print scene-detection heads-up line before the silent CPU pass

Every-frame prediction is free on CUDA (~0.2s) and hopeless on the iGPU
(5.2s/img): a 100-frame faceless scene cost 8.7 min. stride>1 keeps
frame-perfect decode and probes every Nth frame, max M per scene, then
waits for the next scene start. Default stride=1 = original behavior.
Scheduler unit-tested (face-hit, faceless-exhaust, multi-scene).

.gitignore

@@ -6,3 +6,7 @@ output/
 preview/
 *.pt
 *.part
+.venv-ov/
+*.png
+!models/*.png
+kernel.errors.txt

anime_facesaver.py

@@ -119,6 +119,22 @@ def parse_arguments():
              "edges. 0 (default) allows cropped faces; set >0 to reject "
              "faces touching the crop lines.",
     )
+    parser.add_argument(
+        "--probe-stride",
+        type=int,
+        default=1,
+        help="Within a scene still needing a face, run detection every Nth "
+             "frame instead of every frame (1 = original behavior, free on "
+             "CUDA). For slow devices (iGPU ~5s/img) use e.g. 25: decode "
+             "stays frame-perfect, only the expensive inference is thinned.",
+    )
+    parser.add_argument(
+        "--probe-max-per-scene",
+        type=int,
+        default=4,
+        help="With --probe-stride>1, probes before giving up on a scene and "
+             "waiting for the next scene start (default 4).",
+    )
     parser.add_argument(
         "--scene-threshold",
         type=float,
@@ -161,6 +177,10 @@ def resolve_gpu(gpu_arg):
     if gpu_arg.lower() == "cpu":
         os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
         return "cpu"
+    if gpu_arg.lower() == "igpu":
+        # Intel iGPU via OpenVINO (Level Zero) — no CUDA, no torch pinning.
+        # OVFaceDetector takes over model loading in main().
+        return "igpu"
     if gpu_arg.lower() == "auto":
         return None  # let ultralytics/torch decide
     os.environ["CUDA_VISIBLE_DEVICES"] = gpu_arg
@@ -232,6 +252,61 @@ def face_qualifies(box, frame_shape, min_width, min_height,
     return True
 
 
+class AvFrameReader:
+    """Forward-only frame reader with a cv2.VideoCapture-compatible
+    surface (read/release). PyAV ships its own ffmpeg build with libdav1d:
+    the ffmpeg bundled inside opencv-python has no software AV1 decoder,
+    and Alder Lake-N has no AV1 hw decode, so BD AV1 rips (e.g. SavI0r)
+    decode 0 frames through cv2. scenedetect likewise prefers its pyav
+    backend once av is installed, fixing scene detection on the same files.
+    """
+
+    def __init__(self, path):
+        import av
+        self._c = av.open(path)
+        self._vs = self._c.streams.video[0]
+        self._vs.thread_type = "AUTO"
+        self._iter = self._c.decode(self._vs)
+
+    @property
+    def frame_count(self):
+        vs = self._vs
+        if vs.frames:
+            return int(vs.frames)
+        # MKV headers often leave stream.frames unset; estimate from the
+        # container duration (it is this file's total decode work anyway).
+        dur = self._c.duration  # microseconds, or None
+        if dur and vs.average_rate:
+            return int(float(dur) / 1e6 * float(vs.average_rate))
+        return 0
+
+    def read(self):
+        try:
+            frame = next(self._iter)
+        except StopIteration:
+            return False, None
+        return True, frame.to_ndarray(format="bgr24")
+
+    def release(self):
+        try:
+            self._c.close()
+        except Exception:
+            pass
+
+
+def open_reader(video_path):
+    """PyAV when available (AV1-capable), else plain cv2."""
+    import cv2
+    try:
+        return AvFrameReader(video_path)
+    except ImportError:
+        pass
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        return None
+    return cap
+
+
 def process_video(video_path, output_dir, args, model, device, draw,
                   output_start=0):
     from scenedetect import open_video, SceneManager, ContentDetector
@@ -240,6 +315,12 @@ def process_video(video_path, output_dir, args, model, device, draw,
 
