47 lines
1.3 KiB
Python
47 lines
1.3 KiB
Python
import cv2
|
|
|
|
from lama_cleaner.helper import download_model
|
|
|
|
|
|
class RealESRGANUpscaler:
|
|
name = "RealESRGAN"
|
|
|
|
def __init__(self, device):
|
|
super().__init__()
|
|
from basicsr.archs.rrdbnet_arch import RRDBNet
|
|
from realesrgan import RealESRGANer
|
|
|
|
scale = 4
|
|
model = RRDBNet(
|
|
num_in_ch=3,
|
|
num_out_ch=3,
|
|
num_feat=64,
|
|
num_block=23,
|
|
num_grow_ch=32,
|
|
scale=4,
|
|
)
|
|
url = "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth"
|
|
model_md5 = "99ec365d4afad750833258a1a24f44ca"
|
|
model_path = download_model(url, model_md5)
|
|
|
|
self.model = RealESRGANer(
|
|
scale=scale,
|
|
model_path=model_path,
|
|
model=model,
|
|
half=True if "cuda" in str(device) else False,
|
|
tile=640,
|
|
tile_pad=10,
|
|
pre_pad=10,
|
|
device=device,
|
|
)
|
|
|
|
def __call__(self, rgb_np_img, files, form):
|
|
bgr_np_img = cv2.cvtColor(rgb_np_img, cv2.COLOR_RGB2BGR)
|
|
scale = float(form["scale"])
|
|
return self.forward(bgr_np_img, scale)
|
|
|
|
def forward(self, bgr_np_img, scale: float):
|
|
# 输出是 BGR
|
|
upsampled = self.model.enhance(bgr_np_img, outscale=scale)[0]
|
|
return upsampled
|