update examples

app.py

@@ -23,14 +23,12 @@
 # import imageio
 # from PIL import Image, ImageOps
 # from transformers import pipeline
-
 # from huggingface_hub import hf_hub_download
 
 # import diffusers3
 # print("[BOOT] diffusers3 loaded from:", getattr(diffusers3, "__file__", "<?>"), flush=True)
 
 # from diffusers import UniPCMultistepScheduler, AutoencoderKL, UNet2DConditionModel
-
 # from diffusers3.models.controlnet import ControlNetModel
 # from diffusers3.pipelines.controlnet.pipeline_controlnet_sd_xl_img2img_img import (
 #     StableDiffusionXLControlNetImg2ImgPipeline,
@@ -89,7 +87,7 @@
 # @dataclass
 # class Paths:
 #     person_path: str
-#     depth_path: Optional[str]  # ✅ sketch(guide) optional
+#     depth_path: Optional[str]  # sketch(guide) optional
 #     style_path: str
 #     output_path: str
 
@@ -101,22 +99,58 @@
 #     return img
 
 
+# def _pad_or_crop_to_width_np(arr: np.ndarray, target_width: int, pad_value):
+#     """
+#     arr: HxWxC or HxW
+#     target_width로 center crop 또는 좌/우 padding(비대칭 포함)해서 정확히 맞춤.
+#     """
+#     if arr.ndim not in (2, 3):
+#         raise ValueError(f"arr must be 2D or 3D, got shape={arr.shape}")
+
+#     h = arr.shape[0]
+#     w = arr.shape[1]
+
+#     if w == target_width:
+#         return arr
+
+#     if w > target_width:
+#         left = (w - target_width) // 2
+#         return arr[:, left:left + target_width] if arr.ndim == 2 else arr[:, left:left + target_width, :]
+
+#     # w < target_width: pad
+#     total = target_width - w
+#     left = total // 2
+#     right = total - left  # ✅ remainder를 오른쪽이 먹어서 항상 정확히 target_width
+
+#     if arr.ndim == 2:
+#         return cv2.copyMakeBorder(
+#             arr, 0, 0, left, right,
+#             borderType=cv2.BORDER_CONSTANT,
+#             value=pad_value,
+#         )
+#     else:
+#         # 3채널일 때 value는 스칼라 or [b,g,r]/[r,g,b] 모두 허용되는데,
+#         # 여기선 arr가 RGB/BGR인지 호출자가 정해줌.
+#         return cv2.copyMakeBorder(
+#             arr, 0, 0, left, right,
+#             borderType=cv2.BORDER_CONSTANT,
+#             value=pad_value,
+#         )
+
+
 # def apply_parsing_white_mask_to_person_cv2(
 #     person_pil: Image.Image,
 #     parsing_img: Image.Image
 # ) -> np.ndarray:
 #     person_rgb = np.array(person_pil.convert("RGB"), dtype=np.uint8)
-
 #     mask = np.array(parsing_img.convert("L"), dtype=np.uint8)
 
 #     if mask.shape[:2] != person_rgb.shape[:2]:
 #         mask = cv2.resize(mask, (person_rgb.shape[1], person_rgb.shape[0]), interpolation=cv2.INTER_NEAREST)
 
 #     white_mask = (mask == 255)
-
 #     result_rgb = np.full_like(person_rgb, 255, dtype=np.uint8)
 #     result_rgb[white_mask] = person_rgb[white_mask]
-
 #     result_bgr = cv2.cvtColor(result_rgb, cv2.COLOR_RGB2BGR)
 #     return result_bgr
 
@@ -134,19 +168,15 @@
 #     if not isinstance(parsing_img, Image.Image):
 #         raise TypeError("parsing_img must be a PIL.Image.Image")
 
-#     img_l = parsing_img.convert("L")
-#     arr = np.array(img_l, dtype=np.uint8)
-
+#     arr = np.array(parsing_img.convert("L"), dtype=np.uint8)
 #     mask = np.where(arr >= white_threshold, 255, 0).astype(np.uint8)
 
 #     num_labels, labels, stats, _ = cv2.connectedComponentsWithStats(mask, connectivity=8)
-
 #     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
 
