app.py

app.py

@@ -7,7 +7,7 @@ import os
 import time
 
 from utils.dl_utils import dl_cn_model, dl_cn_config, dl_tagger_model, dl_lora_model
-from utils.image_utils import resize_image_aspect_ratio, base_generation, canny_process
+from utils.image_utils import resize_image_aspect_ratio, base_generation, line_process
 
 from utils.prompt_utils import execute_prompt, remove_color, remove_duplicates
 from utils.tagger import modelLoad, analysis
@@ -32,6 +32,7 @@ def load_model(lora_dir, cn_dir):
     dtype = torch.float16
     vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
     controlnet = ControlNetModel.from_pretrained(cn_dir, torch_dtype=dtype, use_safetensors=True)
+
     pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
         "cagliostrolab/animagine-xl-3.1", controlnet=controlnet, vae=vae, torch_dtype=torch.float16
     )
@@ -94,10 +95,10 @@ class Img2Img:
             tags_list = remove_color(tags)
         return tags_list
 
-    def _make_canny(self, img_path, canny_threshold1, canny_threshold2):
-        threshold1 = int(canny_threshold1)
-        threshold2 = int(canny_threshold2)
-        return canny_process(img_path, threshold1, threshold2)
+    def _make_line(self, img_path, sigma, gamma):
+        sigma = float(sigma )
+        gamma = float(gamma)
+        return line_process(img_path, sigma, gamma)
 
     def layout(self):
         css = """
@@ -113,9 +114,8 @@ class Img2Img:
                     self.input_image_path = gr.Image(label="input_image", type='filepath')
                     self.canny_image = gr.Image(label="canny_image", type='pil')
                     with gr.Row():
-                        canny_threshold1 = gr.Slider(minimum=0, value=20, maximum=253, show_label=False)
-                        gr.HTML(value="<span>/</span>", show_label=False)
-                        canny_threshold2 = gr.Slider(minimum=0, value=120, maximum=254, show_label=False)
+                        line_sigma = gr.Slider(minimum=0.1, value=1.4, maximum=3.0, show_label=False)
+                        line_gamma = gr.Slider(minimum=0.5, value=0.98, maximum=2.0, show_label=False)
                         canny_generate_button = gr.Button("canny_generate")
 
                     self.prompt = gr.Textbox(label="prompt", lines=3)
@@ -130,8 +130,8 @@ class Img2Img:
                     self.output_image = gr.Image(type="pil", label="output_image")
 
             canny_generate_button.click(
-                        self._make_canny,
-                        inputs=[self.input_image_path, canny_threshold1, canny_threshold2],
+                        self._make_line,
+                        inputs=[self.input_image_path, line_sigma, line_gamma],
                         outputs=self.canny_image
             )
 

utils/dl_utils.py

@@ -11,7 +11,7 @@ import cv2
 def dl_cn_model(model_dir):
     folder = model_dir
     file_name = 'diffusion_pytorch_model.safetensors'
-    url = "https://huggingface.co/kataragi/controlnet_canny/resolve/main/Kataragi_cannyXL-fp16.safetensors"
+    url = "https://huggingface.co/2vXpSwA7/iroiro-lora/resolve/main/test_controlnet2/CN-anytest_v3-50000_fp16.safetensors"
     file_path = os.path.join(folder, file_name)
     if not os.path.exists(file_path):
         response = requests.get(url, allow_redirects=True)

utils/image_utils.py

@@ -27,6 +27,50 @@ def canny_process(image_path, threshold1, threshold2):
     return canny
 
 
+def line_process(image_path, sigma, gamma):
+    def DoG_filter(image, kernel_size=0, sigma=1.0, k_sigma=2.0, gamma=1.5):
+        g1 = cv2.GaussianBlur(image, (kernel_size, kernel_size), sigma)
+        g2 = cv2.GaussianBlur(image, (kernel_size, kernel_size), sigma * k_sigma)
+        return g1 - gamma * g2
+
+    def XDoG_filter(image, kernel_size=0, sigma=1.4, k_sigma=1.6, epsilon=0, phi=10, gamma=0.98):
+        epsilon /= 255
+        dog = DoG_filter(image, kernel_size, sigma, k_sigma, gamma)
+        dog /= dog.max()
+        e = 1 + np.tanh(phi * (dog - epsilon))
+        e[e >= 1] = 1
+        return (e * 255).astype('uint8')
+
+    def binarize_image(image):
+        _, binarized = cv2.threshold(image, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
+        return binarized
+
+    def process_XDoG(image, kernel_size=0, sigma=1.4, k_sigma=1.6, epsilon=0, phi=10, gamma=0.98):
+        xdog_image = XDoG_filter(image, kernel_size, sigma, k_sigma, epsilon, phi, gamma)
+        binarized_image = binarize_image(xdog_image)
+        final_image_pil = Image.fromarray(binarized_image)
+        return final_image_pil
+
+    # 画像を開き、RGBA形式に変換して透過情報を保持
+    img = Image.open(image_path)
+    img = img.convert("RGBA")
+
+    canvas_image = Image.new('RGBA', img.size, (255, 255, 255, 255))
+    
+    # 画像をキャンバスにペーストし、透過部分が白色になるように設定
+    canvas_image.paste(img, (0, 0), img)
+
+    # RGBAからRGBに変換し、透過部分を白色にする
+    image_pil = canvas_image.convert("RGB")
+    
+    # OpenCVが扱える形式に変換
+    image_cv = cv2.cvtColor(np.array(image_pil), cv2.COLOR_RGB2BGR)
+    image_gray = cv2.cvtColor(image_cv, cv2.COLOR_BGR2GRAY)
+    inv_Line = process_XDoG(image_gray, kernel_size=0, sigma=sigma, k_sigma=1.6, epsilon=0, phi=10, gamma=gamma)
+    return inv_Line
+
+
+
 def resize_image_aspect_ratio(image):
     # 元の画像サイズを取得
     original_width, original_height = image.size