa
|
146 |
+
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
|
147 |
+
# ff
|
148 |
+
FeedForward(dim=dim, mult=ff_mult),
|
149 |
+
# adaLN
|
150 |
+
nn.Sequential(nn.SiLU(), nn.Linear(dim, 4 * dim, bias=True))
|
151 |
+
]
|
152 |
+
)
|
153 |
+
)
|
154 |
+
|
155 |
+
# time
|
156 |
+
self.time_proj = Timesteps(timestep_in_dim, timestep_flip_sin_to_cos, timestep_freq_shift)
|
157 |
+
self.time_embedding = TimestepEmbedding(timestep_in_dim, dim, act_fn="silu")
|
158 |
+
|
159 |
+
|
160 |
+
def forward(self, x, timestep, need_temb=False, extra_feature=None):
|
161 |
+
timestep_emb = self.embedding_time(x, timestep) # bs, dim
|
162 |
+
|
163 |
+
latents = self.latents.repeat(x.size(0), 1, 1)
|
164 |
+
|
165 |
+
if self.expand_token:
|
166 |
+
x = self.expand_proj(x)
|
167 |
+
|
168 |
+
x = self.proj_in(x)
|
169 |
+
|
170 |
+
if self.extra_feature:
|
171 |
+
extra_feature = self.extra_proj_in(extra_feature)
|
172 |
+
x = self.proj_in_norm(x)
|
173 |
+
x = torch.cat([x, extra_feature], dim=1)
|
174 |
+
|
175 |
+
x = x + timestep_emb[:, None]
|
176 |
+
|
177 |
+
for attn, ff, adaLN_modulation in self.layers:
|
178 |
+
shift_msa, scale_msa, shift_mlp, scale_mlp = adaLN_modulation(timestep_emb).chunk(4, dim=1)
|
179 |
+
latents = attn(x, latents, shift_msa, scale_msa) + latents
|
180 |
+
|
181 |
+
res = latents
|
182 |
+
for idx_ff in range(len(ff)):
|
183 |
+
layer_ff = ff[idx_ff]
|
184 |
+
latents = layer_ff(latents)
|
185 |
+
if idx_ff == 0 and isinstance(layer_ff, nn.LayerNorm): # adaLN
|
186 |
+
latents = latents * (1 + scale_mlp.unsqueeze(1)) + shift_mlp.unsqueeze(1)
|
187 |
+
latents = latents + res
|
188 |
+
|
189 |
+
# latents = ff(latents) + latents
|
190 |
+
|
191 |
+
latents = self.proj_out(latents)
|
192 |
+
latents = self.norm_out(latents)
|
193 |
+
|
194 |
+
if need_temb:
|
195 |
+
return latents, timestep_emb
|
196 |
+
else:
|
197 |
+
return latents
|
198 |
+
|
199 |
+
|
200 |
+
def embedding_time(self, sample, timestep):
|
201 |
+
|
202 |
+
# 1. time
|
203 |
+
timesteps = timestep
|
204 |
+
if not torch.is_tensor(timesteps):
|
205 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
206 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
207 |
+
is_mps = sample.device.type == "mps"
|
208 |
+
if isinstance(timestep, float):
|
209 |
+
dtype = torch.float32 if is_mps else torch.float64
|
210 |
+
else:
|
211 |
+
dtype = torch.int32 if is_mps else torch.int64
|
212 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
213 |
+
elif len(timesteps.shape) == 0:
|
214 |
+
timesteps = timesteps[None].to(sample.device)
|
215 |
+
|
216 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
217 |
+
timesteps = timesteps.expand(sample.shape[0])
|
218 |
+
|
219 |
+
t_emb = self.time_proj(timesteps)
|
220 |
+
|
221 |
+
# timesteps does not contain any weights and will always return f32 tensors
|
222 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
223 |
+
# there might be better ways to encapsulate this.
|
224 |
+
t_emb = t_emb.to(dtype=sample.dtype)
|
225 |
+
|
226 |
+
emb = self.time_embedding(t_emb, None)
|
227 |
+
return emb
|
228 |
+
|
229 |
+
|
230 |
+
class CrossLayerCrossScaleProjector(nn.Module):
|
231 |
+
def __init__(
|
232 |
+
self,
|
233 |
+
inner_dim=2688,
|
234 |
+
num_attention_heads=42,
|
235 |
+
attention_head_dim=64,
|
236 |
+
cross_attention_dim=2688,
|
237 |
+
num_layers=4,
|
238 |
+
|
239 |
+
# resampler
|
240 |
+
dim=1280,
|
241 |
+
depth=4,
|
242 |
+
dim_head=64,
|
243 |
+
heads=20,
|
244 |
+
num_queries=1024,
|
245 |
+
embedding_dim=1152 + 1536,
|
246 |
+
output_dim=4096,
|
247 |
+
ff_mult=4,
|
248 |
+
timestep_in_dim=320,
|
249 |
+
timestep_flip_sin_to_cos=True,
|
250 |
+
timestep_freq_shift=0,
|
251 |
+
):
|
252 |
+
super().__init__()
|
253 |
+
|
254 |
+
self.cross_layer_blocks = nn.ModuleList(
|
255 |
+
[
|
256 |
+
BasicTransformerBlock(
|
257 |
+
inner_dim,
|
258 |
+
num_attention_heads,
|
259 |
+
attention_head_dim,
|
260 |
+
dropout=0,
|
261 |
+
cross_attention_dim=cross_attention_dim,
|
262 |
+
activation_fn="geglu",
|
263 |
+
num_embeds_ada_norm=None,
|
264 |
+
attention_bias=False,
|
265 |
+
only_cross_attention=False,
|
266 |
+
double_self_attention=False,
|
267 |
+
upcast_attention=False,
|