 #     def _odd_or_one(k: int) -> int:
@@ -219,20 +249,7 @@
 #     _, m = cv2.threshold(m, 127, 255, cv2.THRESH_BINARY)
 
 #     target_width = 1024
-#     h, w = m.shape[:2]
-
-#     if w < target_width:
-#         total_padding = target_width - w
-#         left_padding = total_padding // 2
-#         right_padding = total_padding - left_padding
-#         m = cv2.copyMakeBorder(
-#             m, 0, 0, left_padding, right_padding,
-#             borderType=cv2.BORDER_CONSTANT,
-#             value=0,
-#         )
-#     elif w > target_width:
-#         left = (w - target_width) // 2
-#         m = m[:, left:left + target_width]
+#     m = _pad_or_crop_to_width_np(m, target_width, pad_value=0)
 
 #     kernel = np.ones((17, 17), np.uint8)
 #     m = cv2.dilate(m, kernel, iterations=1)
@@ -249,7 +266,6 @@
 #         raise RuntimeError("Global H/W not set. Call run_one() first.")
 
 #     depth_img = _imread_or_raise(depth_path, 0)
-
 #     inverted_depth = cv2.bitwise_not(depth_img)
 #     contours, _ = cv2.findContours(inverted_depth, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
 
@@ -257,19 +273,9 @@
 #     cv2.drawContours(filled_depth, contours, -1, (255), thickness=cv2.FILLED)
 
 #     filled_depth = cv2.resize(filled_depth, (W, H), interpolation=cv2.INTER_AREA)
+#     filled_depth = _pad_or_crop_to_width_np(filled_depth, 1024, pad_value=0)
 
-#     height, width = filled_depth.shape
-#     total_padding = 1024 - width
-#     left_padding = total_padding // 2
-#     right_padding = total_padding - left_padding
-
-#     padded_depth = cv2.copyMakeBorder(
-#         filled_depth, 0, 0, left_padding, right_padding,
-#         borderType=cv2.BORDER_CONSTANT,
-#         value=0,
-#     )
-
-#     inverted_image = ImageOps.invert(Image.fromarray(padded_depth))
+#     inverted_image = ImageOps.invert(Image.fromarray(filled_depth))
 
 #     with torch.inference_mode():
 #         image_depth = depth_estimator(inverted_image)["depth"]
@@ -294,7 +300,6 @@
 #         raise RuntimeError("Global H/W not set. Call run_one() first.")
 
 #     depth_img = _edges_from_parsing(parsing_img)
-
 # #     inverted_depth = cv2.bitwise_not(depth_img)
 #     contours, _ = cv2.findContours(depth_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
 
@@ -302,19 +307,9 @@
 #     cv2.drawContours(filled_depth, contours, -1, (255), thickness=cv2.FILLED)
 
 #     filled_depth = cv2.resize(filled_depth, (W, H), interpolation=cv2.INTER_AREA)
+#     filled_depth = _pad_or_crop_to_width_np(filled_depth, 1024, pad_value=0)
 
-#     height, width = filled_depth.shape
-#     total_padding = 1024 - width
-#     left_padding = total_padding // 2
-#     right_padding = total_padding - left_padding
-
-#     padded_depth = cv2.copyMakeBorder(
-#         filled_depth, 0, 0, left_padding, right_padding,
-#         borderType=cv2.BORDER_CONSTANT,
-#         value=0,
-#     )
-
-#     inverted_image = ImageOps.invert(Image.fromarray(padded_depth))
+#     inverted_image = ImageOps.invert(Image.fromarray(filled_depth))
 
 #     with torch.inference_mode():
 #         image_depth = depth_estimator(inverted_image)["depth"]
@@ -402,7 +397,7 @@
 #     return pipe, device, dtype
 
 
-# # ✅ UI 표기 → 내부 extractor category 문자열 매핑
+# # UI 표기 → 내부 extractor category 문자열 매핑
 # _UI_TO_EXTRACTOR_CATEGORY = {
 #     "Upper-body": "Upper-cloth",
 #     "Lower-body": "Bottom",
@@ -411,16 +406,12 @@
 
