Improve image transformation
#1
by
patrickvonplaten
- opened
- image_transformation.py +12 -36
image_transformation.py
CHANGED
|
@@ -15,7 +15,7 @@ if is_vision_available():
|
|
| 15 |
from PIL import Image
|
| 16 |
|
| 17 |
if is_diffusers_available():
|
| 18 |
-
from diffusers import
|
| 19 |
|
| 20 |
if is_opencv_available():
|
| 21 |
import cv2
|
|
@@ -29,8 +29,7 @@ IMAGE_TRANSFORMATION_DESCRIPTION = (
|
|
| 29 |
|
| 30 |
|
| 31 |
class ImageTransformationTool(Tool):
|
| 32 |
-
default_stable_diffusion_checkpoint = "
|
| 33 |
-
default_controlnet_checkpoint = "lllyasviel/sd-controlnet-canny"
|
| 34 |
description = IMAGE_TRANSFORMATION_DESCRIPTION
|
| 35 |
inputs = ['image', 'text']
|
| 36 |
outputs = ['image']
|
|
@@ -47,13 +46,7 @@ class ImageTransformationTool(Tool):
|
|
| 47 |
|
| 48 |
super().__init__()
|
| 49 |
|
| 50 |
-
|
| 51 |
-
controlnet = self.default_controlnet_checkpoint
|
| 52 |
-
self.controlnet_checkpoint = controlnet
|
| 53 |
-
|
| 54 |
-
if stable_diffusion is None:
|
| 55 |
-
stable_diffusion = self.default_stable_diffusion_checkpoint
|
| 56 |
-
self.stable_diffusion_checkpoint = stable_diffusion
|
| 57 |
|
| 58 |
self.device = device
|
| 59 |
self.hub_kwargs = hub_kwargs
|
|
@@ -62,37 +55,20 @@ class ImageTransformationTool(Tool):
|
|
| 62 |
if self.device is None:
|
| 63 |
self.device = get_default_device()
|
| 64 |
|
| 65 |
-
self.
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
self.pipeline.enable_model_cpu_offload()
|
| 71 |
|
| 72 |
self.is_initialized = True
|
| 73 |
|
| 74 |
-
def __call__(self, image, prompt):
|
| 75 |
if not self.is_initialized:
|
| 76 |
self.setup()
|
| 77 |
|
| 78 |
-
initial_prompt = "super-hero character, best quality, extremely detailed"
|
| 79 |
-
prompt = initial_prompt + prompt
|
| 80 |
-
|
| 81 |
-
low_threshold = 100
|
| 82 |
-
high_threshold = 200
|
| 83 |
-
|
| 84 |
-
image = np.array(image)
|
| 85 |
-
image = cv2.Canny(image, low_threshold, high_threshold)
|
| 86 |
-
image = image[:, :, None]
|
| 87 |
-
image = np.concatenate([image, image, image], axis=2)
|
| 88 |
-
canny_image = Image.fromarray(image)
|
| 89 |
-
|
| 90 |
-
generator = torch.Generator(device="cpu").manual_seed(2)
|
| 91 |
-
|
| 92 |
return self.pipeline(
|
| 93 |
-
prompt,
|
| 94 |
-
|
| 95 |
-
negative_prompt=
|
| 96 |
-
num_inference_steps=20,
|
| 97 |
-
generator=generator,
|
| 98 |
).images[0]
|
|
|
|
| 15 |
from PIL import Image
|
| 16 |
|
| 17 |
if is_diffusers_available():
|
| 18 |
+
from diffusers import DiffusionPipeline
|
| 19 |
|
| 20 |
if is_opencv_available():
|
| 21 |
import cv2
|
|
|
|
| 29 |
|
| 30 |
|
| 31 |
class ImageTransformationTool(Tool):
|
| 32 |
+
default_stable_diffusion_checkpoint = "timbrooks/instruct-pix2pix"
|
|
|
|
| 33 |
description = IMAGE_TRANSFORMATION_DESCRIPTION
|
| 34 |
inputs = ['image', 'text']
|
| 35 |
outputs = ['image']
|
|
|
|
| 46 |
|
| 47 |
super().__init__()
|
| 48 |
|
| 49 |
+
self.stable_diffusion = self.default_stable_diffusion_checkpoint
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
self.device = device
|
| 52 |
self.hub_kwargs = hub_kwargs
|
|
|
|
| 55 |
if self.device is None:
|
| 56 |
self.device = get_default_device()
|
| 57 |
|
| 58 |
+
self.pipeline = DiffusionPipeline.from_pretrained(self.stable_diffusion)
|
| 59 |
+
|
| 60 |
+
self.pipeline.to(self.device)
|
| 61 |
+
if self.device.type == "cuda":
|
| 62 |
+
self.pipeline.to(torch_dtype=torch.float16)
|
|
|
|
| 63 |
|
| 64 |
self.is_initialized = True
|
| 65 |
|
| 66 |
+
def __call__(self, image, prompt, negative_prompt="low quality, bad quality, deformed, low resolution", added_prompt=" , highest quality, highly realistic, very high resolution"):
|
| 67 |
if not self.is_initialized:
|
| 68 |
self.setup()
|
| 69 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
return self.pipeline(
|
| 71 |
+
prompt + added_prompt,
|
| 72 |
+
image,
|
| 73 |
+
negative_prompt=negative_prompt,
|
|
|
|
|
|
|
| 74 |
).images[0]
|