StableDesign2 / prod_model.py
anbucur
Refactor generate_design method in ProductionDesignModel for improved image handling and variation generation
5d8e518
from model import DesignModel
from PIL import Image
import numpy as np
from typing import List
import random
import time
import torch
from diffusers.pipelines.controlnet import StableDiffusionControlNetInpaintPipeline
from diffusers import ControlNetModel, UniPCMultistepScheduler, AutoPipelineForText2Image
from transformers import AutoImageProcessor, UperNetForSemanticSegmentation, AutoModelForDepthEstimation
import logging
import os
from datetime import datetime
import gc
# Set up logging
log_dir = "logs"
os.makedirs(log_dir, exist_ok=True)
log_file = os.path.join(log_dir, f"prod_model_{datetime.now().strftime('%Y%m%d')}.log")
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
handlers=[
logging.FileHandler(log_file),
logging.StreamHandler()
]
)
class ProductionDesignModel(DesignModel):
def __init__(self):
"""Initialize the production model with advanced architecture"""
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.dtype = torch.float16 if self.device == "cuda" else torch.float32
# Setup logging
logging.basicConfig(filename=f'logs/prod_model_{time.strftime("%Y%m%d")}.log',
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s')
self.seed = 323*111
self.neg_prompt = "window, door, low resolution, banner, logo, watermark, text, deformed, blurry, out of focus, surreal, ugly, beginner"
self.control_items = ["windowpane;window", "door;double;door"]
self.additional_quality_suffix = "interior design, 4K, high resolution, photorealistic"
try:
logging.info(f"Initializing models on {self.device} with {self.dtype}")
self._initialize_models()
logging.info("Models initialized successfully")
except Exception as e:
logging.error(f"Error initializing models: {e}")
raise
def _initialize_models(self):
"""Initialize all required models and pipelines"""
# Initialize ControlNet models
self.controlnet_depth = ControlNetModel.from_pretrained(
"controlnet_depth", torch_dtype=self.dtype, use_safetensors=True
)
self.controlnet_seg = ControlNetModel.from_pretrained(
"own_controlnet", torch_dtype=self.dtype, use_safetensors=True
)
# Initialize main pipeline
self.pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
"SG161222/Realistic_Vision_V5.1_noVAE",
controlnet=[self.controlnet_depth, self.controlnet_seg],
safety_checker=None,
torch_dtype=self.dtype
)
# Setup IP-Adapter
self.pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models",
weight_name="ip-adapter_sd15.bin")
self.pipe.set_ip_adapter_scale(0.4)
self.pipe.scheduler = UniPCMultistepScheduler.from_config(self.pipe.scheduler.config)
self.pipe = self.pipe.to(self.device)
# Initialize guide pipeline
self.guide_pipe = AutoPipelineForText2Image.from_pretrained(
"segmind/SSD-1B",
torch_dtype=self.dtype,
use_safetensors=True,
variant="fp16"
).to(self.device)
# Initialize segmentation and depth models
self.seg_processor, self.seg_model = self._init_segmentation()
self.depth_processor, self.depth_model = self._init_depth()
self.depth_model = self.depth_model.to(self.device)
def _init_segmentation(self):
"""Initialize segmentation models"""
processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-small")
model = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-convnext-small")
return processor, model
def _init_depth(self):
"""Initialize depth estimation models"""
processor = AutoImageProcessor.from_pretrained(
"LiheYoung/depth-anything-large-hf",
torch_dtype=self.dtype
)
model = AutoModelForDepthEstimation.from_pretrained(
"LiheYoung/depth-anything-large-hf",
torch_dtype=self.dtype
)
return processor, model
def _get_depth_map(self, image: Image.Image) -> Image.Image:
"""Generate depth map for input image"""
image_to_depth = self.depth_processor(images=image, return_tensors="pt").to(self.device)
with torch.inference_mode():
depth_map = self.depth_model(**image_to_depth).predicted_depth
width, height = image.size
depth_map = torch.nn.functional.interpolate(
depth_map.unsqueeze(1).float(),
size=(height, width),
mode="bicubic",
align_corners=False,
)
depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
depth_map = (depth_map - depth_min) / (depth_max - depth_min)
image = torch.cat([depth_map] * 3, dim=1)
image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
return Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
def _segment_image(self, image: Image.Image) -> Image.Image:
"""Generate segmentation map for input image"""
pixel_values = self.seg_processor(image, return_tensors="pt").pixel_values
with torch.inference_mode():
outputs = self.seg_model(pixel_values)
seg = self.seg_processor.post_process_semantic_segmentation(
outputs, target_sizes=[image.size[::-1]])[0]
color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8)
# You'll need to implement the palette mapping here
# This is a placeholder - you should implement proper color mapping
for label in range(seg.max() + 1):
color_seg[seg == label, :] = [label * 30 % 255] * 3
return Image.fromarray(color_seg).convert('RGB')
def _resize_image(self, image: Image.Image, target_size: int) -> Image.Image:
"""Resize image while maintaining aspect ratio"""
width, height = image.size
if width > height:
new_width = target_size
new_height = int(height * (target_size / width))
else:
new_height = target_size
new_width = int(width * (target_size / height))
return image.resize((new_width, new_height), Image.LANCZOS)
def _flush(self):
"""Clear CUDA cache"""
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
def generate_design(self, image, num_variations=1, **kwargs):
"""Generate design variations using the model.
Args:
image: Input image (PIL Image, numpy array, or torch tensor)
num_variations: Number of variations to generate
**kwargs: Additional parameters like prompt, num_steps, guidance_scale, strength
Returns:
List of generated images
"""
try:
# Convert image to PIL Image if needed
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
elif isinstance(image, torch.Tensor):
# Convert tensor to numpy then PIL
image = Image.fromarray((image.cpu().numpy() * 255).astype(np.uint8))
if not isinstance(image, Image.Image):
raise ValueError(f"Unsupported image type: {type(image)}")
# Ensure image is RGB
if image.mode != "RGB":
image = image.convert("RGB")
# Get parameters
prompt = kwargs.get('prompt', '')
num_steps = int(kwargs.get('num_steps', 50))
guidance_scale = float(kwargs.get('guidance_scale', 10.0))
strength = float(kwargs.get('strength', 0.9))
seed_param = kwargs.get('seed')
# Handle seed
base_seed = int(time.time()) if seed_param is None else int(seed_param)
logging.info(f"Using base seed: {base_seed}")
variations = []
for i in range(num_variations):
try:
# Generate distinct seed for each variation
seed = base_seed + i
generator = torch.Generator(device=self.device).manual_seed(seed)
# Generate variation
output = self.pipe(
prompt=prompt,
image=image,
num_inference_steps=num_steps,
guidance_scale=guidance_scale,
strength=strength,
generator=generator,
negative_prompt=self.neg_prompt
).images[0]
variations.append(output)
logging.info(f"Successfully generated variation {i} with seed {seed}")
except Exception as e:
logging.error(f"Error generating variation {i}: {str(e)}")
continue
finally:
# Clear CUDA cache after each variation
if torch.cuda.is_available():
torch.cuda.empty_cache()
if not variations:
logging.warning("No variations were generated successfully")
return [image] # Return original image if no variations generated
return variations
except Exception as e:
logging.error(f"Error in generate_design: {str(e)}")
return [image] # Return original image on error
def __del__(self):
"""Cleanup when the model is deleted"""
self._flush()