 
 # def run_one(paths: Paths, prompt: str, steps: int = DEFAULT_STEPS, category: str = "Dress"):
-#     """
-#     category: UI에서 넘어오는 값(Upper-body/Lower-body/Dress)
-#     """
 #     global H, W
 #     pipe, device, _dtype = get_pipe_and_device()
 #     image_encoder_dir, ip_ckpt, schp_ckpt = get_assets()
 
 #     H, W = compute_hw_from_person(paths.person_path)
 
-#     # ✅ UI category를 extractor가 기대하는 문자열로 변환
 #     extractor_category = _UI_TO_EXTRACTOR_CATEGORY.get(category, "Dress")
 
 #     res = run_simple_extractor(
@@ -456,61 +447,34 @@
 #     merged_img = merge_white_regions_or(parsing_img, sketch_area)
 #     mask_pil = preprocess_mask(merged_img)
 
+#     # person
 #     person_bgr = _imread_or_raise(paths.person_path)
 #     person_bgr = cv2.resize(person_bgr, (W, H), interpolation=cv2.INTER_AREA)
-
-#     target_width = 1024
-#     cur_w = person_bgr.shape[1]
-#     if cur_w < target_width:
-#         total = target_width - cur_w
-#         left = total // 2
-#         right = total - left
-#         padded_person = cv2.copyMakeBorder(
-#             person_bgr, 0, 0, left, right,
-#             borderType=cv2.BORDER_CONSTANT, value=[255, 255, 255]
-#         )
-#     elif cur_w > target_width:
-#         left = (cur_w - target_width) // 2
-#         padded_person = person_bgr[:, left:left + target_width]
-#     else:
-#         padded_person = person_bgr
-
-#     person_rgb = cv2.cvtColor(padded_person, cv2.COLOR_BGR2RGB)
+#     person_bgr = _pad_or_crop_to_width_np(person_bgr, 1024, pad_value=[255, 255, 255])
+#     person_rgb = cv2.cvtColor(person_bgr, cv2.COLOR_BGR2RGB)
 #     person_pil = Image.fromarray(person_rgb)
 
+#     # depth
 #     if use_depth_path:
 #         depth_map = make_depth(paths.depth_path)
 #     else:
 #         depth_map = make_depth_from_parsing_edges(parsing_img)
 
+#     # garment image (✅ 여기서부터가 핵심: 1024 폭 강제)
 #     personn = Image.open(paths.person_path).convert("RGB")
-#     garment_ = apply_parsing_white_mask_to_person_cv2(personn, parsing_img)
-#     garment_rgb = cv2.cvtColor(garment_, cv2.COLOR_BGR2RGB)
+#     garment_bgr = apply_parsing_white_mask_to_person_cv2(personn, parsing_img)
+#     garment_rgb = cv2.cvtColor(garment_bgr, cv2.COLOR_BGR2RGB)
 #     garment_rgb = cv2.resize(garment_rgb, (W, H), interpolation=cv2.INTER_AREA)
 
-#     padding = (target_width - W) // 2 if W < target_width else 0
-#     garment_rgb = cv2.copyMakeBorder(
-#         garment_rgb,
-#         top=0, bottom=0,
-#         left=padding, right=padding,
-#         borderType=cv2.BORDER_CONSTANT,
-#         value=[255, 255, 255],
-#     )
+#     # ✅ 기존 padding=(1024-W)//2 방식 제거 → 비대칭 패딩/크롭으로 정확히 1024
+#     garment_rgb = _pad_or_crop_to_width_np(garment_rgb, 1024, pad_value=[255, 255, 255])
 #     garment_pil = Image.fromarray(garment_rgb)
 