     try:
         import cv2
+        try:
+            import av  # noqa: F401
+            # scenedetect defaults to backend="opencv", which cannot decode
+            # AV1 in this env; pyav ships its own ffmpeg with libdav1d.
+            video = open_video(video_path, backend="pyav")
+        except ImportError:
             video = open_video(video_path)
         scene_manager = SceneManager()
         scene_manager.add_detector(
@@ -249,14 +330,18 @@ def process_video(video_path, output_dir, args, model, device, draw,
               f"{video_path}: {e}")
         return output_start
 
-    cap = cv2.VideoCapture(video_path)
-    if not cap.isOpened():
+    cap = open_reader(video_path)
+    if cap is None:
         print(f"Error opening video file {video_path}")
         return output_start
 
+    frame_count = getattr(cap, "frame_count", 0)
+    if not frame_count and hasattr(cap, "get"):
         frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
 
     try:
+        print(f"{video_name}: detecting scenes "
+              "(silent for several minutes on CPU decode)...", flush=True)
         scene_manager.detect_scenes(video=video)
         scene_list = scene_manager.get_scene_list()
         scene_starts = {scene[0].get_frames() for scene in scene_list}
@@ -272,6 +357,17 @@ def process_video(video_path, output_dir, args, model, device, draw,
     face_in_scene = False
     frame_idx = 0
 
+    # Probe schedule. probe_stride=1 reproduces the original every-frame
+    # loop (right on CUDA, where predict costs ~0.2 s). On slow devices
+    # (iGPU ~5 s/img) predicting every frame is hopeless: decode is ~300 fps,
+    # inference 0.2/s. With stride>1 we decode through frames for free and
+    # predict only at scene starts, retrying every stride frames up to
+    # probe_max attempts per scene, then waiting for the next scene start.
+    probe_stride = getattr(args, "probe_stride", 1)
+    probe_max = getattr(args, "probe_max_per_scene", 4)
+    next_probe = 0
+    probes_used = 0
+
     while True:
         ret, frame = cap.read()
         if not ret:
@@ -279,10 +375,20 @@ def process_video(video_path, output_dir, args, model, device, draw,
 
         if frame_idx in scene_starts:
             face_in_scene = False
-        # membership test on a set; pop matched starts to keep it cheap
-        # (scenes strictly increase, so set stays small either way)
+            next_probe = frame_idx
+            probes_used = 0
+
+        if face_in_scene:
+            frame_idx += 1
+            continue
+        if probe_stride > 1 and (frame_idx < next_probe
+                                 or probes_used >= probe_max):
+            frame_idx += 1
+            continue
 
         if not face_in_scene:
+            probes_used += 1
+            next_probe = frame_idx + probe_stride
             try:
                 results = model.predict(frame, conf=args.conf,
                                         iou=args.iou, imgsz=args.imgsz,
@@ -353,6 +459,22 @@ def main():
     from ultralytics import YOLO
     import cv2
 
+    if device == "igpu":
+        # Intel iGPU path: OpenVINO replaces ultralytics' torch predict loop.
+        # Ultralytics import stays (shared pre/post code paths), but weights
+        # load as ONNX through OVFaceDetector instead of torch.
+        from ov_detector import OVFaceDetector
+        onnx_path = os.path.splitext(args.model)[0] + ".onnx"
+        if not os.path.exists(onnx_path):
+            print(f"Error: --gpu igpu needs an exported ONNX at {onnx_path}\n"
+                  "  (one-time: YOLO('<pt>').export(format='onnx', "
+                  "dynamic=True, simplify=True, opset=17))")
+            sys.exit(1)
+        model = OVFaceDetector(onnx_path, device="GPU", imgsz=args.imgsz)
+        names = model.names
+        print("Device: igpu (OpenVINO, Intel iGPU via Level Zero)")
+        print(f"Model loaded (OpenVINO GPU). Classes: {names}")
+    else:
         if device == "cuda:0" and not torch.cuda.is_available():
             print("CUDA requested but not visible after pinning "
                   f"(--gpu {args.gpu}). Falling back to CPU.")