268 |
+
norm_type='layer_norm',
|
269 |
+
norm_elementwise_affine=True,
|
270 |
+
norm_eps=1e-6,
|
271 |
+
attention_type="default",
|
272 |
+
)
|
273 |
+
for _ in range(num_layers)
|
274 |
+
]
|
275 |
+
)
|
276 |
+
|
277 |
+
self.cross_scale_blocks = nn.ModuleList(
|
278 |
+
[
|
279 |
+
BasicTransformerBlock(
|
280 |
+
inner_dim,
|
281 |
+
num_attention_heads,
|
282 |
+
attention_head_dim,
|
283 |
+
dropout=0,
|
284 |
+
cross_attention_dim=cross_attention_dim,
|
285 |
+
activation_fn="geglu",
|
286 |
+
num_embeds_ada_norm=None,
|
287 |
+
attention_bias=False,
|
288 |
+
only_cross_attention=False,
|
289 |
+
double_self_attention=False,
|
290 |
+
upcast_attention=False,
|
291 |
+
norm_type='layer_norm',
|
292 |
+
norm_elementwise_affine=True,
|
293 |
+
norm_eps=1e-6,
|
294 |
+
attention_type="default",
|
295 |
+
)
|
296 |
+
for _ in range(num_layers)
|
297 |
+
]
|
298 |
+
)
|
299 |
+
|
300 |
+
self.proj = Mlp(
|
301 |
+
in_features=inner_dim,
|
302 |
+
hidden_features=int(inner_dim*2),
|
303 |
+
act_layer=lambda: nn.GELU(approximate="tanh"),
|
304 |
+
drop=0
|
305 |
+
)
|
306 |
+
|
307 |
+
self.proj_cross_layer = Mlp(
|
308 |
+
in_features=inner_dim,
|
309 |
+
hidden_features=int(inner_dim*2),
|
310 |
+
act_layer=lambda: nn.GELU(approximate="tanh"),
|
311 |
+
drop=0
|
312 |
+
)
|
313 |
+
|
314 |
+
self.proj_cross_scale = Mlp(
|
315 |
+
in_features=inner_dim,
|
316 |
+
hidden_features=int(inner_dim*2),
|
317 |
+
act_layer=lambda: nn.GELU(approximate="tanh"),
|
318 |
+
drop=0
|
319 |
+
)
|
320 |
+
|
321 |
+
self.resampler = TimeResampler(
|
322 |
+
dim=dim,
|
323 |
+
depth=depth,
|
324 |
+
dim_head=dim_head,
|
325 |
+
heads=heads,
|
326 |
+
num_queries=num_queries,
|
327 |
+
embedding_dim=embedding_dim,
|
328 |
+
output_dim=output_dim,
|
329 |
+
ff_mult=ff_mult,
|
330 |
+
timestep_in_dim=timestep_in_dim,
|
331 |
+
timestep_flip_sin_to_cos=timestep_flip_sin_to_cos,
|
332 |
+
timestep_freq_shift=timestep_freq_shift,
|
333 |
+
)
|
334 |
+
|
335 |
+
def forward(self, low_res_shallow, low_res_deep, high_res_deep, timesteps, cross_attention_kwargs=None, need_temb=True):
|
336 |
+
'''
|
337 |
+
low_res_shallow [bs, 729*l, c]
|
338 |
+
low_res_deep [bs, 729, c]
|
339 |
+
high_res_deep [bs, 729*4, c]
|
340 |
+
'''
|
341 |
+
|
342 |
+
cross_layer_hidden_states = low_res_deep
|
343 |
+
for block in self.cross_layer_blocks:
|
344 |
+
cross_layer_hidden_states = block(
|
345 |
+
cross_layer_hidden_states,
|
346 |
+
encoder_hidden_states=low_res_shallow,
|
347 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
348 |
+
)
|
349 |
+
cross_layer_hidden_states = self.proj_cross_layer(cross_layer_hidden_states)
|
350 |
+
|
351 |
+
cross_scale_hidden_states = low_res_deep
|
352 |
+
for block in self.cross_scale_blocks:
|
353 |
+
cross_scale_hidden_states = block(
|
354 |
+
cross_scale_hidden_states,
|
355 |
+
encoder_hidden_states=high_res_deep,
|
356 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
357 |
+
)
|
358 |
+
cross_scale_hidden_states = self.proj_cross_scale(cross_scale_hidden_states)
|
359 |
+
|
360 |
+
hidden_states = self.proj(low_res_deep) + cross_scale_hidden_states
|
361 |
+
hidden_states = torch.cat([hidden_states, cross_layer_hidden_states], dim=1)
|
362 |
+
|
363 |
+
hidden_states, timestep_emb = self.resampler(hidden_states, timesteps, need_temb=True)
|
364 |
+
return hidden_states, timestep_emb
|
365 |
+
|
models/utils.py
ADDED
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from safetensors.torch import load_file
|
2 |
+
import torch
|
3 |
+
from tqdm import tqdm
|
4 |
+
|
5 |
+
__all__ = [
|
6 |
+
'flux_load_lora'
|
7 |
+
]
|
8 |
+
|
9 |
+
|
10 |
+
def is_int(d):
|
11 |
+
try:
|
12 |
+
d = int(d)
|
13 |
+
return True
|
14 |
+
except Exception as e:
|
15 |
+
return False
|
16 |
+
|
17 |
+
|
18 |
+
def flux_load_lora(self, lora_file, lora_weight=1.0):
|
19 |
+
device = self.transformer.device
|
20 |
+
|
21 |
+
# DiT 部分
|
22 |
+
state_dict, network_alphas = self.lora_state_dict(lora_file, return_alphas=True)
|
23 |
+
state_dict = {k:v.to(device) for k,v in state_dict.items()}
|
24 |
+
|
25 |
+
model = self.transformer
|
26 |
+
keys = list(state_dict.keys())
|
27 |
+
keys = [k for k in keys if k.startswith('transformer.')]