+#     # garment mask (✅ 동일하게 1024 맞춤)
 #     gm = np.array(parsing_img.convert("L"), dtype=np.uint8)
 #     gm = cv2.resize(gm, (W, H), interpolation=cv2.INTER_NEAREST)
 #     gm = cv2.cvtColor(gm, cv2.COLOR_GRAY2RGB)
-
-#     cur_w2 = gm.shape[1]
-#     if cur_w2 < target_width:
-#         total = target_width - cur_w2
-#         left = total // 2
-#         right = total - left
-#         gm = cv2.copyMakeBorder(gm, 0, 0, left, right, cv2.BORDER_CONSTANT, value=[0, 0, 0])
-#     elif cur_w2 > target_width:
-#         left2 = (cur_w2 - target_width) // 2
-#         gm = gm[:, left2:left2 + target_width]
+#     gm = _pad_or_crop_to_width_np(gm, 1024, pad_value=[0, 0, 0])
 #     garment_mask_pil = Image.fromarray(gm)
 
 #     print(
@@ -531,7 +495,7 @@
 #         device,
 #         mask_pil,
 #         person_pil,
-#         content_scale=0.2,
+#         content_scale=0.3,
 #         style_scale=0.5,
 #         garment_images=garment_pil,
 #         garment_mask=garment_mask_pil,
@@ -547,9 +511,9 @@
 
 #     style_img = Image.open(paths.style_path).convert("RGB")
     
-#     prompt = category + "with " + category
+#     prompt = extractor_category + "with " + prompt
     
-#     print("====prompt? ", prompt)
+#     print("====prompt?  ", prompt)
 
 #     with torch.inference_mode():
 #         images = ip_model.generate(
@@ -613,7 +577,6 @@
 # with gr.Blocks(title="VISTA Demo (HF Spaces)") as demo:
 #     gr.Markdown("## VISTA Demo\nperson / style 필수, sketch(guide)는 선택입니다.")
 
-#     # ✅ UI 표기는 Upper-body/Lower-body/Dress 유지 (기본 Dress)
 #     category_toggle = gr.Radio(
 #         choices=["Dress", "Upper-body", "Lower-body"],
 #         value="Dress",
@@ -621,7 +584,7 @@
 #         interactive=True,
 #     )
 
-#     # ✅ 한 행에 Person / Style / Output 배치
+#     # 한 행에 Person / Style / Output
 #     with gr.Row():
 #         person_in = gr.Image(label="Person Image (required)", type="filepath")
 #         style_in = gr.Image(label="Style Image (required)", type="filepath")
@@ -636,8 +599,6 @@
 #         seed_in = gr.Number(label="Seed (-1=random)", value=-1, precision=0)
 
 #     run_btn = gr.Button("Run")
-
-#     # 파일 다운로드는 Output 아래(다음 행)에 두는 게 일반적으로 보기 좋음
 #     out_file = gr.File(label="Download result.png")
 
 #     gr.Markdown("### Debug Visualizations (mask/depth/etc)")
@@ -662,6 +623,7 @@
 
 import os
 import sys
+import glob
 
 # ---------------------------------------------------------
 # 0) Make sure local packages (diffusers3, preprocess, etc.) are importable on HF Spaces
@@ -676,7 +638,7 @@ print("[BOOT] sys.path[:5] =", sys.path[:5], flush=True)
 import tempfile
 from dataclasses import dataclass
 from functools import lru_cache
-from typing import Optional, Tuple
+from typing import Optional, Tuple, List
 
 import gradio as gr
 import torch
@@ -739,6 +701,52 @@ def get_assets():
     return image_encoder_dir, ip_ckpt, schp_ckpt
 
