Spaces:
Running
on
Zero
Running
on
Zero
Update IP_Composer/perform_swap.py
Browse files- IP_Composer/perform_swap.py +14 -41
IP_Composer/perform_swap.py
CHANGED
@@ -1,64 +1,37 @@
|
|
1 |
import torch
|
2 |
import numpy as np
|
3 |
-
from typing import List, Dict, Optional
|
4 |
-
from PIL.Image import Image as PILImage
|
5 |
-
from IP_Adapter import IPAdapterXL
|
6 |
|
7 |
-
def compute_dataset_embeds_svd(
|
8 |
-
|
9 |
-
rank: int
|
10 |
-
) -> np.ndarray:
|
11 |
# Perform SVD on the combined matrix
|
12 |
-
|
13 |
|
14 |
# Select the top `rank` singular vectors to construct the projection matrix
|
15 |
-
|
16 |
-
projection_matrix =
|
17 |
|
18 |
return projection_matrix
|
19 |
|
20 |
-
def
|
21 |
-
embed
|
22 |
-
projection_matrix: np.ndarray
|
23 |
-
) -> np.ndarray:
|
24 |
-
return embed @ projection_matrix
|
25 |
-
|
26 |
-
def get_embedding_composition(
|
27 |
-
embed: np.ndarray,
|
28 |
-
projections_data: List[Dict[str, np.ndarray]]
|
29 |
-
) -> np.ndarray:
|
30 |
combined_embeds = embed.copy()
|
31 |
|
32 |
for proj_data in projections_data:
|
33 |
-
|
34 |
-
|
|
|
|
|
35 |
|
36 |
return combined_embeds
|
37 |
|
38 |
|
39 |
-
def get_modified_images_embeds_composition(
|
40 |
-
|
41 |
-
projections_data: List[Dict[str, np.ndarray]],
|
42 |
-
ip_model: IPAdapterXL,
|
43 |
-
prompt: Optional[str] = None,
|
44 |
-
scale: float = 1.0,
|
45 |
-
num_samples: int = 3,
|
46 |
-
seed: int = 420
|
47 |
-
) -> List[PILImage]:
|
48 |
final_embeds = get_embedding_composition(embed, projections_data)
|
49 |
clip_embeds = torch.from_numpy(final_embeds)
|
50 |
|
51 |
-
images
|
52 |
-
clip_image_embeds=clip_embeds,
|
53 |
-
prompt=prompt,
|
54 |
-
num_samples=num_samples,
|
55 |
-
num_inference_steps=50,
|
56 |
-
seed=seed,
|
57 |
-
guidance_scale=7.5,
|
58 |
-
scale=scale
|
59 |
-
)
|
60 |
return images
|
61 |
|
62 |
|
63 |
|
64 |
-
|
|
|
1 |
import torch
|
2 |
import numpy as np
|
|
|
|
|
|
|
3 |
|
4 |
+
def compute_dataset_embeds_svd(all_embeds, rank):
|
5 |
+
|
|
|
|
|
6 |
# Perform SVD on the combined matrix
|
7 |
+
u, s, vh = np.linalg.svd(all_embeds, full_matrices=False)
|
8 |
|
9 |
# Select the top `rank` singular vectors to construct the projection matrix
|
10 |
+
vh = vh[:rank] # Top `rank` right singular vectors
|
11 |
+
projection_matrix = vh.T @ vh # Shape: (feature_dim, feature_dim)
|
12 |
|
13 |
return projection_matrix
|
14 |
|
15 |
+
def get_embedding_composition(embed, projections_data):
|
16 |
+
# Initialize the combined embedding with the input embed
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
combined_embeds = embed.copy()
|
18 |
|
19 |
for proj_data in projections_data:
|
20 |
+
|
21 |
+
# Add the combined projection to the result
|
22 |
+
combined_embeds -= embed @ proj_data["projection_matrix"]
|
23 |
+
combined_embeds += proj_data["embed"] @ proj_data["projection_matrix"]
|
24 |
|
25 |
return combined_embeds
|
26 |
|
27 |
|
28 |
+
def get_modified_images_embeds_composition(embed, projections_data, ip_model, prompt=None, scale=1.0, num_samples=3, seed=420, num_inference_steps=50):
|
29 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
final_embeds = get_embedding_composition(embed, projections_data)
|
31 |
clip_embeds = torch.from_numpy(final_embeds)
|
32 |
|
33 |
+
images = ip_model.generate(clip_image_embeds=clip_embeds, prompt=prompt, num_samples=num_samples, num_inference_steps=num_inference_steps, seed=seed, guidance_scale=7.5, scale=scale)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
return images
|
35 |
|
36 |
|
37 |
|
|