Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -15,6 +15,25 @@ Demo for the WaifuDiffusion tagger models
|
|
15 |
|
16 |
HF_TOKEN = os.environ.get("HF_TOKEN", "")
|
17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
MODEL_FILENAME = "model.onnx"
|
19 |
LABEL_FILENAME = "selected_tags.csv"
|
20 |
|
@@ -25,39 +44,11 @@ def parse_args() -> argparse.Namespace:
|
|
25 |
parser.add_argument("--score-character-threshold", type=float, default=1.0)
|
26 |
return parser.parse_args()
|
27 |
|
28 |
-
def
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
old_tags_list = tuple(map(str.strip, old_tags.lower().split(",")))
|
34 |
-
new_tags_list = [tag.strip() for tag in new_tags.split(",")]
|
35 |
-
rules[old_tags_list] = new_tags_list
|
36 |
-
return rules
|
37 |
-
|
38 |
-
def parse_fallback_rules(fallback_text):
|
39 |
-
fallback_rules = {}
|
40 |
-
for line in fallback_text.strip().split("\n"):
|
41 |
-
if "->" in line:
|
42 |
-
expected_tags, fallback_tag = map(str.strip, line.split("->"))
|
43 |
-
expected_tags_list = tuple(map(str.strip, expected_tags.lower().split(",")))
|
44 |
-
fallback_rules[expected_tags_list] = fallback_tag.strip()
|
45 |
-
return fallback_rules
|
46 |
-
|
47 |
-
def apply_replacements(tags, replacement_rules):
|
48 |
-
tags_set = set(tags)
|
49 |
-
for old_tags, new_tags in replacement_rules.items():
|
50 |
-
if set(old_tags).issubset(tags_set):
|
51 |
-
tags_set.difference_update(old_tags)
|
52 |
-
tags_set.update(new_tags)
|
53 |
-
return list(tags_set)
|
54 |
-
|
55 |
-
def apply_fallbacks(tags, fallback_rules):
|
56 |
-
tags_set = set(tags)
|
57 |
-
for expected_tags, fallback_tag in fallback_rules.items():
|
58 |
-
if not any(tag in tags_set for tag in expected_tags):
|
59 |
-
tags_set.add(fallback_tag)
|
60 |
-
return list(tags_set)
|
61 |
|
62 |
class Predictor:
|
63 |
def __init__(self):
|
@@ -75,7 +66,7 @@ class Predictor:
|
|
75 |
|
76 |
csv_path, model_path = self.download_model(model_repo)
|
77 |
tags_df = pd.read_csv(csv_path)
|
78 |
-
self.tag_names, self.general_indexes, self.character_indexes =
|
79 |
|
80 |
model = rt.InferenceSession(model_path)
|
81 |
_, height, width, _ = model.get_inputs()[0].shape
|
@@ -83,70 +74,183 @@ class Predictor:
|
|
83 |
self.last_loaded_repo = model_repo
|
84 |
self.model = model
|
85 |
|
86 |
-
def
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
HF_TOKEN = os.environ.get("HF_TOKEN", "")
|
17 |
|
18 |
+
# Dataset v3 series of models:
|
19 |
+
SWINV2_MODEL_DSV3_REPO = "SmilingWolf/wd-swinv2-tagger-v3"
|
20 |
+
CONV_MODEL_DSV3_REPO = "SmilingWolf/wd-convnext-tagger-v3"
|
21 |
+
VIT_MODEL_DSV3_REPO = "SmilingWolf/wd-vit-tagger-v3"
|
22 |
+
VIT_LARGE_MODEL_DSV3_REPO = "SmilingWolf/wd-vit-large-tagger-v3"
|
23 |
+
EVA02_LARGE_MODEL_DSV3_REPO = "SmilingWolf/wd-eva02-large-tagger-v3"
|
24 |
+
|
25 |
+
# Dataset v2 series of models:
|
26 |
+
MOAT_MODEL_DSV2_REPO = "SmilingWolf/wd-v1-4-moat-tagger-v2"
|
27 |
+
SWIN_MODEL_DSV2_REPO = "SmilingWolf/wd-v1-4-swinv2-tagger-v2"
|
28 |
+
CONV_MODEL_DSV2_REPO = "SmilingWolf/wd-v1-4-convnext-tagger-v2"
|
29 |
+
CONV2_MODEL_DSV2_REPO = "SmilingWolf/wd-v1-4-convnextv2-tagger-v2"
|
30 |
+
VIT_MODEL_DSV2_REPO = "SmilingWolf/wd-v1-4-vit-tagger-v2"
|
31 |
+
|
32 |
+
# IdolSankaku series of models:
|
33 |
+
EVA02_LARGE_MODEL_IS_DSV1_REPO = "deepghs/idolsankaku-eva02-large-tagger-v1"
|
34 |
+
SWINV2_MODEL_IS_DSV1_REPO = "deepghs/idolsankaku-swinv2-tagger-v1"
|
35 |
+
|
36 |
+
# Files to download from the repos
|
37 |
MODEL_FILENAME = "model.onnx"
|
38 |
LABEL_FILENAME = "selected_tags.csv"
|
39 |
|
|
|
44 |
parser.add_argument("--score-character-threshold", type=float, default=1.0)
|
45 |
return parser.parse_args()