 
+# =========================
+# Example assets for Gradio UI
+# =========================
+def _is_image_file(p: str) -> bool:
+    ext = os.path.splitext(p.lower())[1]
+    return ext in (".png", ".jpg", ".jpeg", ".webp")
+
+
+def build_ui_examples(root_dir: str = ROOT) -> List[List[object]]:
+    """
+    Returns a list of [person_fp, sketch_fp_or_None, style_fp] example rows.
+
+    Put your example files under:
+      - {root}/examples/person/*
+      - {root}/examples/style/*
+      - {root}/examples/sketch/*   (optional)
+
+    If folders are missing or empty, this returns [] and the UI will hide Examples.
+    """
+    person_dir = os.path.join(root_dir, "examples", "person")
+    style_dir = os.path.join(root_dir, "examples", "style")
+    sketch_dir = os.path.join(root_dir, "examples", "sketch")
+
+    persons = [p for p in sorted(glob.glob(os.path.join(person_dir, "*"))) if _is_image_file(p)]
+    styles = [p for p in sorted(glob.glob(os.path.join(style_dir, "*"))) if _is_image_file(p)]
+    sketches = [p for p in sorted(glob.glob(os.path.join(sketch_dir, "*"))) if _is_image_file(p)]
+
+    if not persons or not styles:
+        return []
+
+    n_basic = min(len(persons), len(styles))
+    rows: List[List[object]] = []
+
+    # Examples without sketch (sketch is optional)
+    for i in range(n_basic):
+        rows.append([persons[i], None, styles[i]])
+
+    # Examples with sketch (if available)
+    if sketches:
+        n_sk = min(len(persons), len(styles), len(sketches))
+        for i in range(n_sk):
+            rows.append([persons[i], sketches[i], styles[i]])
+
+    return rows
+
+
 DEFAULT_STEPS = 40
 DEBUG_SAVE = False
 
@@ -791,8 +799,6 @@ def _pad_or_crop_to_width_np(arr: np.ndarray, target_width: int, pad_value):
             value=pad_value,
         )
     else:
-        # 3채널일 때 value는 스칼라 or [b,g,r]/[r,g,b] 모두 허용되는데,
-        # 여기선 arr가 RGB/BGR인지 호출자가 정해줌.
         return cv2.copyMakeBorder(
             arr, 0, 0, left, right,
             borderType=cv2.BORDER_CONSTANT,
@@ -962,7 +968,6 @@ def make_depth_from_parsing_edges(parsing_img: Image.Image) -> Image.Image:
         raise RuntimeError("Global H/W not set. Call run_one() first.")
 
     depth_img = _edges_from_parsing(parsing_img)
-#     inverted_depth = cv2.bitwise_not(depth_img)
     contours, _ = cv2.findContours(depth_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
 
     filled_depth = depth_img.copy()
@@ -1128,7 +1133,6 @@ def run_one(paths: Paths, prompt: str, steps: int = DEFAULT_STEPS, category: str
     garment_rgb = cv2.cvtColor(garment_bgr, cv2.COLOR_BGR2RGB)
     garment_rgb = cv2.resize(garment_rgb, (W, H), interpolation=cv2.INTER_AREA)
 
-    # ✅ 기존 padding=(1024-W)//2 방식 제거 → 비대칭 패딩/크롭으로 정확히 1024
     garment_rgb = _pad_or_crop_to_width_np(garment_rgb, 1024, pad_value=[255, 255, 255])
     garment_pil = Image.fromarray(garment_rgb)
 
@@ -1172,6 +1176,10 @@ def run_one(paths: Paths, prompt: str, steps: int = DEFAULT_STEPS, category: str
             pass
 
     style_img = Image.open(paths.style_path).convert("RGB")
+    
+    prompt = extractor_category + "with " + prompt
+    
+    print("==== prompt? ", prompt)
 
     with torch.inference_mode():
         images = ip_model.generate(
@@ -1251,6 +1259,22 @@ with gr.Blocks(title="VISTA Demo (HF Spaces)") as demo:
     with gr.Accordion("Sketch / Guide (optional)", open=False):
         sketch_in = gr.Image(label="Sketch / Guide", type="filepath")
 
+    # ✅ Examples: click to auto-fill Person/Sketch/Style inputs
+    examples_data = build_ui_examples(ROOT)
+    if examples_data:
+        gr.Markdown("### Examples (click to load)")
+        gr.Examples(
+            examples=examples_data,
+            inputs=[person_in, sketch_in, style_in],
+            examples_per_page=8,
+        )
+    else:
+        gr.Markdown(
+            "### Examples\n"
+            "Add example files under `./examples/person/` and `./examples/style/` "
+            "(optional: `./examples/sketch/`) to show clickable examples here."
+        )
+
     with gr.Row():
         prompt_in = gr.Textbox(label="Prompt", value="", lines=2)
         steps_in = gr.Slider(1, 80, value=DEFAULT_STEPS, step=1, label="Steps")