Spaces:
Running
on
L40S
Running
on
L40S
Rishi Desai
commited on
Commit
·
8308bbd
1
Parent(s):
632672e
init dump
Browse files
.env
ADDED
File without changes
|
main.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from utils import crop_face, upscale_image
|
4 |
+
|
5 |
+
def parse_args():
|
6 |
+
parser = argparse.ArgumentParser(description='Face Enhancement Tool')
|
7 |
+
parser.add_argument('--input', type=str, required=True, help='Path to the input image')
|
8 |
+
parser.add_argument('--crop', action='store_true', help='Whether to crop the image')
|
9 |
+
parser.add_argument('--upscale', action='store_true', help='Whether to upscale the image')
|
10 |
+
parser.add_argument('--output', type=str, required=True, help='Path to save the output image')
|
11 |
+
args = parser.parse_args()
|
12 |
+
|
13 |
+
# Validate input file exists
|
14 |
+
if not os.path.exists(args.input):
|
15 |
+
parser.error(f"Input file does not exist: {args.input}")
|
16 |
+
|
17 |
+
# Validate output directory exists
|
18 |
+
output_dir = os.path.dirname(args.output)
|
19 |
+
if output_dir and not os.path.exists(output_dir):
|
20 |
+
parser.error(f"Output directory does not exist: {output_dir}")
|
21 |
+
|
22 |
+
return args
|
23 |
+
|
24 |
+
def main():
|
25 |
+
args = parse_args()
|
26 |
+
print(f"Processing image: {args.input}")
|
27 |
+
print(f"Crop enabled: {args.crop}")
|
28 |
+
print(f"Upscale enabled: {args.upscale}")
|
29 |
+
print(f"Output will be saved to: {args.output}")
|
30 |
+
|
31 |
+
face_image = args.input
|
32 |
+
if args.crop:
|
33 |
+
crop_face(args.input, "./scratch/cropped_face.png")
|
34 |
+
face_image = "./scratch/cropped_face.png"
|
35 |
+
|
36 |
+
if args.upscale:
|
37 |
+
upscale_image(face_image, "./scratch/upscaled_face.png")
|
38 |
+
face_image = "./scratch/upscaled_face.png"
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
if __name__ == "__main__":
|
43 |
+
main()
|
utils.py
ADDED
@@ -0,0 +1,190 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
from PIL import Image
|
5 |
+
import sys
|
6 |
+
import cv2
|
7 |
+
import base64
|
8 |
+
import aiohttp
|
9 |
+
from fal import Client as FalClient
|
10 |
+
sys.path.append('./ComfyUI_AutoCropFaces')
|
11 |
+
from dotenv import load_dotenv
|
12 |
+
load_dotenv()
|
13 |
+
from Pytorch_Retinaface.pytorch_retinaface import Pytorch_RetinaFace
|
14 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
15 |
+
from transformers import CLIPProcessor, CLIPModel
|
16 |
+
import gc
|
17 |
+
|
18 |
+
|
19 |
+
CACHE_DIR = '/workspace/huggingface_cache'
|
20 |
+
|
21 |
+
os.environ["HF_HOME"] = CACHE_DIR
|
22 |
+
os.makedirs(CACHE_DIR, exist_ok=True)
|
23 |
+
|
24 |
+
device = "cuda"
|
25 |
+
|
26 |
+
def clear_cuda_memory():
|
27 |
+
"""Aggressively clear CUDA memory"""
|
28 |
+
gc.collect()
|
29 |
+
torch.cuda.empty_cache()
|
30 |
+
torch.cuda.synchronize()
|
31 |
+
|
32 |
+
|
33 |
+
def load_vision_models():
|
34 |
+
print("Loading CLIP and Florence models...")
|
35 |
+
# Load CLIP
|
36 |
+
clip_model = CLIPModel.from_pretrained(
|
37 |
+
"openai/clip-vit-large-patch14",
|
38 |
+
cache_dir=CACHE_DIR
|
39 |
+
).to(device)
|
40 |
+
clip_processor = CLIPProcessor.from_pretrained(
|
41 |
+
"openai/clip-vit-large-patch14",
|
42 |
+
cache_dir=CACHE_DIR
|
43 |
+
)
|
44 |
+
|
45 |
+
# Load Florence
|
46 |
+
florence_model = AutoModelForCausalLM.from_pretrained(
|
47 |
+
"microsoft/Florence-2-large",
|
48 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
49 |
+
trust_remote_code=True,
|
50 |
+
cache_dir=CACHE_DIR
|
51 |
+
).to(device)
|
52 |
+
florence_processor = AutoProcessor.from_pretrained(
|
53 |
+
"microsoft/Florence-2-large",
|
54 |
+
trust_remote_code=True,
|
55 |
+
cache_dir=CACHE_DIR
|
56 |
+
)
|
57 |
+
|
58 |
+
return {
|
59 |
+
'clip_model': clip_model,
|
60 |
+
'clip_processor': clip_processor,
|
61 |
+
'florence_model': florence_model,
|
62 |
+
'florence_processor': florence_processor,
|
63 |
+
}
|
64 |
+
|
65 |
+
|
66 |
+
def generate_caption(image):
|
67 |
+
vision_models = load_vision_models()
|
68 |
+
|
69 |
+
# Ensure the image is a PIL Image
|
70 |
+
if not isinstance(image, Image.Image):
|
71 |
+
image = Image.fromarray(image)
|
72 |
+
|
73 |
+
# Convert the image to RGB if it has an alpha channel
|
74 |
+
if image.mode == 'RGBA':
|
75 |
+
image = image.convert('RGB')
|
76 |
+
|
77 |
+
prompt = "<DETAILED_CAPTION>"
|
78 |
+
inputs = vision_models['florence_processor'](
