def upscale_image(cfg, src, scale_model, alpha_model=None):
dst, alpha = split_alpha(src, scale_model)
for i in range(int(np.ceil(np.log2(cfg.scale_ratio)))):
six.print_('2.0x upscaling...', end=' ', flush=True)
model = scale_model if i == 0 or alpha_model is None else alpha_model
if model.inner_scale == 1:
dst = iproc.nn_scaling(dst, 2) # Nearest neighbor 2x scaling
alpha = iproc.nn_scaling(alpha, 2) # Nearest neighbor 2x scaling
if cfg.tta:
dst = reconstruct.image_tta(dst, model, cfg.tta_level,
cfg.block_size, cfg.batch_size)
else:
dst = reconstruct.image(dst, model, cfg.block_size, cfg.batch_size)
if alpha_model is None:
alpha = reconstruct.image(alpha, scale_model, cfg.block_size,
cfg.batch_size)
else:
alpha = reconstruct.image(alpha, alpha_model, cfg.block_size,
cfg.batch_size)
six.print_('OK')
dst_w = int(np.round(src.size[0] * cfg.scale_ratio))
dst_h = int(np.round(src.size[1] * cfg.scale_ratio))
if dst_w != dst.size[0] or dst_h != dst.size[1]:
six.print_('Resizing...', end=' ', flush=True)
dst = dst.resize((dst_w, dst_h), Image.LANCZOS)
six.print_('OK')
if alpha is not None:
if alpha.size[0] != dst_w or alpha.size[1] != dst_h:
alpha = alpha.resize((dst_w, dst_h), Image.LANCZOS)
dst.putalpha(alpha)
return dst
def upscale_image(cfg, src, model):
dst = src.copy()
six.print_('2.0x scaling...', end=' ', flush=True)
if model.inner_scale == 1:
dst = iproc.nn_scaling(dst, 2) # Nearest neighbor 2x scaling
if cfg.tta:
dst = reconstruct.image_tta(dst, model, cfg.tta_level, cfg.block_size,
cfg.batch_size)
else:
dst = reconstruct.image(dst, model, cfg.block_size, cfg.batch_size)
six.print_('OK')
return dst
def pairwise_transform(src, cfg):
unstable_region_offset_x = 8
unstable_region_offset_y = unstable_region_offset_x * cfg.inner_scale
top = cfg.offset
bottom = cfg.crop_size - top
y = preprocess(src, cfg)
if cfg.method == 'scale':
x = scale(y, cfg.downsampling_filters, cfg.resize_blur_min,
cfg.resize_blur_max, cfg.inner_scale == 1)
elif cfg.method == 'noise':
if cfg.inner_scale != 1:
raise ValueError('inner_scale must be 1')
x = noise(y, cfg.nr_rate, cfg.chroma_subsampling_rate, cfg.noise_level)
elif cfg.method == 'noise_scale':
if cfg.inner_scale == 1:
raise ValueError('inner_scale must be > 1')
x = noise_scale(y, cfg.downsampling_filters, cfg.resize_blur_min,
cfg.resize_blur_max, False, cfg.nr_rate,
cfg.chroma_subsampling_rate, cfg.noise_level)
y = y[unstable_region_offset_y:y.shape[0] - unstable_region_offset_y,
unstable_region_offset_y:y.shape[1] - unstable_region_offset_y]
x = x[unstable_region_offset_x:x.shape[0] - unstable_region_offset_x,
unstable_region_offset_x:x.shape[1] - unstable_region_offset_x]
lowres_y = y.copy()
if cfg.crop_size != cfg.in_size:
lowres_y = iproc.nn_scaling(y, 1 / cfg.inner_scale)
patch_x = np.zeros((cfg.patches, cfg.ch, cfg.in_size, cfg.in_size),
dtype=np.uint8)
patch_y = np.zeros((cfg.patches, cfg.ch, cfg.out_size, cfg.out_size),
dtype=np.uint8)
for i in range(cfg.patches):
crop_x, crop_y = active_cropping(x, y, lowres_y, cfg.crop_size,
cfg.inner_scale,
cfg.active_cropping_rate,
cfg.active_cropping_tries)
if cfg.ch == 1:
ycbcr_x = Image.fromarray(crop_x).convert('YCbCr')
ycbcr_y = Image.fromarray(crop_y).convert('YCbCr')
crop_x = np.array(ycbcr_x)[:, :, 0]
crop_y = np.array(ycbcr_y)[top:bottom, top:bottom, 0]
patch_x[i] = crop_x[np.newaxis, :, :]
patch_y[i] = crop_y[np.newaxis, :, :]
elif cfg.ch == 3:
crop_y = crop_y[top:bottom, top:bottom, :]
patch_x[i] = crop_x.transpose(2, 0, 1)
patch_y[i] = crop_y.transpose(2, 0, 1)
return patch_x, patch_y