|
46 |
|
47 |
+
def load_labels(dataframe) -> list[str]:
|
48 |
+
tag_names = dataframe["name"].tolist()
|
49 |
+
general_indexes = list(np.where(dataframe["category"] == 0)[0])
|
50 |
+
character_indexes = list(np.where(dataframe["category"] == 4)[0])
|
51 |
+
return tag_names, general_indexes, character_indexes
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
|
53 |
class Predictor:
|
54 |
def __init__(self):
|
|
|
66 |
|
67 |
csv_path, model_path = self.download_model(model_repo)
|
68 |
tags_df = pd.read_csv(csv_path)
|
69 |
+
self.tag_names, self.general_indexes, self.character_indexes = load_labels(tags_df)
|
70 |
|
71 |
model = rt.InferenceSession(model_path)
|
72 |
_, height, width, _ = model.get_inputs()[0].shape
|
|
|
74 |
self.last_loaded_repo = model_repo
|
75 |
self.model = model
|
76 |
|
77 |
+
def prepare_image(self, image):
|
78 |
+
# Create a white canvas with the same size as the input image
|
79 |
+
canvas = Image.new("RGBA", image.size, (255, 255, 255))
|
80 |
+
|
81 |
+
# Ensure the input image has an alpha channel for compositing
|
82 |
+
if image.mode != "RGBA":
|
83 |
+
image = image.convert("RGBA")
|
84 |
+
|
85 |
+
# Composite the input image onto the canvas
|
86 |
+
canvas.alpha_composite(image)
|
87 |
+
|
88 |
+
# Convert to RGB (alpha channel is no longer needed)
|
89 |
+
image = canvas.convert("RGB")
|
90 |
+
|
91 |
+
# Resize the image to a square of size (model_target_size x model_target_size)
|
92 |
+
max_dim = max(image.size)
|
93 |
+
padded_image = Image.new("RGB", (max_dim, max_dim), (255, 255, 255))
|
94 |
+
pad_left = (max_dim - image.width) // 2
|
95 |
+
pad_top = (max_dim - image.height) // 2
|
96 |
+
padded_image.paste(image, (pad_left, pad_top))
|
97 |
+
padded_image = padded_image.resize((self.model_target_size, self.model_target_size), Image.BICUBIC)
|
98 |
+
|
99 |
+
# Convert the image to a NumPy array
|
100 |
+
image_array = np.asarray(padded_image, dtype=np.float32)[:, :, ::-1]
|
101 |
+
return np.expand_dims(image_array, axis=0)
|
102 |
+
|
103 |
+
|
104 |
+
def predict(self, images, model_repo, general_thresh, character_thresh):
|
105 |
+
self.load_model(model_repo)
|
106 |
+
results = []
|
107 |
+
|
108 |
+
for image in images:
|
109 |
+
image = self.prepare_image(image)
|
110 |
+
input_name = self.model.get_inputs()[0].name
|
111 |
+
label_name = self.model.get_outputs()[0].name
|
112 |
+
preds = self.model.run([label_name], {input_name: image})[0]
|
113 |
+
|
114 |
+
labels = list(zip(self.tag_names, preds[0].astype(float)))
|
115 |
+
general_res = [x[0] for i, x in enumerate(labels) if i in self.general_indexes and x[1] > general_thresh]
|
116 |
+
character_res = [x[0] for i, x in enumerate(labels) if i in self.character_indexes and x[1] > character_thresh]
|
117 |
+
results.append((general_res, character_res))
|
118 |
+
|
119 |
+
return results
|
120 |
+
|
121 |
+
def main():
|
122 |
+
args = parse_args()
|
123 |
+
predictor = Predictor()
|
124 |
+
|
125 |
+
model_repos = [
|
126 |
+
SWINV2_MODEL_DSV3_REPO,
|
127 |
+
CONV_MODEL_DSV3_REPO,
|
128 |
+
VIT_MODEL_DSV3_REPO,
|
129 |
+
VIT_LARGE_MODEL_DSV3_REPO,
|
130 |
+
EVA02_LARGE_MODEL_DSV3_REPO,
|
131 |
+
# ---
|
132 |
+
MOAT_MODEL_DSV2_REPO,
|
133 |
+
SWIN_MODEL_DSV2_REPO,
|
134 |
+
CONV_MODEL_DSV2_REPO,
|
135 |
+
CONV2_MODEL_DSV2_REPO,
|
136 |
+
VIT_MODEL_DSV2_REPO,
|
137 |
+
# ---
|
138 |
+
SWINV2_MODEL_IS_DSV1_REPO,
|
139 |
+
EVA02_LARGE_MODEL_IS_DSV1_REPO,
|
140 |
+
]
|
141 |
+
|
142 |
+
predefined_tags = ["loli",
|
143 |
+
"oppai_loli",
|
144 |
+
"onee-shota",
|
145 |
+
"incest",
|
146 |
+
"furry",
|
147 |
+
"furry_female",
|
148 |
+
"shota",
|
149 |
+
"male_focus",
|
150 |
+
"signature",
|
151 |
+