|
79 |
+
text=prompt,
|
80 |
+
images=image,
|
81 |
+
return_tensors="pt"
|
82 |
+
).to(device, torch.float16 if torch.cuda.is_available() else torch.float32)
|
83 |
+
|
84 |
+
generated_ids = vision_models['florence_model'].generate(
|
85 |
+
input_ids=inputs["input_ids"],
|
86 |
+
pixel_values=inputs["pixel_values"],
|
87 |
+
max_new_tokens=1024,
|
88 |
+
num_beams=3,
|
89 |
+
do_sample=False
|
90 |
+
)
|
91 |
+
generated_text = vision_models['florence_processor'].batch_decode(generated_ids, skip_special_tokens=True)[0]
|
92 |
+
parsed_answer = vision_models['florence_processor'].post_process_generation(
|
93 |
+
generated_text, task="<DETAILED_CAPTION>",
|
94 |
+
image_size=(image.width, image.height)
|
95 |
+
)
|
96 |
+
|
97 |
+
clear_cuda_memory()
|
98 |
+
return parsed_answer['<DETAILED_CAPTION>']
|
99 |
+
|
100 |
+
|
101 |
+
def crop_face(image_path, output_dir, output_name, scale_factor=4.0):
|
102 |
+
image = Image.open(image_path).convert("RGB")
|
103 |
+
|
104 |
+
img_raw = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
|
105 |
+
img_raw = img_raw.astype(np.float32)
|
106 |
+
|
107 |
+
rf = Pytorch_RetinaFace(
|
108 |
+
cfg='mobile0.25',
|
109 |
+
pretrained_path='./weights/mobilenet0.25_Final.pth',
|
110 |
+
confidence_threshold=0.02,
|
111 |
+
nms_threshold=0.4,
|
112 |
+
vis_thres=0.6
|
113 |
+
)
|
114 |
+
|
115 |
+
dets = rf.detect_faces(img_raw)
|
116 |
+
print("Dets: ", dets)
|
117 |
+
|
118 |
+
# Instead of asserting, handle multiple faces gracefully
|
119 |
+
if len(dets) == 0:
|
120 |
+
print("No faces detected!")
|
121 |
+
return False
|
122 |
+
|
123 |
+
# If multiple faces detected, use the one with highest confidence
|
124 |
+
if len(dets) > 1:
|
125 |
+
print(f"Warning: {len(dets)} faces detected, using the one with highest confidence")
|
126 |
+
# Assuming dets is a list of [bbox, landmark, score] and we want to sort by score
|
127 |
+
dets = sorted(dets, key=lambda x: x[2], reverse=True) # Sort by confidence score
|
128 |
+
# Just keep the highest confidence detection
|
129 |
+
dets = [dets[0]]
|
130 |
+
|
131 |
+
# Pass the scale_factor to center_and_crop_rescale for adjustable crop size
|
132 |
+
try:
|
133 |
+
# Unpack the tuple correctly - the function returns (cropped_imgs, bbox_infos)
|
134 |
+
cropped_imgs, bbox_infos = rf.center_and_crop_rescale(img_raw, dets, shift_factor=0.45, scale_factor=scale_factor)
|
135 |
+
|
136 |
+
# Check if we got any cropped images
|
137 |
+
if not cropped_imgs or len(cropped_imgs) == 0:
|
138 |
+
print("No cropped images returned")
|
139 |
+
return False
|
140 |
+
|
141 |
+
# Use the first cropped face image directly - it's not nested
|
142 |
+
img_to_save = cropped_imgs[0]
|
143 |
+
|
144 |
+
os.makedirs(output_dir, exist_ok=True)
|
145 |
+
cv2.imwrite(os.path.join(output_dir, output_name), img_to_save)
|
146 |
+
print(f"Saved: {output_name}")
|
147 |
+
return True
|
148 |
+
|
149 |
+
except Exception as e:
|
150 |
+
print(f"Error during face cropping: {e}")
|
151 |
+
return False
|
152 |
+
|
153 |
+
async def upscale_image(image_path, output_path):
|
154 |
+
"""Upscale an image using fal.ai's RealESRGAN model"""
|
155 |
+
fal_client = FalClient()
|
156 |
+
|
157 |
+
# Read and encode the image
|
158 |
+
with open(image_path, "rb") as image_file:
|
159 |
+
encoded_image = base64.b64encode(image_file.read()).decode('utf-8')
|
160 |
+
data_uri = f"data:image/jpeg;base64,{encoded_image}"
|
161 |
+
|
162 |
+
try:
|
163 |
+
# Submit the upscaling request
|
164 |
+
handler = await fal_client.submit_async(
|
165 |
+
"fal-ai/real-esrgan",
|
166 |
+
arguments={
|
167 |
+
"image_url": data_uri,
|
168 |
+
"scale": 2,
|
169 |
+
"model": "RealESRGAN_x4plus",
|
170 |
+
"output_format": "png",
|
171 |
+
"face": True
|
172 |
+
},
|
173 |
+
)
|
174 |
+
result = await handler.get()
|
175 |
+
|
176 |
+
# Download and save the upscaled image
|
177 |
+
image_url = result['image_url']
|
178 |
+
async with aiohttp.ClientSession() as session:
|
179 |
+
async with session.get(image_url) as response:
|
180 |
+
if response.status == 200:
|
181 |
+
with open(output_path, 'wb') as f:
|
182 |
+
f.write(await response.read())
|
183 |
+
return True
|
184 |
+
else:
|
185 |
+
print(f"Failed to download upscaled image: {response.status}")
|
186 |
+
return False
|
187 |
+
|
188 |
+
except Exception as e:
|
189 |
+
print(f"Error during upscaling: {e}")
|
190 |
+
return False
|