|
28 |
+
|
29 |
+
for k_lora in tqdm(keys, total=len(keys), desc=f"loading lora in transformer ..."):
|
30 |
+
v_lora = state_dict[k_lora]
|
31 |
+
|
32 |
+
# 非 up 的都跳过
|
33 |
+
if '.lora_A.weight' in k_lora:
|
34 |
+
continue
|
35 |
+
if '.alpha' in k_lora:
|
36 |
+
continue
|
37 |
+
|
38 |
+
k_lora_name = k_lora.replace("transformer.", "")
|
39 |
+
k_lora_name = k_lora_name.replace(".lora_B.weight", "")
|
40 |
+
attr_name_list = k_lora_name.split('.')
|
41 |
+
|
42 |
+
cur_attr = model
|
43 |
+
latest_attr_name = ''
|
44 |
+
for idx in range(0, len(attr_name_list)):
|
45 |
+
attr_name = attr_name_list[idx]
|
46 |
+
if is_int(attr_name):
|
47 |
+
cur_attr = cur_attr[int(attr_name)]
|
48 |
+
latest_attr_name = ''
|
49 |
+
else:
|
50 |
+
try:
|
51 |
+
if latest_attr_name != '':
|
52 |
+
cur_attr = cur_attr.__getattr__(f"{latest_attr_name}.{attr_name}")
|
53 |
+
else:
|
54 |
+
cur_attr = cur_attr.__getattr__(attr_name)
|
55 |
+
latest_attr_name = ''
|
56 |
+
except Exception as e:
|
57 |
+
if latest_attr_name != '':
|
58 |
+
latest_attr_name = f"{latest_attr_name}.{attr_name}"
|
59 |
+
else:
|
60 |
+
latest_attr_name = attr_name
|
61 |
+
|
62 |
+
up_w = v_lora
|
63 |
+
down_w = state_dict[k_lora.replace('.lora_B.weight', '.lora_A.weight')]
|
64 |
+
|
65 |
+
# 赋值
|
66 |
+
einsum_a = f"ijabcdefg"
|
67 |
+
einsum_b = f"jkabcdefg"
|
68 |
+
einsum_res = f"ikabcdefg"
|
69 |
+
length_shape = len(up_w.shape)
|
70 |
+
einsum_str = f"{einsum_a[:length_shape]},{einsum_b[:length_shape]}->{einsum_res[:length_shape]}"
|
71 |
+
dtype = cur_attr.weight.data.dtype
|
72 |
+
d_w = torch.einsum(einsum_str, up_w.to(torch.float32), down_w.to(torch.float32)).to(dtype)
|
73 |
+
cur_attr.weight.data = cur_attr.weight.data + d_w * lora_weight
|
74 |
+
|
75 |
+
|
76 |
+
|
77 |
+
# text encoder 部分
|
78 |
+
raw_state_dict = load_file(lora_file)
|
79 |
+
raw_state_dict = {k:v.to(device) for k,v in raw_state_dict.items()}
|
80 |
+
|
81 |
+
# text encoder
|
82 |
+
state_dict = {k:v for k,v in raw_state_dict.items() if 'lora_te1_' in k}
|
83 |
+
model = self.text_encoder
|
84 |
+
keys = list(state_dict.keys())
|
85 |
+
keys = [k for k in keys if k.startswith('lora_te1_')]
|
86 |
+
|
87 |
+
for k_lora in tqdm(keys, total=len(keys), desc=f"loading lora in text_encoder ..."):
|
88 |
+
v_lora = state_dict[k_lora]
|
89 |
+
|
90 |
+
# 非 up 的都跳过
|
91 |
+
if '.lora_down.weight' in k_lora:
|
92 |
+
continue
|
93 |
+
if '.alpha' in k_lora:
|
94 |
+
continue
|
95 |
+
|
96 |
+
k_lora_name = k_lora.replace("lora_te1_", "")
|
97 |
+
k_lora_name = k_lora_name.replace(".lora_up.weight", "")
|
98 |
+
attr_name_list = k_lora_name.split('_')
|
99 |
+
|
100 |
+
cur_attr = model
|
101 |
+
latest_attr_name = ''
|
102 |
+
for idx in range(0, len(attr_name_list)):
|
103 |
+
attr_name = attr_name_list[idx]
|
104 |
+
if is_int(attr_name):
|
105 |
+
cur_attr = cur_attr[int(attr_name)]
|
106 |
+
latest_attr_name = ''
|
107 |
+
else:
|
108 |
+
try:
|
109 |
+
if latest_attr_name != '':
|
110 |
+
cur_attr = cur_attr.__getattr__(f"{latest_attr_name}_{attr_name}")
|
111 |
+
else:
|
112 |
+
cur_attr = cur_attr.__getattr__(attr_name)
|
113 |
+
latest_attr_name = ''
|
114 |
+
except Exception as e:
|
115 |
+
if latest_attr_name != '':
|
116 |
+
latest_attr_name = f"{latest_attr_name}_{attr_name}"
|
117 |
+
else:
|
118 |
+
latest_attr_name = attr_name
|
119 |
+
|
120 |
+
up_w = v_lora
|
121 |
+
down_w = state_dict[k_lora.replace('.lora_up.weight', '.lora_down.weight')]
|
122 |
+
|
123 |
+
alpha = state_dict.get(k_lora.replace('.lora_up.weight', '.alpha'), None)
|
124 |
+
if alpha is None:
|
125 |
+
lora_scale = 1
|
126 |
+
else:
|
127 |
+
rank = up_w.shape[1]
|
128 |
+
lora_scale = alpha / rank
|
129 |
+
|
130 |
+
# 赋值
|
131 |
+
einsum_a = f"ijabcdefg"
|
132 |
+
einsum_b = f"jkabcdefg"
|
133 |
+
einsum_res = f"ikabcdefg"
|
134 |
+
length_shape = len(up_w.shape)
|
135 |
+
einsum_str = f"{einsum_a[:length_shape]},{einsum_b[:length_shape]}->{einsum_res[:length_shape]}"
|
136 |
+
dtype = cur_attr.weight.data.dtype
|
137 |
+
d_w = torch.einsum(einsum_str, up_w.to(torch.float32), down_w.to(torch.float32)).to(dtype)
|
138 |
+
cur_attr.weight.data = cur_attr.weight.data + d_w * lora_scale * lora_weight
|
139 |
+
|
pipeline.py
ADDED
@@ -0,0 +1,550 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2025 Tencent InstantX Team. All rights reserved.