"lolita_hairband",
|
152 |
+
"otoko_no_ko",
|
153 |
+
"minigirl",
|
154 |
+
"patreon_username",
|
155 |
+
"babydoll",
|
156 |
+
"monochrome",
|
157 |
+
"happy_birthday",
|
158 |
+
"happy_new_year",
|
159 |
+
"dated",
|
160 |
+
"thought_bubble",
|
161 |
+
"greyscale",
|
162 |
+
"speech_bubble",
|
163 |
+
"english_text",
|
164 |
+
"copyright_name",
|
165 |
+
"twitter_username",
|
166 |
+
"patreon username",
|
167 |
+
"patreon logo",
|
168 |
+
"cover",
|
169 |
+
"content_rating"
|
170 |
+
"cover_page",
|
171 |
+
"doujin_cover",
|
172 |
+
"sex",
|
173 |
+
"artist_name",
|
174 |
+
"watermark",
|
175 |
+
"censored",
|
176 |
+
"bar_censor",
|
177 |
+
"blank_censor",
|
178 |
+
"blur_censor",
|
179 |
+
"light_censor",
|
180 |
+
"mosaic_censoring"]
|
181 |
+
|
182 |
+
with gr.Blocks(title=TITLE) as demo:
|
183 |
+
gr.Markdown(f"<h1 style='text-align: center;'>{TITLE}</h1>")
|
184 |
+
gr.Markdown(DESCRIPTION)
|
185 |
+
|
186 |
+
with gr.Row():
|
187 |
+
with gr.Column():
|
188 |
+
image_files = gr.File(
|
189 |
+
file_types=["image"], label="Upload Images", file_count="multiple",
|
190 |
+
)
|
191 |
+
|
192 |
+
# Wrap the model selection and sliders in an Accordion
|
193 |
+
with gr.Accordion("Advanced Settings", open=False): # Collapsible by default
|
194 |
+
model_repo = gr.Dropdown(
|
195 |
+
model_repos,
|
196 |
+
value=VIT_MODEL_DSV3_REPO,
|
197 |
+
label="Select Model",
|
198 |
+
)
|
199 |
+
general_thresh = gr.Slider(
|
200 |
+
0, 1, step=args.score_slider_step, value=args.score_general_threshold, label="General Tags Threshold"
|
201 |
+
)
|
202 |
+
character_thresh = gr.Slider(
|
203 |
+
0, 1, step=args.score_slider_step, value=args.score_character_threshold, label="Character Tags Threshold"
|
204 |
+
)
|
205 |
+
filter_tags = gr.Textbox(
|
206 |
+
value=", ".join(predefined_tags),
|
207 |
+
label="Filter Tags (comma-separated)",
|
208 |
+
placeholder="Add tags to filter out (e.g., winter, red, from above)",
|
209 |
+
lines=3
|
210 |
+
)
|
211 |
+
|
212 |
+
submit = gr.Button(
|
213 |
+
value="Process Images", variant="primary"
|
214 |
+
)
|
215 |
+
|
216 |
+
with gr.Column():
|
217 |
+
output = gr.Textbox(label="Output", lines=10)
|
218 |
+
|
219 |
+
def process_images(files, model_repo, general_thresh, character_thresh, filter_tags):
|
220 |
+
images = [Image.open(file.name) for file in files]
|
221 |
+
results = predictor.predict(images, model_repo, general_thresh, character_thresh)
|
222 |
+
|
223 |
+
# Parse filter tags
|
224 |
+
filter_set = set(tag.strip().lower() for tag in filter_tags.split(","))
|
225 |
+
|
226 |
+
# Generate formatted output
|
227 |
+
prompts = []
|
228 |
+
for i, (general_tags, character_tags) in enumerate(results):
|
229 |
+
# Replace underscores with spaces for both character and general tags
|
230 |
+
character_part = ", ".join(
|
231 |
+
tag.replace('_', ' ') for tag in character_tags if tag.lower() not in filter_set
|
232 |
+
)
|
233 |
+
general_part = ", ".join(
|
234 |
+
tag.replace('_', ' ') for tag in general_tags if tag.lower() not in filter_set
|
235 |
+
)
|
236 |
+
|
237 |
+
# Construct the prompt based on the presence of character_part
|
238 |
+
if character_part:
|
239 |
+
prompts.append(f"{character_part}, {general_part}")
|
240 |
+
else:
|
241 |
+
prompts.append(general_part)
|
242 |
+
|
243 |
+
# Join all prompts with blank lines
|
244 |
+
return "\n\n".join(prompts)
|
245 |
+
|
246 |
+
submit.click(
|
247 |
+
process_images,
|
248 |
+
inputs=[image_files, model_repo, general_thresh, character_thresh, filter_tags],
|
249 |
+
outputs=output
|
250 |
+
)
|
251 |
+
|
252 |
+
demo.queue(max_size=10)
|
253 |
+
demo.launch()
|
254 |
+
|
255 |
+
if __name__ == "__main__":
|
256 |
+
main()
|