image-evaluation-tool / models /aesthetics_evaluator.py
VOIDER's picture
Upload 14 files
83b7522 verified
raw
history blame
12.6 kB
import torch
import torch.nn as nn
import numpy as np
from PIL import Image
import torchvision.transforms as transforms
from transformers import AutoModel, AutoProcessor
import logging
logger = logging.getLogger(__name__)
class AestheticsEvaluator:
"""Image aesthetics assessment using multiple SOTA models"""
def __init__(self):
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.models = {}
self.processors = {}
self.load_models()
def load_models(self):
"""Load aesthetics assessment models"""
try:
# Load UNIAA model (primary)
logger.info("Loading UNIAA model...")
self.load_uniaa()
# Load MUSIQ model (secondary)
logger.info("Loading MUSIQ model...")
self.load_musiq()
# Load anime-specific aesthetic model
logger.info("Loading anime aesthetic model...")
self.load_anime_aesthetic_model()
except Exception as e:
logger.error(f"Error loading aesthetic models: {str(e)}")
self.use_fallback_implementation()
def load_uniaa(self):
"""Load UNIAA model"""
try:
# Placeholder implementation for UNIAA
self.models['uniaa'] = self.create_mock_aesthetic_model()
self.processors['uniaa'] = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
except Exception as e:
logger.warning(f"Could not load UNIAA: {str(e)}")
def load_musiq(self):
"""Load MUSIQ model"""
try:
# Placeholder implementation for MUSIQ
self.models['musiq'] = self.create_mock_aesthetic_model()
self.processors['musiq'] = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
except Exception as e:
logger.warning(f"Could not load MUSIQ: {str(e)}")
def load_anime_aesthetic_model(self):
"""Load anime-specific aesthetic model"""
try:
# Placeholder for anime-specific model
self.models['anime_aesthetic'] = self.create_mock_aesthetic_model()
self.processors['anime_aesthetic'] = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
except Exception as e:
logger.warning(f"Could not load anime aesthetic model: {str(e)}")
def create_mock_aesthetic_model(self):
"""Create a mock aesthetic model for demonstration"""
class MockAestheticModel(nn.Module):
def __init__(self):
super().__init__()
self.backbone = torch.nn.Sequential(
torch.nn.Conv2d(3, 64, 3, padding=1),
torch.nn.ReLU(),
torch.nn.Conv2d(64, 128, 3, padding=1),
torch.nn.ReLU(),
torch.nn.AdaptiveAvgPool2d((1, 1)),
torch.nn.Flatten(),
torch.nn.Linear(128, 64),
torch.nn.ReLU(),
torch.nn.Linear(64, 1),
torch.nn.Sigmoid()
)
def forward(self, x):
return self.backbone(x) * 10 # Scale to 0-10
model = MockAestheticModel().to(self.device)
model.eval()
return model
def use_fallback_implementation(self):
"""Use simple fallback aesthetic assessment"""
logger.info("Using fallback aesthetic assessment implementation")
self.fallback_mode = True
def evaluate_with_uniaa(self, image: Image.Image) -> float:
"""Evaluate aesthetics using UNIAA"""
try:
if 'uniaa' not in self.models:
return self.fallback_aesthetic_score(image)
# Preprocess image
tensor = self.processors['uniaa'](image).unsqueeze(0).to(self.device)
# Get prediction
with torch.no_grad():
score = self.models['uniaa'](tensor).item()
return max(0.0, min(10.0, score))
except Exception as e:
logger.error(f"Error in UNIAA evaluation: {str(e)}")
return self.fallback_aesthetic_score(image)
def evaluate_with_musiq(self, image: Image.Image) -> float:
"""Evaluate aesthetics using MUSIQ"""
try:
if 'musiq' not in self.models:
return self.fallback_aesthetic_score(image)
# Preprocess image
tensor = self.processors['musiq'](image).unsqueeze(0).to(self.device)
# Get prediction
with torch.no_grad():
score = self.models['musiq'](tensor).item()
return max(0.0, min(10.0, score))
except Exception as e:
logger.error(f"Error in MUSIQ evaluation: {str(e)}")
return self.fallback_aesthetic_score(image)
def evaluate_with_anime_model(self, image: Image.Image) -> float:
"""Evaluate aesthetics using anime-specific model"""
try:
if 'anime_aesthetic' not in self.models:
return self.fallback_aesthetic_score(image)
# Preprocess image
tensor = self.processors['anime_aesthetic'](image).unsqueeze(0).to(self.device)
# Get prediction
with torch.no_grad():
score = self.models['anime_aesthetic'](tensor).item()
return max(0.0, min(10.0, score))
except Exception as e:
logger.error(f"Error in anime aesthetic evaluation: {str(e)}")
return self.fallback_aesthetic_score(image)
def evaluate_composition_rules(self, image: Image.Image) -> float:
"""Evaluate based on composition rules (rule of thirds, etc.)"""
