feat product 后处理图片size改为320*700

fix

app/service/generate_image/service_generate_product_image.py

@@ -15,7 +15,7 @@ import cv2
 import numpy as np
 import redis
 import tritonclient.grpc as grpcclient
-from PIL import Image, ImageOps
+from PIL import Image
 from tritonclient.utils import np_to_triton_dtype
 
 from app.core.config import *
@@ -41,7 +41,7 @@ class GenerateProductImage:
         self.batch_size = 1
         self.product_type = request_data.product_type
         self.prompt = request_data.prompt
-        self.image, self.image_size = pre_processing_image(request_data.image_url)
+        self.image, self.image_size, self.left, self.top = pre_processing_image(request_data.image_url)
         self.tasks_id = request_data.tasks_id
         self.user_id = self.tasks_id[self.tasks_id.rfind('-') + 1:]
         self.gen_product_data = {'tasks_id': self.tasks_id, 'status': 'PENDING', 'message': "pending", 'image_url': ''}
@@ -55,12 +55,10 @@ class GenerateProductImage:
             self.redis_client.set(self.tasks_id, json.dumps(self.gen_product_data))
         else:
             # pil图像转成numpy数组
-            if self.product_type == "single":
+            image = result.as_numpy("generated_inpaint_image")
-                image = result.as_numpy("generated_cnet_image")
-            else:
-                image = result.as_numpy("generated_inpaint_image")
             image_result = Image.fromarray(np.squeeze(image.astype(np.uint8))).resize(self.image_size)
-            image_url = upload_SDXL_image(image_result, user_id=self.user_id, category=f"{self.category}", file_name=f"{self.tasks_id}.png")
+            cropped_image = post_processing_image(image_result, self.left, self.top)
+            image_url = upload_SDXL_image(cropped_image, user_id=self.user_id, category=f"{self.category}", file_name=f"{self.tasks_id}.png")
             self.gen_product_data['status'] = "SUCCESS"
             self.gen_product_data['message'] = "success"
             self.gen_product_data['image_url'] = str(image_url)
@@ -74,16 +72,16 @@ class GenerateProductImage:
         try:
             prompts = [self.prompt] * self.batch_size
             self.image = cv2.cvtColor(self.image, cv2.COLOR_BGR2RGB)
-            self.image = cv2.resize(self.image, (512, 768))
+            self.image = cv2.resize(self.image, (1024, 1024))
             images = [self.image.astype(np.uint8)] * self.batch_size
 
             if self.product_type == "single":
                 text_obj = np.array(prompts, dtype="object").reshape(-1, 1)
-                image_obj = np.array(images, dtype=np.uint8).reshape((-1, 768, 512, 3))
+                image_obj = np.array(images, dtype=np.uint8).reshape((-1, 1024, 1024, 3))
                 image_strength_obj = np.array(self.image_strength, dtype=np.float32).reshape(-1, 1)
             else:
-                text_obj = np.array(prompts, dtype="object").reshape(1)
+                text_obj = np.array(prompts, dtype="object").reshape((1))
-                image_obj = np.array(images, dtype=np.uint8).reshape((768, 512, 3))
+                image_obj = np.array(images, dtype=np.uint8).reshape((1024, 1024, 3))
                 image_strength_obj = np.array(self.image_strength, dtype=np.float32).reshape((1))
 
             # 假设 prompts、images 和 self.image_strength 已经定义
@@ -94,11 +92,12 @@ class GenerateProductImage:
 
             input_text.set_data_from_numpy(text_obj)
             input_image.set_data_from_numpy(image_obj)
-            inputs = [input_text, input_image, input_image_strength]
             input_image_strength.set_data_from_numpy(image_strength_obj)
 
+            inputs = [input_text, input_image, input_image_strength]
+
             if self.product_type == "single":
-                ctx = self.grpc_client.async_infer(model_name=GPI_MODEL_NAME_SINGLE, inputs=inputs, callback=self.callback)
+                ctx = self.grpc_client.async_infer(model_name="stable_diffusion_xl_cnet_inpaint", inputs=inputs, callback=self.callback)
             else:
                 ctx = self.grpc_client.async_infer(model_name=GPI_MODEL_NAME_OVERALL, inputs=inputs, callback=self.callback)
 
@@ -136,22 +135,13 @@ def infer_cancel(tasks_id):
 
 def pre_processing_image(image_url):
     image = oss_get_image(bucket=image_url.split('/')[0], object_name=image_url[image_url.find('/') + 1:], data_type="PIL")
+    # resize 原图至1024*1024
+    image = image.resize((int(1024 / image.height * image.width), 1024))
+
     # 原始图片的尺寸
     width, height = image.size
 
-    # 计算长宽比为 3:2 的新尺寸
+    new_height, new_width = 1024, 1024
-    desired_ratio = 2 / 3
-    current_ratio = width / height
-
-    if current_ratio > desired_ratio:
-        # 原始图片更宽，需要在上下添加 padding
-        new_width = width
-        new_height = int(width / desired_ratio)
-    else:
-        # 原始图片更高或者长宽比已经为 3:2
-        new_height = height
-        new_width = int(height * desired_ratio)
-
     # 创建一个新的画布，大小为添加 padding 后的尺寸，并设置为白色背景
     pad_image = Image.new('RGBA', (new_width, new_height), (0, 0, 0, 0))
 
@@ -160,9 +150,9 @@ def pre_processing_image(image_url):
     top = (new_height - height) // 2
     pad_image.paste(image, (left, top))
 
-    # 将画布 resize 成宽度 500，长度 750
+    # 将画布 resize 成宽度 1024，长度 1024
-    resized_image = pad_image.resize((500, 750))
+    resized_image = pad_image.resize((1024, 1024))
-    image_size = (512, 768)
+    image_size = (1024, 1024)
 
     if resized_image.mode in ('RGBA', 'LA') or (resized_image.mode == 'P' and 'transparency' in resized_image.info):
         # 创建白色背景
@@ -171,16 +161,29 @@ def pre_processing_image(image_url):
         background.paste(resized_image, mask=resized_image.split()[3])
         image = np.array(background)
         image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-    return image, image_size
+    return image, image_size, left, top
+
+
+def post_processing_image(image, left, top):
+    resized_image = image.resize((int(image.width * (768 / image.height)), 768))
+    # 计算裁剪的坐标
+    left = (resized_image.width - 512) // 2
+    upper = 0
+    right = left + 512
+    lower = 768
+
+    # 进行裁剪
+    cropped_image = resized_image.crop((left, upper, right, lower))
+    return cropped_image
 
 
 if __name__ == '__main__':
     rd = GenerateProductImageModel(
         tasks_id="123-89",
         # prompt="",
-        image_strength=0.9,
+        image_strength=0.7,
-        prompt=" the best quality, masterpiece. detailed, high-res, simple background, studio photography, extremely detailed, updo, detailed face, face, close-up, HDR, UHD, 8K realistic, Highly detailed, simple background, Studio lighting",
+        prompt="The best quality, masterpiece,outwear, 8K realistic, HUD",
-        image_url="aida-results/result_00097282-ebb2-11ee-a822-b48351119060.png",
+        image_url="aida-results/result_53381ada-ac64-11ef-ae9d-0242ac150002.png",
         product_type="overall"
     )
     server = GenerateProductImage(rd)