first commit

app.py

@@ -0,0 +1,80 @@
+import gradio as gr
+import sys
+from utils import load_lora_model
+from bg_alpha import adjust_transparency
+from lineart import get_pipe
+import os
+from PIL import Image
+import spaces
+
+path = os.getcwd()
+output_dir = f"{path}/output"
+lora_dir = f"{path}/models/lora"
+
+load_lora_model(lora_dir)
+
+pipe = get_pipe(lora_dir)
+
+@spaces.GPU()
+def generate(prompt, negative_prompt):
+    default_pos = "((white background)), lineart, <lora:sdxl_BWLine:1.0>, monochrome, "
+    default_neg = ""
+    prompt = default_pos + prompt 
+    negative_prompt = default_neg + negative_prompt
+
+    width, height = 1024, 1024
+    color = (255, 255, 255)
+    white_bg =  Image.new("RGB", (width, height), color)
+
+    image = pipe(
+                prompt=prompt,
+                negative_prompt = negative_prompt,
+                image=[white_bg],
+                num_inference_steps=50,
+                controlnet_conditioning_scale=[0.1]
+            ).images[0]
+    
+    return image
+
+
+class webui:
+    def __init__(self):
+        self.demo = gr.Blocks()
+
+    def process(self, pos_prompt, neg_prompt):
+        image = generate(pos_prompt, neg_prompt)
+        image = adjust_transparency(image)
+        return [image]
+
+    def launch(self, share):
+        with self.demo:
+            with gr.Column():
+                    with gr.Tab("output"):
+                        output_0 = gr.Gallery(format="png")
+                    #output_file = gr.File()
+            with gr.Column():
+                pos_prompt = gr.Textbox(value="1girl, cute, kawaii,  medium breasts,  medium hair,  smile,  mini skirt,  best quality, very aesthetic," ,max_lines=1000, label="positive prompt")                    
+                neg_prompt = gr.Textbox(value="bold line, multiple people," ,max_lines=1000, label="negative prompt")
+
+                submit = gr.Button(value="Start")
+                
+                
+            submit.click(
+                self.process, 
+                inputs=[pos_prompt, neg_prompt], #[input_image, pos_prompt, neg_prompt, alpha_th, thickness, reference_image], 
+                outputs=[output_0]
+            )
+
+        self.demo.queue()
+        self.demo.launch(share=share)
+
+
+if __name__ == "__main__":
+    ui = webui()
+    if len(sys.argv) > 1:
+        if sys.argv[1] == "share":
+            ui.launch(share=True)
+        else:
+            ui.launch(share=False)
+    else:
+        ui.launch(share=False)

bg_alpha.py

@@ -0,0 +1,60 @@
+from PIL import Image, ImageFilter
+import numpy as np
+
+def erode(image, iterations=1):
+    # モルフォロジー操作のためのカーネルを定義
+    kernel = np.array([[0, 1, 0],
+                       [1, 1, 1],
+                       [0, 1, 0]], dtype=np.uint8)
+    
+    image_np = np.array(image)
+    for _ in range(iterations):
+        eroded_image = np.zeros_like(image_np)
+        for i in range(1, image_np.shape[0] - 1):
+            for j in range(1, image_np.shape[1] - 1):
+                region = image_np[i-1:i+2, j-1:j+2]
+                eroded_image[i, j] = np.min(region + (1 - kernel) * 255)
+        image_np = eroded_image
+
+    return Image.fromarray(image_np.astype(np.uint8))
+
+def convert_non_white_to_black(image):
+    # 画像をNumPy配列に変換
+    image_np = np.array(image)
+    
+    # 完全に白でないピクセルをすべて黒にする
+    #image_np[image_np < 250] = 0
+    image_np[image_np > 200] = 255
+    
+    # NumPy配列を画像に変換
+    return Image.fromarray(image_np)
+
+
+
+def adjust_transparency(image):
+    # 画像を読み込み、グレースケールに変換
+    image.save("tmp.png")
+    image = image.convert('L')
+    image = convert_non_white_to_black(image)
+    image = image.filter(ImageFilter.SMOOTH)
+    
+    result = Image.new('RGBA', image.size)
+    for x in range(image.width):
+        for y in range(image.height):
+            # グレースケール値を取得
+            gray = image.getpixel((x, y))
+
+            alpha = 255
+            # 透明度の設定
+            if gray == 0:
+                alpha = 0  # 完全に透明
+            else:
+                alpha = 255-gray  # 完全に不透明
+
+            # 新しい画像にピクセルを設定
+            result.putpixel((x, y), (0, 0, 0, alpha))
+
+    return result
+
+def remove_bg(image):
+    image = adjust_transparency(image)

lineart.py

@@ -0,0 +1,21 @@
+from diffusers import AutoPipelineForText2Image
+import torch
+
+from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
+from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel, AutoencoderKL
+import torch
+
+from PIL import Image
+
+
+def get_pipe(lora_dir):
+  controlnets = [ControlNetModel.from_pretrained("mattyamonaca/white2line", torch_dtype=torch.float16)]
+
+  vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+  pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
+      "cagliostrolab/animagine-xl-3.1", controlnet=controlnets, vae=vae, torch_dtype=torch.float16
+  )
+
+  pipe.enable_model_cpu_offload()
+  pipe.load_lora_weights(lora_dir, weight_name="sdxl_BWLine.safetensors")
+  return pipe

utils.py

@@ -0,0 +1,33 @@
+import random
+import string
+import os
+
+import requests
+from tqdm import tqdm
+
+
+def randomname(n):
+   randlst = [random.choice(string.ascii_letters + string.digits) for i in range(n)]
+   return ''.join(randlst)
+
+def load_lora_model(model_dir):
+  folder = model_dir
+  file_name = 'sdxl_BWLine.safetensors'
+  url = "https://huggingface.co/tori29umai/lineart/resolve/main/sdxl_BWLine.safetensors"
+  
+
+  file_path = os.path.join(folder, file_name)
+  if not os.path.exists(file_path):
+    response = requests.get(url, stream=True)
+
+    total_size = int(response.headers.get('content-length', 0))
+    with open(file_path, 'wb') as f, tqdm(
+            desc=file_name,
+            total=total_size,
+            unit='iB',
+            unit_scale=True,
+            unit_divisor=1024,
+        ) as bar:
+        for data in response.iter_content(chunk_size=1024):
+            size = f.write(data)
+            bar.update(size)