Update handler.py
Browse files- handler.py +85 -74
handler.py
CHANGED
@@ -1,116 +1,127 @@
|
|
1 |
-
from typing import List,
|
2 |
import base64
|
3 |
from PIL import Image
|
4 |
from io import BytesIO
|
5 |
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
|
6 |
import torch
|
|
|
|
|
|
|
|
|
7 |
import controlnet_hinter
|
8 |
|
9 |
# set device
|
10 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
11 |
if device.type != 'cuda':
|
12 |
-
raise ValueError("
|
13 |
# set mixed precision dtype
|
14 |
dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
|
15 |
|
16 |
-
# controlnet mapping for
|
17 |
CONTROLNET_MAPPING = {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
"depth": {
|
19 |
"model_id": "lllyasviel/sd-controlnet-depth",
|
20 |
"hinter": controlnet_hinter.hint_depth
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
}
|
22 |
}
|
23 |
|
|
|
24 |
class EndpointHandler():
|
25 |
def __init__(self, path=""):
|
26 |
# define default controlnet id and load controlnet
|
27 |
self.control_type = "depth"
|
28 |
-
self.controlnet = ControlNetModel.from_pretrained(CONTROLNET_MAPPING[self.control_type]["model_id"],
|
29 |
-
|
30 |
-
# Load StableDiffusionControlNetPipeline
|
31 |
self.stable_diffusion_id = "runwayml/stable-diffusion-v1-5"
|
32 |
-
self.pipe = StableDiffusionControlNetPipeline.from_pretrained(self.stable_diffusion_id,
|
33 |
-
controlnet=self.controlnet,
|
34 |
torch_dtype=dtype,
|
35 |
safety_checker=None).to(device)
|
36 |
# Define Generator with seed
|
37 |
self.generator = torch.Generator(device="cpu").manual_seed(3)
|
38 |
|
39 |
-
def __call__(self, data: Any) -> Dict[str,
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
"cfg_scale": 7,
|
48 |
-
"alwayson_scripts": {
|
49 |
-
"controlnet": {
|
50 |
-
"args": [
|
51 |
-
{
|
52 |
-
"enabled": True,
|
53 |
-
"input_image": "image in base64",
|
54 |
-
"model": "control_sd15_depth [fef5e48e]",
|
55 |
-
"control_mode": "Balanced"
|
56 |
-
}
|
57 |
-
]
|
58 |
-
}
|
59 |
-
}
|
60 |
-
}
|
61 |
-
|
62 |
-
# Extract parameters from the payload
|
63 |
-
prompt = data.get("prompt", None)
|
64 |
-
negative_prompt = data.get("negative_prompt", None)
|
65 |
-
width = data.get("width", None)
|
66 |
-
height = data.get("height", None)
|
67 |
-
num_inference_steps = data.get("steps", 30)
|
68 |
-
guidance_scale = data.get("cfg_scale", 7)
|
69 |
|
70 |
-
#
|
71 |
-
|
72 |
-
|
73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
out = self.pipe(
|
75 |
-
prompt=prompt,
|
76 |
negative_prompt=negative_prompt,
|
77 |
-
|
|
|
78 |
guidance_scale=guidance_scale,
|
79 |
num_images_per_prompt=1,
|
80 |
height=height,
|
81 |
width=width,
|
82 |
-
controlnet_conditioning_scale=
|
83 |
-
generator=self.generator
|
84 |
)
|
85 |
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
input_image_base64 = controlnet_config.get("input_image", "")
|
92 |
-
input_image = self.decode_base64_image(input_image_base64)
|
93 |
-
controlnet_model = controlnet_config.get("model", "")
|
94 |
-
controlnet_control_mode = controlnet_config.get("control_mode", "")
|
95 |
-
|
96 |
-
processed_image = self.process_with_controlnet(generated_image, input_image, controlnet_model, controlnet_control_mode)
|
97 |
-
else:
|
98 |
-
processed_image = generated_image
|
99 |
-
|
100 |
-
# Return the final processed image as base64
|
101 |
-
return {"image": self.encode_base64_image(processed_image)}
|
102 |
-
|
103 |
-
def process_with_controlnet(self, generated_image, input_image, model, control_mode):