|
2 |
+
#
|
3 |
+
|
4 |
+
from PIL import Image
|
5 |
+
from einops import rearrange
|
6 |
+
import torch
|
7 |
+
from diffusers.pipelines.flux.pipeline_flux import *
|
8 |
+
from transformers import SiglipVisionModel, SiglipImageProcessor, AutoModel, AutoImageProcessor
|
9 |
+
|
10 |
+
from models.attn_processor import FluxIPAttnProcessor
|
11 |
+
from models.resampler import CrossLayerCrossScaleProjector
|
12 |
+
from models.utils import flux_load_lora
|
13 |
+
|
14 |
+
|
15 |
+
# TODO
|
16 |
+
EXAMPLE_DOC_STRING = """
|
17 |
+
Examples:
|
18 |
+
```py
|
19 |
+
>>> import torch
|
20 |
+
>>> from diffusers import FluxPipeline
|
21 |
+
|
22 |
+
>>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
|
23 |
+
>>> pipe.to("cuda")
|
24 |
+
>>> prompt = "A cat holding a sign that says hello world"
|
25 |
+
>>> # Depending on the variant being used, the pipeline call will slightly vary.
|
26 |
+
>>> # Refer to the pipeline documentation for more details.
|
27 |
+
>>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0]
|
28 |
+
>>> image.save("flux.png")
|
29 |
+
```
|
30 |
+
"""
|
31 |
+
|
32 |
+
|
33 |
+
class InstantCharacterFluxPipeline(FluxPipeline):
|
34 |
+
|
35 |
+
|
36 |
+
@torch.inference_mode()
|
37 |
+
def encode_siglip_image_emb(self, siglip_image, device, dtype):
|
38 |
+
siglip_image = siglip_image.to(device, dtype=dtype)
|
39 |
+
res = self.siglip_image_encoder(siglip_image, output_hidden_states=True)
|
40 |
+
|
41 |
+
siglip_image_embeds = res.last_hidden_state
|
42 |
+
|
43 |
+
siglip_image_shallow_embeds = torch.cat([res.hidden_states[i] for i in [7, 13, 26]], dim=1)
|
44 |
+
|
45 |
+
return siglip_image_embeds, siglip_image_shallow_embeds
|
46 |
+
|
47 |
+
|
48 |
+
@torch.inference_mode()
|
49 |
+
def encode_dinov2_image_emb(self, dinov2_image, device, dtype):
|
50 |
+
dinov2_image = dinov2_image.to(device, dtype=dtype)
|
51 |
+
res = self.dino_image_encoder_2(dinov2_image, output_hidden_states=True)
|
52 |
+
|
53 |
+
dinov2_image_embeds = res.last_hidden_state[:, 1:]
|
54 |
+
|
55 |
+
dinov2_image_shallow_embeds = torch.cat([res.hidden_states[i][:, 1:] for i in [9, 19, 29]], dim=1)
|
56 |
+
|
57 |
+
return dinov2_image_embeds, dinov2_image_shallow_embeds
|
58 |
+
|
59 |
+
|
60 |
+
@torch.inference_mode()
|
61 |
+
def encode_image_emb(self, siglip_image, device, dtype):
|
62 |
+
object_image_pil = siglip_image
|
63 |
+
object_image_pil_low_res = [object_image_pil.resize((384, 384))]
|
64 |
+
object_image_pil_high_res = object_image_pil.resize((768, 768))
|
65 |
+
object_image_pil_high_res = [
|
66 |
+
object_image_pil_high_res.crop((0, 0, 384, 384)),
|
67 |
+
object_image_pil_high_res.crop((384, 0, 768, 384)),
|
68 |
+
object_image_pil_high_res.crop((0, 384, 384, 768)),
|
69 |
+
object_image_pil_high_res.crop((384, 384, 768, 768)),
|
70 |
+
]
|
71 |
+
nb_split_image = len(object_image_pil_high_res)
|
72 |
+
|
73 |
+
siglip_image_embeds = self.encode_siglip_image_emb(
|
74 |
+
self.siglip_image_processor(images=object_image_pil_low_res, return_tensors="pt").pixel_values,
|
75 |
+
device,
|
76 |
+
dtype
|
77 |
+
)
|
78 |
+
dinov2_image_embeds = self.encode_dinov2_image_emb(
|
79 |
+
self.dino_image_processor_2(images=object_image_pil_low_res, return_tensors="pt").pixel_values,
|
80 |
+
device,
|
81 |
+
dtype
|
82 |
+
)
|
83 |
+
|
84 |
+
image_embeds_low_res_deep = torch.cat([siglip_image_embeds[0], dinov2_image_embeds[0]], dim=2)
|
85 |
+
image_embeds_low_res_shallow = torch.cat([siglip_image_embeds[1], dinov2_image_embeds[1]], dim=2)
|
86 |
+
|
87 |
+
siglip_image_high_res = self.siglip_image_processor(images=object_image_pil_high_res, return_tensors="pt").pixel_values
|
88 |
+
siglip_image_high_res = siglip_image_high_res[None]
|
89 |
+
siglip_image_high_res = rearrange(siglip_image_high_res, 'b n c h w -> (b n) c h w')
|
90 |
+
siglip_image_high_res_embeds = self.encode_siglip_image_emb(siglip_image_high_res, device, dtype)
|
91 |
+
siglip_image_high_res_deep = rearrange(siglip_image_high_res_embeds[0], '(b n) l c -> b (n l) c', n=nb_split_image)
|
92 |
+
dinov2_image_high_res = self.dino_image_processor_2(images=object_image_pil_high_res, return_tensors="pt").pixel_values
|
93 |
+
dinov2_image_high_res = dinov2_image_high_res[None]
|
94 |
+
dinov2_image_high_res = rearrange(dinov2_image_high_res, 'b n c h w -> (b n) c h w')
|
95 |
+
dinov2_image_high_res_embeds = self.encode_dinov2_image_emb(dinov2_image_high_res, device, dtype)
|
96 |
+
dinov2_image_high_res_deep = rearrange(dinov2_image_high_res_embeds[0], '(b n) l c -> b (n l) c', n=nb_split_image)
|
97 |
+
image_embeds_high_res_deep = torch.cat([siglip_image_high_res_deep, dinov2_image_high_res_deep], dim=2)
|
98 |
+
|
99 |
+
image_embeds_dict = dict(
|
100 |
+
image_embeds_low_res_shallow=image_embeds_low_res_shallow,
|
101 |
+
image_embeds_low_res_deep=image_embeds_low_res_deep,
|
102 |
+
