feat: add Real-ESRGAN AI upscaling support
- New module: modules/ai_upscaler.py - Real-ESRGAN integration for AI-powered upscaling - Designed specifically for low-res game textures/icons - 4x upscale factor with neural network enhancement - Falls back gracefully if model not available - Updated TGA Converter UI: - New "AI Enhanced (Real-ESRGAN)" option in upscale dropdown - Only shown if Real-ESRGAN is installed - AI upscaling happens before canvas placement - Shows progress during AI model loading Real-ESRGAN is much better than basic methods for rendered (non-pixel) icons! To use AI upscaling: 1. pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu 2. pip install realesrgan 3. Download RealESRGAN_x4plus.pth model 4. Select "AI Enhanced" in the UI
This commit is contained in:
parent
0034cc9453
commit
5374eba08f
|
|
@ -0,0 +1,220 @@
|
|||
"""
|
||||
Lemontropia Suite - AI Image Upscaling with Real-ESRGAN
|
||||
Optional AI-powered upscaling for game icons and textures.
|
||||
|
||||
Real-ESRGAN is specifically designed for low-resolution game graphics
|
||||
and produces excellent results for rendered icons (not pixel art).
|
||||
"""
|
||||
|
||||
import logging
|
||||
from pathlib import Path
|
||||
from typing import Optional, Tuple
|
||||
import numpy as np
|
||||
|
||||
try:
|
||||
from PIL import Image
|
||||
PIL_AVAILABLE = True
|
||||
except ImportError:
|
||||
PIL_AVAILABLE = False
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Try to import Real-ESRGAN
|
||||
try:
|
||||
import torch
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet
|
||||
from realesrgan import RealESRGANer
|
||||
REALESRGAN_AVAILABLE = True
|
||||
except ImportError:
|
||||
REALESRGAN_AVAILABLE = False
|
||||
logger.info("Real-ESRGAN not available. Install with: pip install realesrgan")
|
||||
|
||||
|
||||
class AIIconUpscaler:
|
||||
"""
|
||||
AI-powered upscaler for game icons using Real-ESRGAN.
|
||||
|
||||
Real-ESRGAN is trained specifically on game textures and produces
|
||||
excellent results for low-resolution rendered graphics (not pixel art).
|
||||
|
||||
Usage:
|
||||
upscaler = AIIconUpscaler()
|
||||
if upscaler.is_available():
|
||||
result = upscaler.upscale(image, scale=4)
|
||||
"""
|
||||
|
||||
# Model download URLs
|
||||
MODEL_URLS = {
|
||||
'RealESRGAN_x4plus': 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth',
|
||||
'RealESRGAN_x4plus_anime_6B': 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus_anime_6B.pth',
|
||||
'RealESRGAN_x2plus': 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x2plus.pth',
|
||||
}
|
||||
|
||||
def __init__(self, model_name: str = 'RealESRGAN_x4plus', device: str = 'cpu'):
|
||||
"""
|
||||
Initialize AI upscaler.
|
||||
|
||||
Args:
|
||||
model_name: Which model to use
|
||||
device: 'cpu' or 'cuda' (GPU)
|
||||
"""
|
||||
self.model_name = model_name
|
||||
self.device = device
|
||||
self.upsampler = None
|
||||
|
||||
if REALESRGAN_AVAILABLE:
|
||||
self._init_model()
|
||||
|
||||
def _init_model(self):
|
||||
"""Initialize the Real-ESRGAN model."""
|
||||
try:
|
||||
# Model parameters
|
||||
if 'anime' in self.model_name:
|
||||
# Anime model (6B parameters)
|
||||
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64,
|
||||
num_block=6, num_grow_ch=32, scale=4)
|
||||
elif 'x2' in self.model_name:
|
||||
# 2x upscale
|
||||
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64,
|
||||
num_block=23, num_grow_ch=32, scale=2)
|
||||
else:
|
||||
# 4x upscale (default)
|
||||
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64,
|
||||
num_block=23, num_grow_ch=32, scale=4)
|
||||
|
||||
# Get model path
|
||||
model_path = self._get_model_path()
|
||||
|
||||
if not model_path.exists():
|
||||
logger.warning(f"Model not found: {model_path}")
|
||||
logger.info(f"Download from: {self.MODEL_URLS.get(self.model_name)}")
|
||||
return
|
||||
|
||||
# Initialize upsampler
|
||||
self.upsampler = RealESRGANer(
|
||||
scale=4 if 'x4' in self.model_name else 2,
|
||||
model_path=str(model_path),
|
||||
model=model,
|
||||
tile=0, # No tiling (process whole image)
|
||||
tile_pad=10,
|
||||
pre_pad=0,
|
||||
half=False if self.device == 'cpu' else True,
|
||||
device=torch.device(self.device)
|
||||
)
|
||||
|
||||
logger.info(f"Real-ESRGAN initialized: {self.model_name} on {self.device}")
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Failed to initialize Real-ESRGAN: {e}")
|
||||
self.upsampler = None
|
||||
|
||||
def _get_model_path(self) -> Path:
|
||||
"""Get path to model file."""