|
104 |
-
# Simulated controlnet processing (replace with actual implementation)
|
105 |
-
# Here, we're just using the input_image as-is. Replace this with your controlnet logic.
|
106 |
-
return input_image
|
107 |
-
|
108 |
-
def encode_base64_image(self, image):
|
109 |
-
# Encode the PIL Image to base64
|
110 |
-
buffer = BytesIO()
|
111 |
-
image.save(buffer, format="PNG")
|
112 |
-
return base64.b64encode(buffer.getvalue()).decode("utf-8")
|
113 |
-
|
114 |
def decode_base64_image(self, image_string):
|
115 |
base64_image = base64.b64decode(image_string)
|
116 |
buffer = BytesIO(base64_image)
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
import base64
|
3 |
from PIL import Image
|
4 |
from io import BytesIO
|
5 |
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
|
6 |
import torch
|
7 |
+
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
import cv2
|
11 |
import controlnet_hinter
|
12 |
|
13 |
# set device
|
14 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
15 |
if device.type != 'cuda':
|
16 |
+
raise ValueError("need to run on GPU")
|
17 |
# set mixed precision dtype
|
18 |
dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
|
19 |
|
20 |
+
# controlnet mapping for controlnet id and control hinter
|
21 |
CONTROLNET_MAPPING = {
|
22 |
+
"canny_edge": {
|
23 |
+
"model_id": "lllyasviel/sd-controlnet-canny",
|
24 |
+
"hinter": controlnet_hinter.hint_canny
|
25 |
+
},
|
26 |
+
"pose": {
|
27 |
+
"model_id": "lllyasviel/sd-controlnet-openpose",
|
28 |
+
"hinter": controlnet_hinter.hint_openpose
|
29 |
+
},
|
30 |
"depth": {
|
31 |
"model_id": "lllyasviel/sd-controlnet-depth",
|
32 |
"hinter": controlnet_hinter.hint_depth
|
33 |
+
},
|
34 |
+
"scribble": {
|
35 |
+
"model_id": "lllyasviel/sd-controlnet-scribble",
|
36 |
+
"hinter": controlnet_hinter.hint_scribble,
|
37 |
+
},
|
38 |
+
"segmentation": {
|
39 |
+
"model_id": "lllyasviel/sd-controlnet-seg",
|
40 |
+
"hinter": controlnet_hinter.hint_segmentation,
|
41 |
+
},
|
42 |
+
"normal": {
|
43 |
+
"model_id": "lllyasviel/sd-controlnet-normal",
|
44 |
+
"hinter": controlnet_hinter.hint_normal,
|
45 |
+
},
|
46 |
+
"hed": {
|
47 |
+
"model_id": "lllyasviel/sd-controlnet-hed",
|
48 |
+
"hinter": controlnet_hinter.hint_hed,
|
49 |
+
},
|
50 |
+
"hough": {
|
51 |
+
"model_id": "lllyasviel/sd-controlnet-mlsd",
|
52 |
+
"hinter": controlnet_hinter.hint_hough,
|
53 |
}
|
54 |
}
|
55 |
|
56 |
+
|
57 |
class EndpointHandler():
|
58 |
def __init__(self, path=""):
|
59 |
# define default controlnet id and load controlnet
|
60 |
self.control_type = "depth"
|
61 |
+
self.controlnet = ControlNetModel.from_pretrained(CONTROLNET_MAPPING[self.control_type]["model_id"],torch_dtype=dtype).to(device)
|
62 |
+
|
63 |
+
# Load StableDiffusionControlNetPipeline
|
64 |
self.stable_diffusion_id = "runwayml/stable-diffusion-v1-5"
|
65 |
+
self.pipe = StableDiffusionControlNetPipeline.from_pretrained(self.stable_diffusion_id,
|
66 |
+
controlnet=self.controlnet,
|
67 |
torch_dtype=dtype,
|
68 |
safety_checker=None).to(device)
|
69 |
# Define Generator with seed
|
70 |
self.generator = torch.Generator(device="cpu").manual_seed(3)
|
71 |
|
72 |
+
def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
|
73 |
+
"""
|
74 |
+
:param data: A dictionary contains `inputs` and optional `image` field.
|
75 |
+
:return: A dictionary with `image` field contains image in base64.
|
76 |
+
"""
|
77 |
+
prompt = data.pop("inputs", None)
|
78 |
+
image = data.pop("image", None)
|
79 |
+
controlnet_type = data.pop("controlnet_type", None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
|
81 |
+
# Check if neither prompt nor image is provided
|
82 |
+
if prompt is None and image is None:
|
83 |
+
return {"error": "Please provide a prompt and base64 encoded image."}
|
84 |
+
|
85 |
+
# Check if a new controlnet is provided
|
86 |
+
if controlnet_type is not None and controlnet_type != self.control_type:
|
87 |
+
print(f"changing controlnet from {self.control_type} to {controlnet_type} using {CONTROLNET_MAPPING[controlnet_type]['model_id']} model")
|
88 |
+
self.control_type = controlnet_type
|
89 |
+
self.controlnet = ControlNetModel.from_pretrained(CONTROLNET_MAPPING[self.control_type]["model_id"],
|
90 |
+
torch_dtype=dtype).to(device)
|
91 |
+
self.pipe.controlnet = self.controlnet
|
92 |
+
|
93 |
+
|
94 |
+
# hyperparamters
|
95 |
+
num_inference_steps = data.pop("num_inference_steps", 30)
|
96 |
+
guidance_scale = data.pop("guidance_scale", 7.5)
|
97 |
+
negative_prompt = data.pop("negative_prompt", None)
|
98 |
+
height = data.pop("height", None)
|
99 |
+
width = data.pop("width", None)
|
100 |
+
controlnet_conditioning_scale = data.pop("controlnet_conditioning_scale", 1.0)
|
101 |
+
|
102 |
+
# process image
|
103 |
+
image = self.decode_base64_image(image)
|
104 |
+
control_image = CONTROLNET_MAPPING[self.control_type]["hinter"](image)
|
105 |
+
|
106 |
+
# run inference pipeline
|
107 |
out = self.pipe(
|
108 |
+
prompt=prompt,
|
109 |
negative_prompt=negative_prompt,
|
110 |
+
image=control_image,
|
111 |
+
num_inference_steps=num_inference_steps,
|
112 |
guidance_scale=guidance_scale,
|
113 |
num_images_per_prompt=1,
|
114 |
height=height,
|
115 |
width=width,
|
116 |
+
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
117 |
+
generator=self.generator
|
118 |
)
|
119 |
|
120 |
+
|
121 |
+
# return first generate PIL image
|
122 |
+
return out.images[0]
|
123 |
+
|
124 |
+
# helper to decode input image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
125 |
def decode_base64_image(self, image_string):
|
126 |
base64_image = base64.b64decode(image_string)
|
127 |
buffer = BytesIO(base64_image)
|