Upload 4 files

README.md

@@ -1,6 +1,19 @@
 ---
-title: Face Swap Demo
+title: Face Swap
+emoji: 🎭
+colorFrom: red
+colorTo: indigo
 sdk: gradio
-python_version: "3.11"
-sdk_version: "6.5.1"
+sdk_version: 4.44.0
+app_file: app.py
+pinned: false
+license: mit
 ---
+
+# Face Swap — InsightFace inswapper_128
+
+Upload or snap a webcam photo of your face, click **Swap**, and it replaces the face in the target image (default: a Denmark fan). Runs on the free CPU tier.
+
+Models:
+- Detection + embedding: `buffalo_l` (auto-downloaded by insightface)
+- Swap: `inswapper_128.onnx` (pulled from `ezioruan/inswapper_128.onnx`)

app.py

@@ -1,82 +1,98 @@
-import gradio as gr
+"""
+Hugging Face Space — Face Swap with InsightFace inswapper_128.
+Gradio app: webcam or upload a source face, swap it onto a target image.
+"""
+
+import os
 import cv2
 import numpy as np
-import mediapipe as mp
-
-mp_fd = mp.solutions.face_detection.FaceDetection(model_selection=1, min_detection_confidence=0.6)
-
-def read_rgb(image):
-    # gradioはnumpy(H,W,3) RGBで来る想定
-    if image is None:
-        return None
-    return image.copy()
-
-def detect_face_bbox(rgb):
-    h, w, _ = rgb.shape
-    bgr = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
-    res = mp_fd.process(cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB))
-    if not res.detections:
-        return None
-    # 最もスコアが高い顔を使う
-    det = sorted(res.detections, key=lambda d: d.score[0], reverse=True)[0]
-    bb = det.location_data.relative_bounding_box
-    x1 = max(int(bb.xmin * w), 0)
-    y1 = max(int(bb.ymin * h), 0)
-    x2 = min(int((bb.xmin + bb.width) * w), w - 1)
-    y2 = min(int((bb.ymin + bb.height) * h), h - 1)
-    if x2 <= x1 or y2 <= y1:
-        return None
-    return (x1, y1, x2, y2)
-
-def naive_paste(source_rgb, target_rgb):
-    """
-    まずは“動く”こと優先の簡易版：
-    targetの顔bboxにsourceの顔bboxをリサイズして貼り付け、簡易フェザーでなじませる。
-    （高品質にするには、この後にFaceMesh + 三角ワープ + seamlessCloneへ進む）
-    """
-    sb = detect_face_bbox(source_rgb)
-    tb = detect_face_bbox(target_rgb)
-    if sb is None:
-        raise gr.Error("source画像から顔を検出できませんでした。正面顔・明るさを調整して再試行してください。")
-    if tb is None:
-        raise gr.Error("target画像から顔を検出できませんでした。正面顔・明るさを調整して再試行してください。")
-
-    sx1, sy1, sx2, sy2 = sb
-    tx1, ty1, tx2, ty2 = tb
-
-    s_face = source_rgb[sy1:sy2, sx1:sx2]
-    t_h, t_w = (ty2 - ty1), (tx2 - tx1)
-    s_face_rs = cv2.resize(s_face, (t_w, t_h), interpolation=cv2.INTER_LINEAR)
-
-    out = target_rgb.copy()
-
-    # マスク（楕円 + ぼかし）で簡易ブレンド
-    mask = np.zeros((t_h, t_w), dtype=np.uint8)
-    cv2.ellipse(mask, (t_w // 2, t_h // 2), (int(t_w * 0.45), int(t_h * 0.50)), 0, 0, 360, 255, -1)
-    mask = cv2.GaussianBlur(mask, (0, 0), sigmaX=max(t_w, t_h) * 0.02)
-
-    roi = out[ty1:ty2, tx1:tx2]
-    # alpha合成
-    alpha = (mask.astype(np.float32) / 255.0)[..., None]
-    blended = (alpha * s_face_rs + (1 - alpha) * roi).astype(np.uint8)
-    out[ty1:ty2, tx1:tx2] = blended
-
-    return out
-
-def run(source_img, target_img):
-    s = read_rgb(source_img)
-    t = read_rgb(target_img)
-    if s is None or t is None:
-        raise gr.Error("source/target の両方の画像をアップロードしてください。")
-    return naive_paste(s, t)
-
-with gr.Blocks() as demo:
-    gr.Markdown("# Face Swap (HF Spaces / Python)\nアップロードした2枚の画像から顔を検出し、target側にsource顔を当て込みます。")
+import gradio as gr
+import insightface
+from insightface.app import FaceAnalysis
+from huggingface_hub import hf_hub_download
