update sketch input

app.py

@@ -125,6 +125,92 @@ def apply_parsing_white_mask_to_person_cv2(
     result_bgr = cv2.cvtColor(result_rgb, cv2.COLOR_RGB2BGR)
     return result_bgr
 
+from typing import Optional, Tuple
+import numpy as np
+from PIL import Image
+
+def clean_and_smooth_parsing_mask(
+    parsing_img: Image.Image,
+    *,
+    white_threshold: int = 128,
+    min_white_area: int = 300,
+    close_ksize: int = 7,
+    open_ksize: int = 3,
+    morph_iters: int = 1,
+    blur_ksize: int = 0,
+) -> Image.Image:
+    """
+    Clean small white blobs and smooth boundaries on a grayscale (0/255) PIL mask.
+
+    Args:
+        parsing_img: PIL.Image in 'L' mode recommended. White=foreground, Black=background.
+        white_threshold: threshold to binarize. >= threshold -> white(255), else black(0).
+        min_white_area: remove connected white components smaller than this pixel area.
+        close_ksize: kernel size for morphological closing (fill small holes, smooth edges).
+        open_ksize: kernel size for morphological opening (remove small spikes/noise).
+        morph_iters: number of iterations for close/open.
+        blur_ksize: optional gaussian blur kernel size (odd number, e.g. 7). 0 disables blur.
+
+    Returns:
+        A PIL.Image (mode 'L') cleaned + smoothed (values 0 or 255).
+    """
+    if not isinstance(parsing_img, Image.Image):
+        raise TypeError("parsing_img must be a PIL.Image.Image")
+
+    # Convert to grayscale
+    img_l = parsing_img.convert("L")
+    arr = np.array(img_l, dtype=np.uint8)
+
+    # Binarize -> 0/255
+    mask = np.where(arr >= white_threshold, 255, 0).astype(np.uint8)
+
+    # --- connected components: remove small white regions ---
+    try:
+        import cv2
+    except ImportError as e:
+        raise ImportError(
+            "This function requires opencv-python (cv2). Install with: pip install opencv-python"
+        ) from e
+
+    # Connected components on binary mask
+    # Note: cv2.connectedComponentsWithStats expects 0/255 (or 0/1), uint8.
+    num_labels, labels, stats, _ = cv2.connectedComponentsWithStats(mask, connectivity=8)
+
+    # Keep only components with area >= min_white_area (label 0 is background)
+    keep = np.zeros_like(mask)
+    for lab in range(1, num_labels):
+        area = int(stats[lab, cv2.CC_STAT_AREA])
+        if area >= min_white_area:
+            keep[labels == lab] = 255
+
+    mask = keep
+
+    # --- boundary smoothing via morphology ---
+    def _odd_or_one(k: int) -> int:
+        k = int(k)
+        if k <= 1:
+            return 1
+        return k if (k % 2 == 1) else (k + 1)
+
+    close_k = _odd_or_one(close_ksize)
+    open_k = _odd_or_one(open_ksize)
+
+    if close_k > 1:
+        k_close = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (close_k, close_k))
+        mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, k_close, iterations=int(morph_iters))
+
+    if open_k > 1:
+        k_open = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (open_k, open_k))
+        mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, k_open, iterations=int(morph_iters))
+
+    # --- optional blur for extra smooth boundary (then re-threshold) ---
+    if blur_ksize and int(blur_ksize) > 1:
+        b = _odd_or_one(int(blur_ksize))
+        mask_blur = cv2.GaussianBlur(mask, (b, b), 0)
+        mask = np.where(mask_blur >= 128, 255, 0).astype(np.uint8)
+
+    return Image.fromarray(mask, mode="L")
+
 
 def compute_hw_from_person(person_path: str):
     img = _imread_or_raise(person_path)
@@ -402,6 +488,15 @@ def run_one(paths: Paths, prompt: str, steps: int = DEFAULT_STEPS):
     parsing_img = res["images"][0] if res.get("images") else None
     if parsing_img is None:
         raise RuntimeError("run_simple_extractor returned no parsing images.")
+        
+    parsing_img = clean_and_smooth_parsing_mask(
+        parsing_img,
+        min_white_area=300,   # 작은 흰색 덩어리 제거 강도
+        close_ksize=9,        # 경계 매끈 + 작은 구멍 메움
+        open_ksize=3,         # 잔 노이즈 제거
+        morph_iters=1,
+        blur_ksize=7,         # 더 부드럽게 (원치 않으면 0)
+    )
 
     # -------------------------------------------------
     # ✅ (2) UI sketch 업로드는 optional