Spaces:
Runtime error
Runtime error
Vivien Chappelier
commited on
Commit
Β·
464ec84
1
Parent(s):
fbe5687
use packaged VAEs
Browse files
app.py
CHANGED
|
@@ -7,7 +7,7 @@ import numpy as np
|
|
| 7 |
|
| 8 |
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
|
| 9 |
|
| 10 |
-
from diffusers import DiffusionPipeline
|
| 11 |
import torchvision.transforms as transforms
|
| 12 |
|
| 13 |
from copy import deepcopy
|
|
@@ -26,42 +26,35 @@ class BZHStableSignatureDemo(object):
|
|
| 26 |
|
| 27 |
def __init__(self, *args, **kwargs):
|
| 28 |
super().__init__(*args, **kwargs)
|
|
|
|
| 29 |
self.pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16").to("cuda")
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
for name,
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
("extreme", "models/checkpoint_000.pth.1500000")):
|
| 43 |
-
sd2 = torch.load(patched_decoder_ckpt)['ldm_decoder']
|
| 44 |
-
msg = self.pipe.vae.load_state_dict(sd2, strict=False)
|
| 45 |
-
print(f"loaded LDM decoder state_dict with message\n{msg}")
|
| 46 |
-
print("you should check that the decoder keys are correctly matched")
|
| 47 |
-
decoders[name] = sd2
|
| 48 |
self.decoders = decoders
|
| 49 |
|
| 50 |
def generate(self, mode, seed, prompt):
|
| 51 |
generator = torch.Generator(device=device)
|
| 52 |
-
#if seed:
|
| 53 |
torch.manual_seed(seed)
|
| 54 |
|
| 55 |
-
# load the patched VAE
|
| 56 |
-
|
| 57 |
-
self.pipe.vae
|
| 58 |
|
| 59 |
output = self.pipe(prompt, num_inference_steps=4, guidance_scale=0.0, output_type="pil")
|
| 60 |
-
return output.images[0]
|
| 61 |
|
| 62 |
def attack_detect(self, img, jpeg_compression, downscale, crop, saturation):
|
| 63 |
|
| 64 |
-
#img = img_edit["composite"]
|
| 65 |
img = img.convert("RGB")
|
| 66 |
|
| 67 |
# attack
|
|
@@ -69,6 +62,7 @@ class BZHStableSignatureDemo(object):
|
|
| 69 |
size = img.size
|
| 70 |
size = (int(size[0] / downscale), int(size[1] / downscale))
|
| 71 |
img = img.resize(size, Image.Resampling.LANCZOS)
|
|
|
|
| 72 |
if crop != 0:
|
| 73 |
width, height = img.size
|
| 74 |
area = width * height
|
|
@@ -108,17 +102,15 @@ class BZHStableSignatureDemo(object):
|
|
| 108 |
|
| 109 |
mf.seek(0)
|
| 110 |
img0 = Image.open(mf) # reload to show JPEG attack
|
| 111 |
-
|
| 112 |
result = "No watermark detected."
|
| 113 |
-
chances = int(1 / pvalue + 1)
|
| 114 |
rpv = 10**int(math.log10(pvalue))
|
| 115 |
if pvalue < 1e-3:
|
| 116 |
-
result = "Watermark detected with low confidence (p-value<%.0e)" % rpv
|
| 117 |
if pvalue < 1e-9:
|
| 118 |
-
result = "Watermark detected with high confidence (p-value<%.0e)" % rpv
|
| 119 |
return (img0, result)
|
| 120 |
|
| 121 |
-
|
| 122 |
def interface():
|
| 123 |
prompt = "sailing ship in storm by Rembrandt"
|
| 124 |
|
|
|
|
| 7 |
|
| 8 |
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
|
| 9 |
|
| 10 |
+
from diffusers import DiffusionPipeline, AutoencoderKL
|
| 11 |
import torchvision.transforms as transforms
|
| 12 |
|
| 13 |
from copy import deepcopy
|
|
|
|
| 26 |
|
| 27 |
def __init__(self, *args, **kwargs):
|
| 28 |
super().__init__(*args, **kwargs)
|
| 29 |
+
|
| 30 |
self.pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16").to("cuda")
|
| 31 |
+
|
| 32 |
+
# disable invisible-watermark
|
| 33 |
+
self.pipe.watermark = None
|
| 34 |
+
|
| 35 |
+
# save the original VAE
|
| 36 |
+
decoders = OrderedDict([("no watermark", self.pipe.vae)])
|
| 37 |
+
|
| 38 |
+
# load the patched VAEs
|
| 39 |
+
for name in ("weak", "medium", "strong", "extreme"):
|
| 40 |
+
vae = AutoencoderKL.from_pretrained(f"imatag/stable-signature-bzh-sdxl-vae-{name}", torch_dtype=torch.float16).to("cuda")
|
| 41 |
+
decoders[name] = vae
|
| 42 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
self.decoders = decoders
|
| 44 |
|
| 45 |
def generate(self, mode, seed, prompt):
|
| 46 |
generator = torch.Generator(device=device)
|
|
|
|
| 47 |
torch.manual_seed(seed)
|
| 48 |
|
| 49 |
+
# load the patched VAE
|
| 50 |
+
vae = self.decoders[mode]
|
| 51 |
+
self.pipe.vae = vae
|
| 52 |
|
| 53 |
output = self.pipe(prompt, num_inference_steps=4, guidance_scale=0.0, output_type="pil")
|
| 54 |
+
return output.images[0]
|
| 55 |
|
| 56 |
def attack_detect(self, img, jpeg_compression, downscale, crop, saturation):
|
| 57 |
|
|
|
|
| 58 |
img = img.convert("RGB")
|
| 59 |
|
| 60 |
# attack
|
|
|
|
| 62 |
size = img.size
|
| 63 |
size = (int(size[0] / downscale), int(size[1] / downscale))
|
| 64 |
img = img.resize(size, Image.Resampling.LANCZOS)
|
| 65 |
+
|
| 66 |
if crop != 0:
|
| 67 |
width, height = img.size
|
| 68 |
area = width * height
|
|
|
|
| 102 |
|
| 103 |
mf.seek(0)
|
| 104 |
img0 = Image.open(mf) # reload to show JPEG attack
|
| 105 |
+
|
| 106 |
result = "No watermark detected."
|
|
|
|
| 107 |
rpv = 10**int(math.log10(pvalue))
|
| 108 |
if pvalue < 1e-3:
|
| 109 |
+
result = "Watermark detected with low confidence (p-value<%.0e)" % rpv
|
| 110 |
if pvalue < 1e-9:
|
| 111 |
+
result = "Watermark detected with high confidence (p-value<%.0e)" % rpv
|
| 112 |
return (img0, result)
|
| 113 |
|
|
|
|
| 114 |
def interface():
|
| 115 |
prompt = "sailing ship in storm by Rembrandt"
|
| 116 |
|