File size: 12,152 Bytes
bb7edb9
 
 
2b755c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c1b5cda
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3494400
c1b5cda
 
 
3494400
c1b5cda
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d955f0e
c1b5cda
 
 
 
2b755c2
 
 
 
 
 
be7a25b
2b755c2
 
 
bb7edb9
 
 
 
 
 
 
 
 
 
 
2b755c2
 
 
bb7edb9
 
 
2b755c2
 
 
 
bb7edb9
2b755c2
bb7edb9
 
 
 
2b755c2
 
c1b5cda
2b755c2
 
bb7edb9
2b755c2
 
 
bb7edb9
 
 
 
2b755c2
 
 
 
 
bb7edb9
2b755c2
bb7edb9
 
 
 
2b755c2
 
 
 
 
bb7edb9
 
 
 
2b755c2
 
 
 
bb7edb9
2b755c2
 
 
bb7edb9
 
 
 
2b755c2
 
1878ed0
2b755c2
 
bb7edb9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2b755c2
 
 
 
 
 
 
 
 
bb7edb9
2b755c2
 
 
 
bb7edb9
2b755c2
 
 
 
 
 
 
 
 
8f873ac
2b755c2
8f873ac
2b755c2
 
 
 
3494400
 
 
5e63b46
2b755c2
bb7edb9
2b755c2
 
bb7edb9
2b755c2
 
bb7edb9
2b755c2
 
bb7edb9
 
 
 
 
 
 
 
 
2b755c2
 
bb7edb9
 
 
 
 
2b755c2
 
 
 
 
bb7edb9
 
 
2b755c2
 
 
 
 
 
 
 
 
 
 
 
 
bb7edb9
2b755c2
c628a76
 
 
2b755c2
 
 
bb7edb9
 
2b755c2
 
 
 
 
 
bb7edb9
 
 
2b755c2
 
 
c1b5cda
 
2b755c2
 
 
 
 
 
 
336be84
2b755c2
 
 
 
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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
import gc
import os
from abc import ABC, abstractmethod

import PIL.Image
import torch
from controlnet_aux import (
    CannyDetector,
    LineartDetector,
    MidasDetector,
    PidiNetDetector,
    ZoeDetector,
)
from diffusers import (
    AutoencoderKL,
    EulerAncestralDiscreteScheduler,
    StableDiffusionXLAdapterPipeline,
    T2IAdapter,
)

SD_XL_BASE_RATIOS = {
    "0.5": (704, 1408),
    "0.52": (704, 1344),
    "0.57": (768, 1344),
    "0.6": (768, 1280),
    "0.68": (832, 1216),
    "0.72": (832, 1152),
    "0.78": (896, 1152),
    "0.82": (896, 1088),
    "0.88": (960, 1088),
    "0.94": (960, 1024),
    "1.0": (1024, 1024),
    "1.07": (1024, 960),
    "1.13": (1088, 960),
    "1.21": (1088, 896),
    "1.29": (1152, 896),
    "1.38": (1152, 832),
    "1.46": (1216, 832),
    "1.67": (1280, 768),
    "1.75": (1344, 768),
    "1.91": (1344, 704),
    "2.0": (1408, 704),
    "2.09": (1472, 704),
    "2.4": (1536, 640),
    "2.5": (1600, 640),
    "2.89": (1664, 576),
    "3.0": (1728, 576),
}


def find_closest_aspect_ratio(target_width, target_height):
    target_ratio = target_width / target_height
    closest_ratio = None
    min_difference = float("inf")

    for ratio_str, (width, height) in SD_XL_BASE_RATIOS.items():
        ratio = width / height
        difference = abs(target_ratio - ratio)

        if difference < min_difference:
            min_difference = difference
            closest_ratio = ratio_str

    return closest_ratio


def resize_to_closest_aspect_ratio(image):
    target_width, target_height = image.size
    closest_ratio = find_closest_aspect_ratio(target_width, target_height)

    # Get the dimensions from the closest aspect ratio in the dictionary
    new_width, new_height = SD_XL_BASE_RATIOS[closest_ratio]

    # Resize the image to the new dimensions while preserving the aspect ratio
    resized_image = image.resize((new_width, new_height), PIL.Image.LANCZOS)

    return resized_image


ADAPTER_NAMES = [
    "TencentARC/t2i-adapter-canny-sdxl-1.0",
    "TencentARC/t2i-adapter-sketch-sdxl-1.0",
    "TencentARC/t2i-adapter-lineart-sdxl-1.0",
    "TencentARC/t2i-adapter-depth-midas-sdxl-1.0",
    "TencentARC/t2i-adapter-depth-zoe-sdxl-1.0",
    # "TencentARC/t2i-adapter-recolor-sdxl-1.0",
]


class Preprocessor(ABC):
    @abstractmethod
    def to(self, device: torch.device | str) -> "Preprocessor":
        pass

