|
|
|
|
|
import os
|
|
|
import sys
|
|
|
import argparse
|
|
|
import json
|
|
|
import datetime
|
|
|
import cv2
|
|
|
import torch
|
|
|
import lpips
|
|
|
from torchvision import transforms
|
|
|
import torch.nn.functional as F
|
|
|
from PIL import Image, UnidentifiedImageError
|
|
|
|
|
|
def verify_image(path, exts=('.png','.jpg','.jpeg','.webp')):
|
|
|
if not os.path.isfile(path):
|
|
|
return False, f'File does not exist: {path}'
|
|
|
if os.path.getsize(path) == 0:
|
|
|
return False, f'File is empty: {path}'
|
|
|
if not path.lower().endswith(exts):
|
|
|
return False, f'Unsupported format: {path}'
|
|
|
try:
|
|
|
img = Image.open(path)
|
|
|
img.verify()
|
|
|
except (UnidentifiedImageError, Exception) as e:
|
|
|
return False, f'Failed to read image: {path} ({e})'
|
|
|
return True, ''
|
|
|
|
|
|
def load_tensor(path):
|
|
|
img = cv2.imread(path, cv2.IMREAD_COLOR)
|
|
|
if img is None:
|
|
|
raise RuntimeError(f'cv2 read failed: {path}')
|
|
|
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
|
|
t = transforms.ToTensor()(img) * 2 - 1
|
|
|
return t.unsqueeze(0)
|
|
|
|
|
|
def main():
|
|
|
p = argparse.ArgumentParser(description='Automated anime effect evaluation script')
|
|
|
p.add_argument('--groundtruth', required=True, help='Original image path')
|
|
|
p.add_argument('--output', required=True, help='Anime-styled output image path')
|
|
|
p.add_argument('--lpips-thresh', type=float, default=0.40,
|
|
|
help='LPIPS structural similarity max distance (Pass if <= threshold)')
|
|
|
p.add_argument('--clip-thresh', type=float, default=0.25,
|
|
|
help='CLIP anime style similarity threshold (Pass if > threshold)')
|
|
|
p.add_argument('--result', required=True, help='Result JSONL file path (append mode)')
|
|
|
args = p.parse_args()
|
|
|
|
|
|
process = True
|
|
|
comments = []
|
|
|
|
|
|
|
|
|
for tag, path in [('input', args.groundtruth), ('output', args.output)]:
|
|
|
ok, msg = verify_image(path)
|
|
|
if not ok:
|
|
|
process = False
|
|
|
comments.append(f'[{tag}] {msg}')
|
|
|
|
|
|
lpips_val = None
|
|
|
lpips_pass = True
|
|
|
clip_pass = False
|
|
|
if process:
|
|
|
try:
|
|
|
|
|
|
img0 = load_tensor(args.groundtruth)
|
|
|
img1 = load_tensor(args.output)
|
|
|
_, _, h0, w0 = img0.shape
|
|
|
_, _, h1, w1 = img1.shape
|
|
|
nh, nw = min(h0,h1), min(w0,w1)
|
|
|
img0 = F.interpolate(img0, size=(nh,nw), mode='bilinear', align_corners=False)
|
|
|
img1 = F.interpolate(img1, size=(nh,nw), mode='bilinear', align_corners=False)
|
|
|
|
|
|
loss_fn = lpips.LPIPS(net='vgg').to(torch.device('cpu'))
|
|
|
with torch.no_grad():
|
|
|
lpips_val = float(loss_fn(img0, img1).item())
|
|
|
lpips_pass = lpips_val <= args.lpips_thresh
|
|
|
comments.append(f'LPIPS={lpips_val:.4f} (<= {args.lpips_thresh} → {"OK" if lpips_pass else "FAIL"})')
|
|
|
except Exception as e:
|
|
|
process = False
|
|
|
comments.append(f'Metric calculation error: {e}')
|
|
|
|
|
|
if process:
|
|
|
try:
|
|
|
import clip
|
|
|
import PIL.Image
|
|
|
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
clip_model, clip_preprocess = clip.load("ViT-B/32", device=device)
|
|
|
|
|
|
image = clip_preprocess(PIL.Image.open(args.output)).unsqueeze(0).to(device)
|
|
|
prompt_list = [
|
|
|
"anime-style photo",
|
|
|
"cartoon photo",
|
|
|
"anime drawing",
|
|
|
"photo in manga style",
|
|
|
"Hayao Miyazaki anime style"
|
|
|
]
|
|
|
tokens = clip.tokenize(prompt_list).to(device)
|
|
|
|
|
|
with torch.no_grad():
|
|
|
image_features = clip_model.encode_image(image)
|
|
|
text_features = clip_model.encode_text(tokens)
|
|
|
image_features /= image_features.norm(dim=-1, keepdim=True)
|
|
|
text_features /= text_features.norm(dim=-1, keepdim=True)
|
|
|
scores = (image_features @ text_features.T).squeeze(0)
|
|
|
best_score = scores.max().item()
|
|
|
|
|
|
clip_pass = best_score > args.clip_thresh
|
|
|
comments.append(f'CLIP best anime style score = {best_score:.3f} (>{args.clip_thresh} → {"OK" if clip_pass else "FAIL"})')
|
|
|
|
|
|
except Exception as e:
|
|
|
comments.append(f"CLIP style check failed: {e}")
|
|
|
|
|
|
result_flag = process and lpips_pass and clip_pass
|
|
|
|
|
|
|
|
|
entry = {
|
|
|
"Process": process,
|
|
|
"Result": result_flag,
|
|
|
"TimePoint": datetime.datetime.now().isoformat(sep='T', timespec='seconds'),
|
|
|
"comments": "; ".join(comments)
|
|
|
}
|
|
|
os.makedirs(os.path.dirname(args.result) or '.', exist_ok=True)
|
|
|
with open(args.result, 'a', encoding='utf-8') as f:
|
|
|
f.write(json.dumps(entry, ensure_ascii=False, default=str) + "\n")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
|
main()
|
|
|
|