image_embeds_high_res_deep=image_embeds_high_res_deep,
|
103 |
+
)
|
104 |
+
return image_embeds_dict
|
105 |
+
|
106 |
+
|
107 |
+
@torch.inference_mode()
|
108 |
+
def init_ccp_and_attn_processor(self, *args, **kwargs):
|
109 |
+
subject_ip_adapter_path = kwargs['subject_ip_adapter_path']
|
110 |
+
nb_token = kwargs['nb_token']
|
111 |
+
state_dict = torch.load(subject_ip_adapter_path, map_location="cpu")
|
112 |
+
device, dtype = self.transformer.device, self.transformer.dtype
|
113 |
+
|
114 |
+
print(f"=> init attn processor")
|
115 |
+
attn_procs = {}
|
116 |
+
for idx_attn, (name, v) in enumerate(self.transformer.attn_processors.items()):
|
117 |
+
attn_procs[name] = FluxIPAttnProcessor(
|
118 |
+
hidden_size=self.transformer.config.attention_head_dim * self.transformer.config.num_attention_heads,
|
119 |
+
ip_hidden_states_dim=self.text_encoder_2.config.d_model,
|
120 |
+
).to(device, dtype=dtype)
|
121 |
+
self.transformer.set_attn_processor(attn_procs)
|
122 |
+
tmp_ip_layers = torch.nn.ModuleList(self.transformer.attn_processors.values())
|
123 |
+
key_name = tmp_ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
|
124 |
+
print(f"=> load attn processor: {key_name}")
|
125 |
+
|
126 |
+
print(f"=> init project")
|
127 |
+
image_proj_model = CrossLayerCrossScaleProjector(
|
128 |
+
inner_dim=1152 + 1536,
|
129 |
+
num_attention_heads=42,
|
130 |
+
attention_head_dim=64,
|
131 |
+
cross_attention_dim=1152 + 1536,
|
132 |
+
num_layers=4,
|
133 |
+
dim=1280,
|
134 |
+
depth=4,
|
135 |
+
dim_head=64,
|
136 |
+
heads=20,
|
137 |
+
num_queries=nb_token,
|
138 |
+
embedding_dim=1152 + 1536,
|
139 |
+
output_dim=4096,
|
140 |
+
ff_mult=4,
|
141 |
+
timestep_in_dim=320,
|
142 |
+
timestep_flip_sin_to_cos=True,
|
143 |
+
timestep_freq_shift=0,
|
144 |
+
)
|
145 |
+
image_proj_model.eval()
|
146 |
+
image_proj_model.to(device, dtype=dtype)
|
147 |
+
|
148 |
+
key_name = image_proj_model.load_state_dict(state_dict["image_proj"], strict=False)
|
149 |
+
print(f"=> load project: {key_name}")
|
150 |
+
self.subject_image_proj_model = image_proj_model
|
151 |
+
|
152 |
+
|
153 |
+
@torch.inference_mode()
|
154 |
+
def init_adapter(
|
155 |
+
self,
|
156 |
+
image_encoder_path=None,
|
157 |
+
image_encoder_2_path=None,
|
158 |
+
subject_ipadapter_cfg=None,
|
159 |
+
):
|
160 |
+
device, dtype = self.transformer.device, self.transformer.dtype
|
161 |
+
|
162 |
+
# image encoder
|
163 |
+
print(f"=> loading image_encoder_1: {image_encoder_path}")
|
164 |
+
image_encoder = SiglipVisionModel.from_pretrained(image_encoder_path)
|
165 |
+
image_processor = SiglipImageProcessor.from_pretrained(image_encoder_path)
|
166 |
+
image_encoder.eval()
|
167 |
+
image_encoder.to(device, dtype=dtype)
|
168 |
+
self.siglip_image_encoder = image_encoder
|
169 |
+
self.siglip_image_processor = image_processor
|
170 |
+
|
171 |
+
# image encoder 2
|
172 |
+
print(f"=> loading image_encoder_2: {image_encoder_2_path}")
|
173 |
+
image_encoder_2 = AutoModel.from_pretrained(image_encoder_2_path)
|
174 |
+
image_processor_2 = AutoImageProcessor.from_pretrained(image_encoder_2_path)
|
175 |
+
image_encoder_2.eval()
|
176 |
+
image_encoder_2.to(device, dtype=dtype)
|
177 |
+
image_processor_2.crop_size = dict(height=384, width=384)
|
178 |
+
image_processor_2.size = dict(shortest_edge=384)
|
179 |
+
self.dino_image_encoder_2 = image_encoder_2
|
180 |
+
self.dino_image_processor_2 = image_processor_2
|
181 |
+
|
182 |
+
# ccp and adapter
|
183 |
+
self.init_ccp_and_attn_processor(**subject_ipadapter_cfg)
|
184 |
+
|
185 |
+
|
186 |
+
@torch.no_grad()
|
187 |
+
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
188 |
+
def __call__(
|
189 |
+
self,
|
190 |
+
prompt: Union[str, List[str]] = None,
|
191 |
+
prompt_2: Optional[Union[str, List[str]]] = None,
|
192 |
+
negative_prompt: Union[str, List[str]] = None,
|
193 |
+
negative_prompt_2: Optional[Union[str, List[str]]] = None,
|
194 |
+
true_cfg_scale: float = 1.0,
|
195 |
+
height: Optional[int] = None,
|
196 |
+
width: Optional[int] = None,
|
197 |
+
num_inference_steps: int = 28,
|
198 |
+
sigmas: Optional[List[float]] = None,
|
199 |
+
guidance_scale: float = 3.5,
|
200 |
+
num_images_per_prompt: Optional[int] = 1,
|
201 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
202 |
+
latents: Optional[torch.FloatTensor] = None,
|
203 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
204 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
205 |
+
ip_adapter_image: Optional[PipelineImageInput] = None,
|
206 |
+
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
|
207 |
+
negative_ip_adapter_image: Optional[PipelineImageInput] = None,
|
208 |
+
negative_ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
|
209 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
210 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
211 |
+
output_type: Optional[str] = "pil",
|
212 |
+
return_dict: bool = True,
|
213 |
+
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
214 |
+
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
215 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
216 |
+
max_sequence_length: int = 512,
|
217 |
+
subject_image: Image.Image = None,
|
218 |
+
subject_scale: float = 0.8,
|
219 |
+
|
220 |
+
):
|
221 |
+
r"""
|
222 |
+
Function invoked when calling the pipeline for generation.
|
223 |
+
|
224 |
+
Args:
|
225 |
+
prompt (`str` or `List[str]`, *optional*):
|
226 |
+
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
227 |
+
instead.
|
228 |
+
prompt_2 (`str` or `List[str]`, *optional*):
|
229 |
+
The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
230 |
+
will be used instead
|
231 |
+
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
232 |
+
The height in pixels of the generated image. This is set to 1024 by default for the best results.
|
233 |
+
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
234 |
+
The width in pixels of the generated image. This is set to 1024 by default for the best results.
|
235 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
236 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
237 |
+
expense of slower inference.
|
238 |
+
sigmas (`List[float]`, *optional*):
|
239 |
+
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
|
240 |
+
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
|
241 |
+
will be used.
|
242 |
+
guidance_scale (`float`, *optional*, defaults to 7.0):
|
243 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
244 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
245 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
246 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
247 |
+
usually at the expense of lower image quality.
|
248 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
249 |
+
The number of images to generate per prompt.
|
250 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
251 |
+
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
252 |
+
to make generation deterministic.
|
253 |
+
latents (`torch.FloatTensor`, *optional*):
|
254 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
255 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
256 |
+
tensor will ge generated by sampling using the supplied random `generator`.
|
257 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
258 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
259 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
260 |
+
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
261 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
262 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
263 |
+
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
|
264 |
+
ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
|
265 |
+
Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
|
266 |
+
IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not
|
267 |
+
provided, embeddings are computed from the `ip_adapter_image` input argument.
|
268 |
+
negative_ip_adapter_image:
|
269 |
+
(`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
|
270 |
+
negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
|
271 |
+
Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
|
272 |
+
IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not
|
273 |
+
provided, embeddings are computed from the `ip_adapter_image` input argument.
|
274 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
275 |
+
The output format of the generate image. Choose between
|
276 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
277 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
278 |
+
Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple.
|
279 |
+
joint_attention_kwargs (`dict`, *optional*):
|
280 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
281 |
+
`self.processor` in
|
282 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
283 |
+
callback_on_step_end (`Callable`, *optional*):
|
284 |
+
A function that calls at the end of each denoising steps during the inference. The function is called
|
285 |
+
with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
|
286 |
+
callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
|
287 |
+
`callback_on_step_end_tensor_inputs`.
|
288 |
+
callback_on_step_end_tensor_inputs (`List`, *optional*):
|
289 |
+
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
290 |
+
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
291 |
+
`._callback_tensor_inputs` attribute of your pipeline class.
|
292 |
+
max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
|
293 |
+
|
294 |
+
Examples:
|
295 |
+
|
296 |
+
Returns:
|
297 |
+
[`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict`
|
298 |
+
is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated
|
299 |
+
images.
|
300 |
+
"""
|
301 |
+
|
302 |
+
height = height or self.default_sample_size * self.vae_scale_factor
|
303 |
+
width = width or self.default_sample_size * self.vae_scale_factor
|
304 |
+
|
305 |
+
# 1. Check inputs. Raise error if not correct
|
306 |
+
self.check_inputs(
|
307 |
+
prompt,
|
308 |
+
prompt_2,
|
309 |
+
height,
|
310 |
+
width,
|
311 |
+
negative_prompt=negative_prompt,
|
312 |
+
negative_prompt_2=negative_prompt_2,
|
313 |
+
prompt_embeds=prompt_embeds,
|
314 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
315 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
316 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
317 |
+
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
318 |
+
max_sequence_length=max_sequence_length,
|
319 |
+
)
|
320 |
+
|
321 |
+
self._guidance_scale = guidance_scale
|
322 |
+
self._joint_attention_kwargs = joint_attention_kwargs
|
323 |
+
self._interrupt = False
|
324 |
+
|
325 |
+
# 2. Define call parameters
|
326 |
+
if prompt is not None and isinstance(prompt, str):
|
327 |
+
batch_size = 1
|
328 |
+
elif prompt is not None and isinstance(prompt, list):
|
329 |
+
batch_size = len(prompt)
|
330 |
+
else:
|
331 |
+
batch_size = prompt_embeds.shape[0]
|
332 |
+
|
333 |
+
device = self._execution_device
|
334 |
+
dtype = self.transformer.dtype
|
335 |
+
|
336 |
+
lora_scale = (
|
337 |
+
self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
|
338 |
+
)
|
339 |
+
do_true_cfg = true_cfg_scale > 1 and negative_prompt is not None
|
340 |
+
(
|
341 |
+
prompt_embeds,
|
342 |
+
pooled_prompt_embeds,
|
343 |
+
text_ids,
|
344 |
+
) = self.encode_prompt(
|
345 |
+
prompt=prompt,
|
346 |
+
prompt_2=prompt_2,
|
347 |
+
prompt_embeds=prompt_embeds,
|
348 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
349 |
+
device=device,
|
350 |
+
num_images_per_prompt=num_images_per_prompt,
|
351 |
+
max_sequence_length=max_sequence_length,
|
352 |
+
lora_scale=lora_scale,
|
353 |
+
)
|
354 |
+
if do_true_cfg:
|
355 |
+
(
|
356 |
+
negative_prompt_embeds,
|
357 |
+
negative_pooled_prompt_embeds,
|
358 |
+
_,
|
359 |
+
) = self.encode_prompt(
|
360 |
+
prompt=negative_prompt,
|
361 |
+
prompt_2=negative_prompt_2,
|
362 |
+
prompt_embeds=negative_prompt_embeds,
|
363 |
+
pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
364 |
+
device=device,
|
365 |
+
num_images_per_prompt=num_images_per_prompt,
|
366 |
+
max_sequence_length=max_sequence_length,
|
367 |
+
lora_scale=lora_scale,
|
368 |
+
)
|
369 |
+
|
370 |
+
# 3.1 Prepare subject emb
|
371 |
+
if subject_image is not None:
|
372 |
+
subject_image = subject_image.resize((max(subject_image.size), max(subject_image.size)))
|
373 |
+
subject_image_embeds_dict = self.encode_image_emb(subject_image, device, dtype)
|
374 |
+
|
375 |
+
# 4. Prepare latent variables
|
376 |
+
num_channels_latents = self.transformer.config.in_channels // 4
|
377 |
+
latents, latent_image_ids = self.prepare_latents(
|
378 |
+
batch_size * num_images_per_prompt,
|
379 |
+
num_channels_latents,
|
380 |
+
height,
|
381 |
+
width,
|
382 |
+
prompt_embeds.dtype,
|
383 |
+
device,
|
384 |
+
generator,
|
385 |
+
latents,
|
386 |
+
)
|
387 |
+
|
388 |
+
# 5. Prepare timesteps
|
389 |
+
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
|
390 |
+
image_seq_len = latents.shape[1]
|
391 |
+
mu = calculate_shift(
|
392 |
+
image_seq_len,
|
393 |
+
self.scheduler.config.base_image_seq_len,
|
394 |
+
self.scheduler.config.max_image_seq_len,
|
395 |
+
self.scheduler.config.base_shift,
|
396 |
+
self.scheduler.config.max_shift,
|
397 |
+
)
|
398 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
399 |
+
self.scheduler,
|
400 |
+
num_inference_steps,
|
401 |
+
device,
|
402 |
+
sigmas=sigmas,
|
403 |
+
mu=mu,
|
404 |
+
)
|
405 |
+
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
406 |
+
self._num_timesteps = len(timesteps)
|
407 |
+
|
408 |
+
# handle guidance
|
409 |
+
if self.transformer.config.guidance_embeds:
|
410 |
+
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
|
411 |
+
guidance = guidance.expand(latents.shape[0])
|
412 |
+
else:
|
413 |
+
guidance = None
|
414 |
+
|
415 |
+
if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and (
|
416 |
+
negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None
|
417 |
+
):
|
418 |
+
negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8)
|
419 |
+
elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and (
|
420 |
+
negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None
|
421 |
+
):
|
422 |
+
ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8)
|
423 |
+
|
424 |
+
if self.joint_attention_kwargs is None:
|
425 |
+
self._joint_attention_kwargs = {}
|
426 |
+
|
427 |
+
image_embeds = None
|
428 |
+
negative_image_embeds = None
|
429 |
+
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
|
430 |
+
image_embeds = self.prepare_ip_adapter_image_embeds(
|
431 |
+
ip_adapter_image,
|
432 |
+
ip_adapter_image_embeds,
|
433 |
+
device,
|
434 |
+
batch_size * num_images_per_prompt,
|
435 |
+
)
|
436 |
+
if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None:
|
437 |
+
negative_image_embeds = self.prepare_ip_adapter_image_embeds(
|
438 |
+
negative_ip_adapter_image,
|
439 |
+
negative_ip_adapter_image_embeds,
|
440 |
+
device,
|
441 |
+
batch_size * num_images_per_prompt,
|
442 |
+
)
|
443 |
+
|
444 |
+
# 6. Denoising loop
|
445 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
446 |
+
for i, t in enumerate(timesteps):
|
447 |
+
if self.interrupt:
|
448 |
+
continue
|
449 |
+
|
450 |
+
if image_embeds is not None:
|
451 |
+
self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds
|
452 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
453 |
+
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
454 |
+
|
455 |
+
|
456 |
+
# subject adapter
|
457 |
+
if subject_image is not None:
|
458 |
+
subject_image_prompt_embeds = self.subject_image_proj_model(
|
459 |
+
low_res_shallow=subject_image_embeds_dict['image_embeds_low_res_shallow'],
|
460 |
+
low_res_deep=subject_image_embeds_dict['image_embeds_low_res_deep'],
|
461 |
+
high_res_deep=subject_image_embeds_dict['image_embeds_high_res_deep'],
|
462 |
+
timesteps=timestep.to(dtype=latents.dtype),
|
463 |
+
need_temb=True
|
464 |
+
)[0]
|
465 |
+
self._joint_attention_kwargs['emb_dict'] = dict(
|
466 |
+
length_encoder_hidden_states=prompt_embeds.shape[1]
|
467 |
+
)
|
468 |
+
self._joint_attention_kwargs['subject_emb_dict'] = dict(
|
469 |
+
ip_hidden_states=subject_image_prompt_embeds,
|
470 |
+
scale=subject_scale,
|
471 |
+
)
|
472 |
+
|
473 |
+
noise_pred = self.transformer(
|
474 |
+
hidden_states=latents,
|
475 |
+
timestep=timestep / 1000,
|
476 |
+
guidance=guidance,
|
477 |
+
pooled_projections=pooled_prompt_embeds,
|
478 |
+
encoder_hidden_states=prompt_embeds,
|
479 |
+
txt_ids=text_ids,
|
480 |
+
img_ids=latent_image_ids,
|
481 |
+
joint_attention_kwargs=self.joint_attention_kwargs,
|
482 |
+
return_dict=False,
|
483 |
+
)[0]
|
484 |
+
|
485 |
+
if do_true_cfg:
|
486 |
+
if negative_image_embeds is not None:
|
487 |
+
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
|
488 |
+
neg_noise_pred = self.transformer(
|
489 |
+
hidden_states=latents,
|
490 |
+
timestep=timestep / 1000,
|
491 |
+
guidance=guidance,
|
492 |
+
pooled_projections=negative_pooled_prompt_embeds,
|
493 |
+
encoder_hidden_states=negative_prompt_embeds,
|
494 |
+
txt_ids=text_ids,
|
495 |
+
img_ids=latent_image_ids,
|
496 |
+
joint_attention_kwargs=self.joint_attention_kwargs,
|
497 |
+
return_dict=False,
|
498 |
+
)[0]
|
499 |
+
noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
|
500 |
+
|
501 |
+
# compute the previous noisy sample x_t -> x_t-1
|
502 |
+
latents_dtype = latents.dtype
|
503 |
+
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
504 |
+
|
505 |
+
if latents.dtype != latents_dtype:
|
506 |
+
if torch.backends.mps.is_available():
|
507 |
+
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
508 |
+
latents = latents.to(latents_dtype)
|
509 |
+
|
510 |
+
if callback_on_step_end is not None:
|
511 |
+
callback_kwargs = {}
|
512 |
+
for k in callback_on_step_end_tensor_inputs:
|
513 |
+
callback_kwargs[k] = locals()[k]
|
514 |
+
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
515 |
+
|
516 |
+
latents = callback_outputs.pop("latents", latents)
|
517 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
518 |
+
|
519 |
+
# call the callback, if provided
|
520 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
521 |
+
progress_bar.update()
|
522 |
+
|
523 |
+
if XLA_AVAILABLE:
|
524 |
+
xm.mark_step()
|
525 |
+
|
526 |
+
if output_type == "latent":
|
527 |
+
image = latents
|
528 |
+
|
529 |
+
else:
|
530 |
+
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
|
531 |
+
latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
|
532 |
+
image = self.vae.decode(latents, return_dict=False)[0]
|
533 |
+
image = self.image_processor.postprocess(image, output_type=output_type)
|
534 |
+
|
535 |
+
# Offload all models
|
536 |
+
self.maybe_free_model_hooks()
|
537 |
+
|
538 |
+
if not return_dict:
|
539 |
+
return (image,)
|
540 |
+
|
541 |
+
return FluxPipelineOutput(images=image)
|
542 |
+
|
543 |
+
|
544 |
+
def with_style_lora(self, lora_file_path, lora_weight=1.0, trigger='', *args, **kwargs):
|
545 |
+
flux_load_lora(self, lora_file_path, lora_weight)
|
546 |
+
kwargs['prompt'] = f"{trigger}, {kwargs['prompt']}"
|
547 |
+
res = self.__call__(*args, **kwargs)
|
548 |
+
flux_load_lora(self, lora_file_path, -lora_weight)
|
549 |
+
return res
|
550 |
+
|