|
||||
# Store models in user's home directory
|
||||
model_dir = Path.home() / ".lemontropia" / "models"
|
||||
model_dir.mkdir(parents=True, exist_ok=True)
|
||||
return model_dir / f"{self.model_name}.pth"
|
||||
|
||||
def is_available(self) -> bool:
|
||||
"""Check if AI upscaler is available and ready."""
|
||||
return REALESRGAN_AVAILABLE and self.upsampler is not None
|
||||
|
||||
def upscale(self, image: Image.Image, scale: int = 4) -> Optional[Image.Image]:
|
||||
"""
|
||||
Upscale an image using Real-ESRGAN.
|
||||
|
||||
Args:
|
||||
image: PIL Image to upscale
|
||||
scale: Upscale factor (2 or 4)
|
||||
|
||||
Returns:
|
||||
Upscaled PIL Image or None if failed
|
||||
"""
|
||||
if not self.is_available():
|
||||
logger.error("AI upscaler not available")
|
||||
return None
|
||||
|
||||
try:
|
||||
# Convert PIL to numpy
|
||||
img_np = np.array(image)
|
||||
|
||||
# Remove alpha channel if present (Real-ESRGAN expects RGB)
|
||||
has_alpha = img_np.shape[-1] == 4
|
||||
if has_alpha:
|
||||
alpha = img_np[:, :, 3]
|
||||
img_rgb = img_np[:, :, :3]
|
||||
else:
|
||||
img_rgb = img_np
|
||||
|
||||
# Upscale
|
||||
output, _ = self.upsampler.enhance(img_rgb, outscale=scale)
|
||||
|
||||
# Restore alpha channel if needed
|
||||
if has_alpha:
|
||||
# Upscale alpha with simple resize
|
||||
from PIL import Image as PILImage
|
||||
alpha_pil = PILImage.fromarray(alpha)
|
||||
alpha_upscaled = alpha_pil.resize(
|
||||
(output.shape[1], output.shape[0]),
|
||||
PILImage.Resampling.LANCZOS
|
||||
)
|
||||
alpha_np = np.array(alpha_upscaled)
|
||||
output = np.dstack([output, alpha_np])
|
||||
|
||||
# Convert back to PIL
|
||||
result = Image.fromarray(output)
|
||||
return result
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Upscaling failed: {e}")
|
||||
return None
|
||||
|
||||
def get_info(self) -> dict:
|
||||
"""Get information about the upscaler status."""
|
||||
return {
|
||||
'available': self.is_available(),
|
||||
'model': self.model_name,
|
||||
'device': self.device,
|
||||
'model_path': str(self._get_model_path()),
|
||||
'dependencies': {
|
||||
'torch': torch.__version__ if REALESRGAN_AVAILABLE else 'not installed',
|
||||
'realesrgan': REALESRGAN_AVAILABLE
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
# Convenience function
|
||||
def upscale_with_ai(image: Image.Image, scale: int = 4) -> Optional[Image.Image]:
|
||||
"""Quick AI upscale using Real-ESRGAN."""
|
||||
upscaler = AIIconUpscaler()
|
||||
return upscaler.upscale(image, scale)
|
||||
|
||||
|
||||
def check_ai_upscaler():
|
||||
"""Check if AI upscaler is available and print status."""
|
||||
upscaler = AIIconUpscaler()
|
||||
info = upscaler.get_info()
|
||||
|
||||
print("=" * 50)
|
||||
print("AI Upscaler Status (Real-ESRGAN)")
|
||||
print("=" * 50)
|
||||
print(f"Available: {info['available']}")
|
||||
print(f"Model: {info['model']}")
|
||||
print(f"Device: {info['device']}")
|
||||
print(f"Model Path: {info['model_path']}")
|
||||
print(f"PyTorch: {info['dependencies']['torch']}")
|
||||
print("=" * 50)
|
||||
|
||||
if not info['available']:
|
||||
print("\nTo install AI upscaler:")
|
||||
print("1. pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu")
|
||||
print("2. pip install realesrgan")
|
||||
print("3. Download model from:")
|
||||
for name, url in AIIconUpscaler.MODEL_URLS.items():
|
||||
print(f" {name}: {url}")
|
||||
|
||||
return info['available']
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
check_ai_upscaler()
|
||||
|
|
@ -17,6 +17,7 @@ from PyQt6.QtCore import Qt, QThread, pyqtSignal
|
|||
from PyQt6.QtGui import QPixmap
|
||||
|
||||
from modules.tga_converter import TGAConverter
|
||||
from modules.ai_upscaler import AIIconUpscaler, REALESRGAN_AVAILABLE
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
|
@ -59,6 +60,15 @@ class TGAConvertWorker(QThread):
|
|||
total = len(tga_files)
|
||||
success = 0
|
||||
|
||||
# Initialize AI upscaler if needed
|
||||
ai_upscaler = None
|
||||
if self.upscale_method == 'ai' and REALESRGAN_AVAILABLE:
|
||||
self.progress_update.emit("Loading AI upscaler (Real-ESRGAN)...")