try:
# Convert to numpy array
img_array = np.array(image)
height, width = img_array.shape[:2]
# Convert to grayscale for analysis
if len(img_array.shape) == 3:
gray = np.dot(img_array[...,:3], [0.2989, 0.5870, 0.1140])
else:
gray = img_array
# Rule of thirds analysis
third_h, third_w = height // 3, width // 3
# Check for interesting content at rule of thirds intersections
intersections = [
(third_h, third_w), (third_h, 2*third_w),
(2*third_h, third_w), (2*third_h, 2*third_w)
]
composition_score = 0.0
for y, x in intersections:
# Check local variance around intersection points
region = gray[max(0, y-10):min(height, y+10),
max(0, x-10):min(width, x+10)]
if region.size > 0:
composition_score += region.var()
# Normalize composition score
composition_score = min(10.0, composition_score / 1000.0)
# Color harmony analysis
if len(img_array.shape) == 3:
# Calculate color distribution
colors = img_array.reshape(-1, 3)
color_std = np.std(colors, axis=0).mean()
color_harmony_score = min(10.0, color_std / 25.0)
else:
color_harmony_score = 5.0
# Combine scores
final_score = (composition_score * 0.6 + color_harmony_score * 0.4)
return max(0.0, min(10.0, final_score))
except Exception as e:
logger.error(f"Error in composition analysis: {str(e)}")
return 5.0
def fallback_aesthetic_score(self, image: Image.Image) -> float:
"""Simple fallback aesthetic assessment"""
try:
# Basic aesthetic assessment based on image properties
width, height = image.size
# Aspect ratio score (prefer aesthetically pleasing ratios)
aspect_ratio = width / height
golden_ratio = 1.618
if abs(aspect_ratio - golden_ratio) < 0.1 or abs(aspect_ratio - 1/golden_ratio) < 0.1:
aspect_score = 9.0
elif 0.7 <= aspect_ratio <= 1.4: # Square-ish
aspect_score = 7.0
elif 1.4 <= aspect_ratio <= 2.0: # Landscape
aspect_score = 8.0
else:
aspect_score = 5.0
# Resolution score (higher resolution often looks better)
total_pixels = width * height
resolution_score = min(10.0, total_pixels / 200000.0) # Normalize by 2MP
# Color analysis
img_array = np.array(image)
if len(img_array.shape) == 3:
# Color variety score
unique_colors = len(np.unique(img_array.reshape(-1, 3), axis=0))
color_variety_score = min(10.0, unique_colors / 1000.0)
# Brightness distribution
brightness = np.mean(img_array, axis=2)
brightness_score = 10.0 - abs(brightness.mean() - 127.5) / 12.75
else:
color_variety_score = 5.0
brightness_score = 5.0
# Combine scores
aesthetic_score = (aspect_score * 0.3 +
resolution_score * 0.2 +
color_variety_score * 0.3 +
brightness_score * 0.2)
return max(0.0, min(10.0, aesthetic_score))
except Exception:
return 5.0 # Default neutral score
def evaluate(self, image: Image.Image, anime_mode: bool = False) -> float:
"""
Evaluate image aesthetics using ensemble of models
Args:
image: PIL Image to evaluate
anime_mode: Whether to use anime-specific evaluation
Returns:
Aesthetic score from 0-10
"""
try:
scores = []
if anime_mode:
# For anime images, prioritize anime-specific model
anime_score = self.evaluate_with_anime_model(image)
scores.append(anime_score)
# Also use general models but with lower weight
uniaa_score = self.evaluate_with_uniaa(image)
scores.append(uniaa_score)
# Composition rules
composition_score = self.evaluate_composition_rules(image)
scores.append(composition_score)
# Weights for anime mode
weights = [0.5, 0.3, 0.2]
else:
# For realistic images, use general aesthetic models
uniaa_score = self.evaluate_with_uniaa(image)
scores.append(uniaa_score)
musiq_score = self.evaluate_with_musiq(image)
scores.append(musiq_score)
# Composition rules
composition_score = self.evaluate_composition_rules(image)
scores.append(composition_score)
# Weights for realistic mode
weights = [0.4, 0.4, 0.2]
# Ensemble scoring
final_score = sum(score * weight for score, weight in zip(scores, weights))
logger.info(f"Aesthetic scores - Scores: {scores}, Final: {final_score:.2f}")
return max(0.0, min(10.0, final_score))
except Exception as e:
logger.error(f"Error in aesthetic evaluation: {str(e)}")
return self.fallback_aesthetic_score(image)