1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
|
"""Diffusers-backed generation and on-demand ControlNet management."""
from __future__ import annotations
import gc
from typing import Any
import numpy as np
from PIL import Image
from .config import AppConfig
from .errors import GenerationError
from .pipeline_factory import PipelineFactory
from .preprocessors.depth import DepthPreprocessor
from .preprocessors.openpose import OpenPosePreprocessor
from .schema import GuidedImageRequest, InpaintImageRequest, TextToImageRequest
class ModelManager:
"""Own all model state; construction validates model compatibility at startup."""
def __init__(self, config: AppConfig) -> None:
self.config = config
self.factory = PipelineFactory(config)
self.base = self.factory.create_base()
self._control_mode: str | None = None
self._controlnet: Any | None = None
self._control_pipeline: Any | None = None
self._depth = DepthPreprocessor()
self._openpose = OpenPosePreprocessor()
def text_to_image(self, request: TextToImageRequest, seed: int) -> Image.Image:
width, height = request.size.dimensions
prompt_embeds, negative_prompt_embeds = self._embed_prompts(
request.positive_prompt, request.negative_prompt
)
result = self.base(
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
width=width,
height=height,
num_inference_steps=self.config.generation.steps,
guidance_scale=self.config.generation.cfg_scale,
num_images_per_prompt=1,
generator=self._generator(seed),
)
return self._one_image(result)
def guided_image(
self, request: GuidedImageRequest, image: Image.Image, seed: int
) -> Image.Image:
width, height = request.size.dimensions
image = image.resize((width, height), Image.Resampling.LANCZOS)
if request.control_mode == "depth":
control_image = self._depth(image)
else:
control_image = self._openpose(image, full=request.control_mode == "openpose_full")
pipeline = self._pipeline_for(str(request.control_mode), inpaint=False)
prompt_embeds, negative_prompt_embeds = self._embed_prompts(
request.positive_prompt, request.negative_prompt
)
result = pipeline(
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
image=control_image,
width=width,
height=height,
num_inference_steps=self.config.generation.steps,
guidance_scale=self.config.generation.cfg_scale,
num_images_per_prompt=1,
generator=self._generator(seed),
)
return self._one_image(result)
def inpaint_image(
self,
request: InpaintImageRequest,
image: Image.Image,
mask: Image.Image,
seed: int,
) -> Image.Image:
width, height = request.size.dimensions
image = image.resize((width, height), Image.Resampling.LANCZOS)
mask = mask.resize((width, height), Image.Resampling.NEAREST)
control_image = self._inpaint_condition(image, mask)
pipeline = self._pipeline_for("inpaint", inpaint=True)
prompt_embeds, negative_prompt_embeds = self._embed_prompts(
request.positive_prompt, request.negative_prompt
)
result = pipeline(
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
image=image,
mask_image=mask,
control_image=control_image,
width=width,
height=height,
num_inference_steps=self.config.generation.steps,
guidance_scale=self.config.generation.cfg_scale,
num_images_per_prompt=1,
generator=self._generator(seed),
)
return self._one_image(result)
def _pipeline_for(self, mode: str, *, inpaint: bool) -> Any:
model_mode = "openpose" if mode == "openpose_full" else mode
if self._control_mode == model_mode and self._control_pipeline is not None:
return self._control_pipeline
self._release_controlnet()
self._controlnet = self.factory.create_controlnet(self.config.controlnet_path(model_mode))
self._control_pipeline = (
self.factory.create_inpaint_pipeline(self.base, self._controlnet)
if inpaint
else self.factory.create_guided_pipeline(self.base, self._controlnet)
)
self._control_mode = model_mode
return self._control_pipeline
def _release_controlnet(self) -> None:
self._control_pipeline = None
self._controlnet = None
self._control_mode = None
gc.collect()
try:
import torch
torch.cuda.empty_cache()
except ImportError:
pass
@staticmethod
def _generator(seed: int) -> Any:
import torch
return torch.Generator(device="cuda").manual_seed(seed)
@staticmethod
def _inpaint_condition(image: Image.Image, mask: Image.Image) -> Any:
import torch
image_array = np.asarray(image.convert("RGB"), dtype=np.float32) / 255.0
mask_array = np.asarray(mask.convert("L"), dtype=np.float32) / 255.0
image_array[mask_array > 0.5] = -1.0
return torch.from_numpy(image_array[None].transpose(0, 3, 1, 2))
@staticmethod
def _one_image(result: Any) -> Image.Image:
images = getattr(result, "images", None)
if not images or len(images) != 1 or not isinstance(images[0], Image.Image):
raise GenerationError("Diffusers did not return exactly one image")
return images[0]
def _embed_prompts(self, positive: str, negative: str) -> tuple[Any, Any]:
"""Generate prompt embeds via sd_embed to bypass the 77-token limit."""
from sd_embed.embedding_funcs import get_weighted_text_embeddings_sd15
prompt_embeds, neg_prompt_embeds = get_weighted_text_embeddings_sd15(
self.base,
prompt=positive,
neg_prompt=negative or "",
pad_last_block=self.config.generation.pad_last_block,
clip_skip=self.config.generation.clip_skip,
)
return prompt_embeds, neg_prompt_embeds
|