|
||||
ai_upscaler = AIIconUpscaler(device='cpu')
|
||||
if not ai_upscaler.is_available():
|
||||
self.progress_update.emit("AI model not found, falling back to HQ4x")
|
||||
self.upscale_method = 'hq4x'
|
||||
|
||||
canvas_info = f" ({self.canvas_size[0]}x{self.canvas_size[1]} canvas)" if self.canvas_size else ""
|
||||
self.progress_update.emit(f"Found {total} TGA files to convert{canvas_info}")
|
||||
|
||||
|
|
@ -68,12 +78,17 @@ class TGAConvertWorker(QThread):
|
|||
|
||||
self.progress_update.emit(f"[{i+1}/{total}] Converting: {tga_path.name}")
|
||||
|
||||
output_path = self.converter.convert_tga_to_png(
|
||||
tga_path,
|
||||
canvas_size=self.canvas_size,
|
||||
upscale=self.upscale,
|
||||
upscale_method=self.upscale_method
|
||||
)
|
||||
# Handle AI upscaling separately
|
||||
if self.upscale_method == 'ai' and ai_upscaler and ai_upscaler.is_available():
|
||||
output_path = self._convert_with_ai(tga_path, ai_upscaler)
|
||||
else:
|
||||
output_path = self.converter.convert_tga_to_png(
|
||||
tga_path,
|
||||
canvas_size=self.canvas_size,
|
||||
upscale=True,
|
||||
upscale_method=self.upscale_method
|
||||
)
|
||||
|
||||
if output_path:
|
||||
success += 1
|
||||
self.file_converted.emit(tga_path.name, str(output_path))
|
||||
|
|
@ -83,6 +98,44 @@ class TGAConvertWorker(QThread):
|
|||
except Exception as e:
|
||||
self.conversion_error.emit(str(e))
|
||||
|
||||
def _convert_with_ai(self, tga_path: Path, ai_upscaler: AIIconUpscaler) -> Optional[Path]:
|
||||
"""Convert a single TGA file using AI upscaling."""
|
||||
try:
|
||||
from PIL import Image
|
||||
|
||||
# Load TGA
|
||||
image = Image.open(tga_path)
|
||||
if image.mode != 'RGBA':
|
||||
image = image.convert('RGBA')
|
||||
|
||||
# AI upscale (4x)
|
||||
self.progress_update.emit(f" AI upscaling {tga_path.name}...")
|
||||
upscaled = ai_upscaler.upscale(image, scale=4)
|
||||
|
||||
if upscaled is None:
|
||||
return None
|
||||
|
||||
# Apply canvas if requested
|
||||
if self.canvas_size:
|
||||
upscaled = self.converter._apply_canvas(
|
||||
upscaled, self.canvas_size, upscale=False
|
||||
)
|
||||
|
||||
# Save
|
||||
output_path = self.converter.output_dir / f"{tga_path.stem}.png"
|
||||
upscaled.save(output_path, 'PNG')
|
||||
|
||||
return output_path
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"AI conversion failed: {e}")
|
||||
return None
|
||||
|
||||
self.conversion_complete.emit(success, total)
|
||||
|
||||
except Exception as e:
|
||||
self.conversion_error.emit(str(e))
|
||||
|
||||
def stop(self):
|
||||
"""Stop the conversion."""
|
||||
self._is_running = False
|
||||
|
|
@ -204,10 +257,16 @@ class TGAConverterDialog(QDialog):
|
|||
self.upscale_method_combo.addItem("Sharp Pixels (NEAREST)", "nearest")
|
||||
self.upscale_method_combo.addItem("Smooth (HQ4x-style)", "hq4x")
|
||||
self.upscale_method_combo.addItem("Photorealistic (LANCZOS)", "lanczos")
|
||||
|
||||
# Add AI option if Real-ESRGAN is available
|
||||
if REALESRGAN_AVAILABLE:
|
||||
self.upscale_method_combo.addItem("🤖 AI Enhanced (Real-ESRGAN)", "ai")
|
||||
|
||||
self.upscale_method_combo.setToolTip(
|
||||
"NEAREST: Sharp pixel edges (best for pixel art)\n"
|
||||
"HQ4x: Smooth but keeps details (best for game icons)\n"
|
||||
"LANCZOS: Very smooth (best for photos)"
|
||||
"LANCZOS: Very smooth (best for photos)\n"
|
||||
"AI: Neural network upscaling (best quality, requires model)"
|
||||
)
|
||||
canvas_layout.addWidget(self.upscale_method_combo)
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue