Spaces:
Running
on
Zero
Running
on
Zero
Create models.py
Browse files
models.py
ADDED
@@ -0,0 +1,270 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from diffusers import (
|
3 |
+
StableDiffusionXLImg2ImgPipeline,
|
4 |
+
StableDiffusionInpaintPipeline,
|
5 |
+
DDIMScheduler,
|
6 |
+
PNDMScheduler,
|
7 |
+
EulerDiscreteScheduler,
|
8 |
+
DPMSolverMultistepScheduler
|
9 |
+
)
|
10 |
+
from PIL import Image, ImageFilter, ImageEnhance
|
11 |
+
import numpy as np
|
12 |
+
import cv2
|
13 |
+
|
14 |
+
class InteriorDesignerPro:
|
15 |
+
def __init__(self):
|
16 |
+
self.device = torch.device("cuda") # ТОЛЬКО GPU!
|
17 |
+
self.model_name = "RealVisXL V4.0"
|
18 |
+
|
19 |
+
# Проверка GPU
|
20 |
+
gpu_name = torch.cuda.get_device_name(0)
|
21 |
+
self.is_powerful_gpu = any(gpu in gpu_name for gpu in ['A100', 'H100', 'RTX 4090', 'RTX 3090', 'T4', 'A10G'])
|
22 |
+
|
23 |
+
# Основная модель - RealVisXL V4
|
24 |
+
print(f"Loading {self.model_name} on {gpu_name}...")
|
25 |
+
self.pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
|
26 |
+
"SG161222/RealVisXL_V4.0",
|
27 |
+
torch_dtype=torch.float16,
|
28 |
+
use_safetensors=True,
|
29 |
+
variant="fp16"
|
30 |
+
).to(self.device)
|
31 |
+
|
32 |
+
# БЕЗ enable_model_cpu_offload() и enable_vae_slicing() - они замедляют H200!
|
33 |
+
|
34 |
+
# Настройка scheduler для качества
|
35 |
+
self.pipe.scheduler = EulerDiscreteScheduler.from_config(self.pipe.scheduler.config)
|
36 |
+
|
37 |
+
# Inpainting модель
|
38 |
+
try:
|
39 |
+
self.inpaint_pipe = StableDiffusionInpaintPipeline.from_pretrained(
|
40 |
+
"stabilityai/stable-diffusion-2-inpainting",
|
41 |
+
torch_dtype=torch.float16,
|
42 |
+
safety_checker=None,
|
43 |
+
requires_safety_checker=False,
|
44 |
+
local_files_only=False,
|
45 |
+
resume_download=True
|
46 |
+
).to(self.device)
|
47 |
+
print("Inpainting model loaded")
|
48 |
+
except Exception as e:
|
49 |
+
print(f"Warning: Could not load inpainting model: {e}")
|
50 |
+
print("Using img2img as fallback for object removal")
|
51 |
+
self.inpaint_pipe = None
|
52 |
+
|
53 |
+
@torch.inference_mode()
|
54 |
+
def apply_style_pro(self, image, style_name, room_type, strength=0.75, quality="balanced", custom_prompt=None, custom_negative=None):
|
55 |
+
"""Применение стиля к изображению"""
|
56 |
+
from design_styles import DESIGN_STYLES
|
57 |
+
|
58 |
+
# Ресайз для скорости
|
59 |
+
original_size = image.size
|
60 |
+
if image.width > 768 or image.height > 768:
|
61 |
+
image.thumbnail((768, 768), Image.Resampling.LANCZOS)
|
62 |
+
|
63 |
+
if style_name == "custom" and custom_prompt:
|
64 |
+
# Кастомный промпт
|
65 |
+
full_prompt = custom_prompt
|
66 |
+
negative = custom_negative or "low quality, blurry"
|
67 |
+
else:
|
68 |
+
# Предустановленный стиль
|
69 |
+
style = DESIGN_STYLES.get(style_name, DESIGN_STYLES["Современный минимализм"])
|
70 |
+
room_specific = style.get("room_specific", {}).get(room_type, "")
|
71 |
+
full_prompt = f"{style['prompt']}, {room_specific}, {room_type} interior design, professional photo, high quality, 8k, photorealistic"
|
72 |
+
negative = style.get("negative", "low quality, blurry")
|
73 |
+
|
74 |
+
# Настройки качества - оптимизированные для H200
|
75 |
+
quality_settings = {
|
76 |
+
"fast": {"steps": 15, "guidance": 6.0},
|
77 |
+
"balanced": {"steps": 20, "guidance": 7.0},
|
78 |
+
"ultra": {"steps": 30, "guidance": 8.0}
|
79 |
+
}
|
80 |
+
|
81 |
+
settings = quality_settings.get(quality, quality_settings["balanced"])
|
82 |
+
|
83 |
+
# Генерация с SDXL
|
84 |
+
result = self.pipe(
|
85 |
+
prompt=full_prompt,
|
86 |
+
prompt_2=full_prompt, # Для SDXL
|
87 |
+
negative_prompt=negative,
|
88 |
+
negative_prompt_2=negative, # Для SDXL
|
89 |
+
image=image,
|
90 |
+
strength=strength,
|
91 |
+
num_inference_steps=settings["steps"],
|
92 |
+
guidance_scale=settings["guidance"],
|
93 |
+
# SDXL параметры - оптимизированные
|
94 |
+
original_size=(768, 768),
|
95 |
+
target_size=(768, 768)
|
96 |
+
).images[0]
|
97 |
+
|
98 |
+
# Возвращаем к оригинальному размеру если нужно
|
99 |
+
if result.size != original_size and max(original_size) <= 1024:
|
100 |
+
result = result.resize(original_size, Image.Resampling.LANCZOS)
|
101 |
+
|
102 |
+
return result
|
103 |
+
|
104 |
+
def create_variations(self, image, num_variations=4):
|
105 |
+
"""Создание вариаций дизайна"""
|
106 |
+
variations = []
|
107 |
+
base_seed = torch.randint(0, 1000000, (1,)).item()
|
108 |
+
|
109 |
+
# Ресайз для скорости
|
110 |
+
if image.width > 768 or image.height > 768:
|
111 |
+
image.thumbnail((768, 768), Image.Resampling.LANCZOS)
|
112 |
+
|
113 |
+
for i in range(num_variations):
|
114 |
+
torch.manual_seed(base_seed + i)
|
115 |
+
|
116 |
+
var = self.pipe(
|
117 |
+
prompt="interior design variation, same style, different details",
|
118 |
+
prompt_2="interior design variation, same style, different details",
|
119 |
+
image=image,
|
120 |
+
strength=0.4 + (i * 0.05),
|
121 |
+
num_inference_steps=20, # Меньше шагов для скорости
|
122 |
+
guidance_scale=6.0
|
123 |
+
).images[0]
|
124 |
+
|
125 |
+
variations.append(var)
|
126 |
+
|
127 |
+
return variations
|
128 |
+
|
129 |
+
def create_hdr_lighting(self, image, intensity=0.3):
|
130 |
+
"""Улучшение освещения в стиле HDR"""
|
131 |
+
# Конвертируем в numpy
|
132 |
+
img_array = np.array(image)
|
133 |
+
|
134 |
+
# Применяем CLAHE для улучшения контраста
|
135 |
+
lab = cv2.cvtColor(img_array, cv2.COLOR_RGB2LAB)
|
136 |
+
l, a, b = cv2.split(lab)
|
137 |
+
|
138 |
+
clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8,8))
|
139 |
+
l_clahe = clahe.apply(l)
|
140 |
+
|
141 |
+
enhanced_lab = cv2.merge([l_clahe, a, b])
|
142 |
+
enhanced_rgb = cv2.cvtColor(enhanced_lab, cv2.COLOR_LAB2RGB)
|
143 |
+
|
144 |
+
# Смешиваем с оригиналом
|
145 |
+
result = cv2.addWeighted(img_array, 1-intensity, enhanced_rgb, intensity, 0)
|
146 |
+
|
147 |
+
return Image.fromarray(result)
|
148 |
+
|
149 |
+
def enhance_details(self, image):
|
150 |
+
"""Улучшение деталей изображения"""
|
151 |
+
# Увеличиваем резкость
|
152 |
+
enhancer = ImageEnhance.Sharpness(image)
|
153 |
+
sharp = enhancer.enhance(1.5)
|
154 |
+
|
155 |
+
# Немного увеличиваем контраст
|
156 |
+
enhancer = ImageEnhance.Contrast(sharp)
|
157 |
+
contrast = enhancer.enhance(1.1)
|
158 |
+
|
159 |
+
return contrast
|
160 |
+
|
161 |
+
def change_element(self, image, element, value, strength=0.7):
|
162 |
+
"""Изменение отдельного элемента интерьера"""
|
163 |
+
from design_styles import ROOM_ELEMENTS
|
164 |
+
|
165 |
+
# Ресайз для скорости
|
166 |
+
original_size = image.size
|
167 |
+
if image.width > 768 or image.height > 768:
|
168 |
+
image.thumbnail((768, 768), Image.Resampling.LANCZOS)
|
169 |
+
|
170 |
+
element_info = ROOM_ELEMENTS.get(element, {})
|
171 |
+
prompt_add = element_info.get("prompt_add", element.lower())
|
172 |
+
|
173 |
+
prompt = f"interior with {value} {prompt_add}, professional photo"
|
174 |
+
negative = f"old {element}, damaged, ugly"
|
175 |
+
|
176 |
+
result = self.pipe(
|
177 |
+
prompt=prompt,
|
178 |
+
prompt_2=prompt, # Для SDXL
|
179 |
+
negative_prompt=negative,
|
180 |
+
negative_prompt_2=negative,
|
181 |
+
image=image,
|
182 |
+
strength=min(strength, 0.8), # Ограничиваем для скорости
|
183 |
+
num_inference_steps=20, # Оптимизировано для H200
|
184 |
+
guidance_scale=6.0
|
185 |
+
).images[0]
|
186 |
+
|
187 |
+
# Возвращаем к оригинальному размеру
|
188 |
+
if result.size != original_size:
|
189 |
+
result = result.resize(original_size, Image.Resampling.LANCZOS)
|
190 |
+
|
191 |
+
return result
|
192 |
+
|
193 |
+
def create_style_comparison(self, image, styles, quality="fast"):
|
194 |
+
"""Создание сравнения стилей"""
|
195 |
+
results = []
|
196 |
+
|
197 |
+
# Настройки для быстрой генерации
|
198 |
+
steps = 15 if quality == "fast" else 20
|
199 |
+
|
200 |
+
for style in styles:
|
201 |
+
styled = self.apply_style_pro(
|
202 |
+
image,
|
203 |
+
style,
|
204 |
+
"living room", # default
|
205 |
+
strength=0.75,
|
206 |
+
quality=quality
|
207 |
+
)
|
208 |
+
results.append((style, styled))
|
209 |
+
|
210 |
+
return results
|
211 |
+
|
212 |
+
|
213 |
+
class ObjectRemover:
|
214 |
+
"""Класс для удаления объектов - оптимизированный"""
|
215 |
+
|
216 |
+
def __init__(self, inpaint_pipe):
|
217 |
+
self.pipe = inpaint_pipe
|
218 |
+
self.device = torch.device("cuda")
|
219 |
+
|
220 |
+
def remove_objects(self, image, mask):
|
221 |
+
"""Удаление объектов с изображения"""
|
222 |
+
if self.pipe is None:
|
223 |
+
# Fallback на простое заполнение
|
224 |
+
return self.simple_inpaint(image, mask)
|
225 |
+
|
226 |
+
try:
|
227 |
+
# Используем inpainting pipeline с оптимизированными параметрами
|
228 |
+
result = self.pipe(
|
229 |
+
prompt="empty room interior, clean wall, seamless texture",
|
230 |
+
negative_prompt="furniture, objects, people, clutter",
|
231 |
+
image=image,
|
232 |
+
mask_image=mask,
|
233 |
+
strength=0.95, # Немного меньше для скорости
|
234 |
+
num_inference_steps=25, # Оптимизировано!
|
235 |
+
guidance_scale=5.0 # Меньше для скорости
|
236 |
+
).images[0]
|
237 |
+
|
238 |
+
return result
|
239 |
+
except Exception as e:
|
240 |
+
print(f"Inpainting failed: {e}, using OpenCV fallback")
|
241 |
+
return self.simple_inpaint(image, mask)
|
242 |
+
|
243 |
+
def simple_inpaint(self, image, mask):
|
244 |
+
"""Простое заполнение через OpenCV - очень быстро"""
|
245 |
+
img_array = np.array(image)
|
246 |
+
mask_array = np.array(mask.convert('L'))
|
247 |
+
|
248 |
+
# Инпейнтинг через OpenCV
|
249 |
+
result = cv2.inpaint(img_array, mask_array, 3, cv2.INPAINT_TELEA)
|
250 |
+
|
251 |
+
return Image.fromarray(result)
|
252 |
+
|
253 |
+
def generate_mask_from_text(self, image, text_description, precision=0.3):
|
254 |
+
"""Генерация маски на основе текстового описания"""
|
255 |
+
# Простая маска в центре (заглушка)
|
256 |
+
# В реальности тут должен быть CLIP или SAM
|
257 |
+
width, height = image.size
|
258 |
+
mask = Image.new('L', (width, height), 0)
|
259 |
+
|
260 |
+
# Создаем маску в центре
|
261 |
+
center_x, center_y = width // 2, height // 2
|
262 |
+
radius = int(min(width, height) * precision)
|
263 |
+
|
264 |
+
# Рисуем круг
|
265 |
+
from PIL import ImageDraw
|
266 |
+
draw = ImageDraw.Draw(mask)
|
267 |
+
draw.ellipse([center_x - radius, center_y - radius,
|
268 |
+
center_x + radius, center_y + radius], fill=255)
|
269 |
+
|
270 |
+
return mask
|