    @abstractmethod
    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        pass


class CannyPreprocessor(Preprocessor):
    def __init__(self):
        self.model = CannyDetector()

    def to(self, device: torch.device | str) -> Preprocessor:
        return self

    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        return self.model(image, detect_resolution=384, image_resolution=1024)


class LineartPreprocessor(Preprocessor):
    def __init__(self):
        self.model = LineartDetector.from_pretrained("lllyasviel/Annotators")

    def to(self, device: torch.device | str) -> Preprocessor:
        return self.model.to(device)

    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        return self.model(image, detect_resolution=512, image_resolution=1024)


class MidasPreprocessor(Preprocessor):
    def __init__(self):
        self.model = MidasDetector.from_pretrained(
            "valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large"
        )

    def to(self, device: torch.device | str) -> Preprocessor:
        return self.model.to(device)

    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        return self.model(image, detect_resolution=512, image_resolution=1024)


class PidiNetPreprocessor(Preprocessor):
    def __init__(self):
        self.model = PidiNetDetector.from_pretrained("lllyasviel/Annotators")

    def to(self, device: torch.device | str) -> Preprocessor:
        return self.model.to(device)

    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        return self.model(image, detect_resolution=512, image_resolution=1024, apply_filter=True)


class RecolorPreprocessor(Preprocessor):
    def to(self, device: torch.device | str) -> Preprocessor:
        return self

    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        return image.convert("L").convert("RGB")


class ZoePreprocessor(Preprocessor):
    def __init__(self):
        self.model = ZoeDetector.from_pretrained(
            "valhalla/t2iadapter-aux-models", filename="zoed_nk.pth", model_type="zoedepth_nk"
        )

    def to(self, device: torch.device | str) -> Preprocessor:
        return self.model.to(device)

    def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
        return self.model(image, gamma_corrected=True, image_resolution=1024)


PRELOAD_PREPROCESSORS_IN_GPU_MEMORY = os.getenv("PRELOAD_PREPROCESSORS_IN_GPU_MEMORY", "1") == "1"
PRELOAD_PREPROCESSORS_IN_CPU_MEMORY = os.getenv("PRELOAD_PREPROCESSORS_IN_CPU_MEMORY", "0") == "1"
if PRELOAD_PREPROCESSORS_IN_GPU_MEMORY:
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    preprocessors_gpu: dict[str, Preprocessor] = {
        "TencentARC/t2i-adapter-canny-sdxl-1.0": CannyPreprocessor().to(device),
        "TencentARC/t2i-adapter-sketch-sdxl-1.0": PidiNetPreprocessor().to(device),
        "TencentARC/t2i-adapter-lineart-sdxl-1.0": LineartPreprocessor().to(device),
        "TencentARC/t2i-adapter-depth-midas-sdxl-1.0": MidasPreprocessor().to(device),
        "TencentARC/t2i-adapter-depth-zoe-sdxl-1.0": ZoePreprocessor().to(device),
        "TencentARC/t2i-adapter-recolor-sdxl-1.0": RecolorPreprocessor().to(device),
    }

    def get_preprocessor(adapter_name: str) -> Preprocessor:
        return preprocessors_gpu[adapter_name]

elif PRELOAD_PREPROCESSORS_IN_CPU_MEMORY:
    preprocessors_cpu: dict[str, Preprocessor] = {
        "TencentARC/t2i-adapter-canny-sdxl-1.0": CannyPreprocessor(),
        "TencentARC/t2i-adapter-sketch-sdxl-1.0": PidiNetPreprocessor(),
        "TencentARC/t2i-adapter-lineart-sdxl-1.0": LineartPreprocessor(),
        "TencentARC/t2i-adapter-depth-midas-sdxl-1.0": MidasPreprocessor(),
        "TencentARC/t2i-adapter-depth-zoe-sdxl-1.0": ZoePreprocessor(),
        "TencentARC/t2i-adapter-recolor-sdxl-1.0": RecolorPreprocessor(),
    }

    def get_preprocessor(adapter_name: str) -> Preprocessor:
        return preprocessors_cpu[adapter_name]

else:

    def get_preprocessor(adapter_name: str) -> Preprocessor:
        if adapter_name == "TencentARC/t2i-adapter-canny-sdxl-1.0":
            return CannyPreprocessor()
        elif adapter_name == "TencentARC/t2i-adapter-sketch-sdxl-1.0":
            return PidiNetPreprocessor()
        elif adapter_name == "TencentARC/t2i-adapter-lineart-sdxl-1.0":
            return LineartPreprocessor()
        elif adapter_name == "TencentARC/t2i-adapter-depth-midas-sdxl-1.0":
            return MidasPreprocessor()
        elif adapter_name == "TencentARC/t2i-adapter-depth-zoe-sdxl-1.0":
            return ZoePreprocessor()
        elif adapter_name == "TencentARC/t2i-adapter-recolor-sdxl-1.0":
            return RecolorPreprocessor()
        else:
            raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")