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+DEFAULT_TARGET = os.path.join(HERE, "reference.png")
+
+print("[boot] loading FaceAnalysis (buffalo_l)…", flush=True)
+face_app = FaceAnalysis(name="buffalo_l", providers=["CPUExecutionProvider"])
+face_app.prepare(ctx_id=0, det_size=(640, 640))
+
+print("[boot] downloading inswapper_128…", flush=True)
+MODEL_PATH = hf_hub_download(
+    repo_id="ezioruan/inswapper_128.onnx",
+    filename="inswapper_128.onnx",
+)
+
+print("[boot] loading inswapper_128…", flush=True)
+swapper = insightface.model_zoo.get_model(
+    MODEL_PATH, providers=["CPUExecutionProvider"]
+)
+
+print("[boot] ready", flush=True)
+
+
+def largest_face(faces):
+    return max(
+        faces,
+        key=lambda f: (f.bbox[2] - f.bbox[0]) * (f.bbox[3] - f.bbox[1]),
+    )
+
+
+def swap(source_img, target_img):
+    if source_img is None:
+        return None, "Please provide a source face (webcam or upload)."
+
+    # Gradio gives RGB numpy — convert to BGR for OpenCV/InsightFace
+    source_bgr = cv2.cvtColor(source_img, cv2.COLOR_RGB2BGR)
+
+    if target_img is None:
+        target_bgr = cv2.imread(DEFAULT_TARGET)
+        if target_bgr is None:
+            return None, "Default target image is missing."
+    else:
+        target_bgr = cv2.cvtColor(target_img, cv2.COLOR_RGB2BGR)
+
+    source_faces = face_app.get(source_bgr)
+    if not source_faces:
+        return None, "No face detected in source image."
+    src_face = largest_face(source_faces)
+
+    target_faces = face_app.get(target_bgr)
+    if not target_faces:
+        return None, "No face detected in target image."
+    tgt_face = largest_face(target_faces)
+
+    result = swapper.get(target_bgr, tgt_face, src_face, paste_back=True)
+    return cv2.cvtColor(result, cv2.COLOR_BGR2RGB), "Done."
+
+
+with gr.Blocks(title="Face Swap — InsightFace") as demo:
+    gr.Markdown(
+        "# Face Swap — InsightFace `inswapper_128`\n"
+        "Take a webcam photo (or upload), click **Swap**. "
+        "Leave the target empty to use the default Denmark fan, "
+        "or upload your own target image."
+    )
     with gr.Row():
-        src = gr.Image(label="Source（当てたい顔）", type="numpy")
-        tgt = gr.Image(label="Target（当てられる側）", type="numpy")
-    btn = gr.Button("Swap")
-    out = gr.Image(label="Output", type="numpy")
-    btn.click(run, inputs=[src, tgt], outputs=out)
+        with gr.Column():
+            src_in = gr.Image(
+                sources=["webcam", "upload"],
+                type="numpy",
+                label="Your face (webcam or upload)",
+            )
+            tgt_in = gr.Image(
+                sources=["upload"],
+                type="numpy",
+                label="Target image (optional — defaults to reference.png)",
+                value=DEFAULT_TARGET,
+            )
+            btn = gr.Button("Swap", variant="primary")
+        with gr.Column():
+            out_img = gr.Image(label="Result", type="numpy")
+            status = gr.Textbox(label="Status", interactive=False)
+
+    btn.click(swap, inputs=[src_in, tgt_in], outputs=[out_img, status])
 
-demo.launch()
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -1,5 +1,7 @@
-gradio==6.5.1
-opencv-python
-numpy==1.26.4
-protobuf<5
-mediapipe==0.10.14
+gradio==4.44.0
+insightface==0.7.3
+onnxruntime==1.18.1
+opencv-python-headless==4.10.0.84
+numpy<2
+huggingface_hub>=0.24.0
+cython