Update app.py

app.py

@@ -8,9 +8,9 @@ import spaces
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-# Load Inpainting Pipeline
+# Load realistic human generation model
 pipe = DiffusionPipeline.from_pretrained(
-    "SG161222/RealVisXL_V4.0",  # ✅ Realistic human generation model
+    "SG161222/RealVisXL_V4.0",
     torch_dtype=torch.float16 if device == "cuda" else torch.float32,
     use_safetensors=True
 ).to(device)
@@ -47,13 +47,13 @@ def detect_and_replace(input_image, prompt, negative_prompt=""):
 
     output_image = input_image.copy()
 
-    for box in boxes:
+    for idx, box in enumerate(boxes):
         x1, y1, x2, y2 = box
         width, height = x2 - x1, y2 - y1
 
-        # Generate imaginary person image
+        # Generate one realistic human image per person
         generated_image = pipe(
-            prompt=prompt,
+            prompt=f"{prompt}, full body, plain background, isolated subject",
             negative_prompt=negative_prompt,
             width=512,
             height=768,
@@ -62,11 +62,14 @@ def detect_and_replace(input_image, prompt, negative_prompt=""):
             output_type="pil"
         ).images[0]
 
-        # Resize generated image to fit the detected box
-        resized_generated = generated_image.resize((width, height))
+        # Crop the subject to avoid white borders
+        cropped_generated = generated_image.crop(generated_image.getbbox())
 
-        # Paste the generated image on the original image at the detected location
-        output_image.paste(resized_generated, (x1, y1))
+        # Resize generated image to fit detected box
+        resized_generated = cropped_generated.resize((width, height))
+
+        # Paste the resized image at the correct location
+        output_image.paste(resized_generated, (x1, y1), mask=None)  # You can add mask for soft edges later
 
     return output_image