    def download_all_preprocessors():
        for adapter_name in ADAPTER_NAMES:
            get_preprocessor(adapter_name)
        gc.collect()

    download_all_preprocessors()


def download_all_adapters():
    for adapter_name in ADAPTER_NAMES:
        T2IAdapter.from_pretrained(
            adapter_name,
            torch_dtype=torch.float16,
            varient="fp16",
        )
    gc.collect()


download_all_adapters()


class Model:
    MAX_NUM_INFERENCE_STEPS = 50

    def __init__(self, adapter_name: str):
        if adapter_name not in ADAPTER_NAMES:
            raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")

        self.preprocessor_name = adapter_name
        self.adapter_name = adapter_name

        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        if torch.cuda.is_available():
            self.preprocessor = get_preprocessor(adapter_name).to(self.device)

            model_id = "stabilityai/stable-diffusion-xl-base-1.0"
            adapter = T2IAdapter.from_pretrained(
                adapter_name,
                torch_dtype=torch.float16,
                varient="fp16",
            ).to(self.device)
            self.pipe = StableDiffusionXLAdapterPipeline.from_pretrained(
                model_id,
                vae=AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16),
                adapter=adapter,
                scheduler=EulerAncestralDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler"),
                torch_dtype=torch.float16,
                variant="fp16",
            ).to(self.device)
            self.pipe.enable_xformers_memory_efficient_attention()
            self.pipe.load_lora_weights(
                "stabilityai/stable-diffusion-xl-base-1.0", weight_name="sd_xl_offset_example-lora_1.0.safetensors"
            )
            self.pipe.fuse_lora(lora_scale=0.4)
        else:
            self.preprocessor = None  # type: ignore
            self.pipe = None

    def change_preprocessor(self, adapter_name: str) -> None:
        if adapter_name not in ADAPTER_NAMES:
            raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
        if adapter_name == self.preprocessor_name:
            return

        if PRELOAD_PREPROCESSORS_IN_GPU_MEMORY:
            pass
        elif PRELOAD_PREPROCESSORS_IN_CPU_MEMORY:
            self.preprocessor.to("cpu")
        else:
            del self.preprocessor
        self.preprocessor = get_preprocessor(adapter_name).to(self.device)
        self.preprocessor_name = adapter_name
        gc.collect()
        torch.cuda.empty_cache()

    def change_adapter(self, adapter_name: str) -> None:
        if adapter_name not in ADAPTER_NAMES:
            raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
        if adapter_name == self.adapter_name:
            return
        self.pipe.adapter = T2IAdapter.from_pretrained(
            adapter_name,
            torch_dtype=torch.float16,
            varient="fp16",
        ).to(self.device)
        self.adapter_name = adapter_name
        gc.collect()
        torch.cuda.empty_cache()

    def resize_image(self, image: PIL.Image.Image) -> PIL.Image.Image:
        w, h = image.size
        scale = 1024 / max(w, h)
        new_w = int(w * scale)
        new_h = int(h * scale)
        return image.resize((new_w, new_h), PIL.Image.LANCZOS)

    def run(
        self,
        image: PIL.Image.Image,
        prompt: str,
        negative_prompt: str,
        adapter_name: str,
        num_inference_steps: int = 30,
        guidance_scale: float = 5.0,
        adapter_conditioning_scale: float = 1.0,
        cond_tau: float = 1.0,
        seed: int = 0,
        apply_preprocess: bool = True,
    ) -> list[PIL.Image.Image]:
        if not torch.cuda.is_available():
            raise RuntimeError("This demo does not work on CPU.")
        if num_inference_steps > self.MAX_NUM_INFERENCE_STEPS:
            raise ValueError(f"Number of steps must be less than {self.MAX_NUM_INFERENCE_STEPS}")

        # Resize image to avoid OOM
        image = self.resize_image(image)

        self.change_preprocessor(adapter_name)
        self.change_adapter(adapter_name)

        if apply_preprocess:
            image = self.preprocessor(image)

        image = resize_to_closest_aspect_ratio(image)

        generator = torch.Generator(device=self.device).manual_seed(seed)
        out = self.pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            image=image,
            num_inference_steps=num_inference_steps,
            adapter_conditioning_scale=adapter_conditioning_scale,
            adapter_conditioning_factor=cond_tau,
            generator=generator,
            guidance_scale=guidance_scale,
        ).images[0]
        return [image, out]