ov_detector.py

@@ -0,0 +1,136 @@
+"""OpenVINO-backed anime-face detector for anime_facesaver (iGPU path).
+
+Used when --gpu igpu: replaces ultralytics' torch predict loop with a direct
+OpenVINO inference request on Intel GPUs (Level Zero / ANV). Pre/post-processing
+emulates ultralytics YOLO defaults: 1280 letterbox (stride 32), conf/iou gates,
+class-agnostic-free single-class NMS.
+
+On chibi (i3-N305): GPU fp32 measured 5.2 s/img vs CPU 14.5 s/img
+(2026-10-20, models/yolov8x6_animeface.onnx, 25 real stills).
+"""
+import os
+
+import numpy as np
+
+
+def letterbox(img, new_shape=1280, stride=32):
+    """Square ultralytics letterbox (auto=False): fit, then center-pad to
+    new_shape x new_shape. The GPU plugin compiles one static shape this way
+    and sidesteps dynamic-shape concat failures on non-square inputs
+    (measured: (1,3,736,1280) fails inside the graph; 1280x1280 always works).
+
+    Returns (blob CHW float32 /255, scale, (pad_left, pad_top)).
+    Inverse map for boxes: x_orig = (x_lb - pad_left) / scale.
+    """
+    import cv2
+
+    h, w = img.shape[:2]
+    r = min(new_shape / h, new_shape / w)
+    nw, nh = int(round(w * r)), int(round(h * r))
+    interp = cv2.INTER_AREA if r < 1 else cv2.INTER_LINEAR
+    resized = cv2.resize(img, (nw, nh), interpolation=interp)
+    canvas = np.full((new_shape, new_shape, 3), 114, dtype=np.uint8)
+    top, left = (new_shape - nh) // 2, (new_shape - nw) // 2
+    canvas[top:top + nh, left:left + nw] = resized
+    blob = canvas[..., ::-1].transpose(2, 0, 1)[None].astype(np.float32) / 255.0
+    return blob, r, (left, top)
+
+
+def numpy_ceil_div(a, b):
+    return -(-a // b)
+
+
+def nms_numpy(boxes, scores, iou_thres):
+    """Plain greedy NMS; returns kept indices."""
+    if len(boxes) == 0:
+        return []
+    x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
+    areas = (x2 - x1) * (y2 - y1)
+    order = scores.argsort()[::-1]
+    keep = []
+    while order.size:
+        i = order[0]
+        keep.append(int(i))
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+        inter = np.clip(xx2 - xx1, 0, None) * np.clip(yy2 - yy1, 0, None)
+        iou = inter / (areas[i] + areas[order[1:]] - inter + 1e-9)
+        order = order[1:][iou <= iou_thres]
+    return keep
+
+
+class OVFaceDetector:
+    """model.predict(frame_bgr, conf=, iou=, imgsz=, device=) compatible shim.
+
+    Returns a one-element list with .boxes being a Boxes-like object exposing
+    .xyxy (numpy, already denormalized to original-frame pixels) and .conf.
+    """
+
+    class _Boxes:
+        """numpy-backed stand-in for ultralytics Boxes (.xyxy.cpu().numpy())."""
+        class _Lazy:
+            def __init__(self, arr):
+                self._arr = arr
+
+            def cpu(self):
+                return self
+
+            def numpy(self):
+                return self._arr
+
+            def int(self):
+                return OVFaceDetector._Boxes._Lazy(self._arr.astype(np.int32))
+
+            def tolist(self):
+                return self._arr.tolist()
+
+        def __init__(self, xyxy, conf):
+            self.xyxy = self._Lazy(xyxy)
+            self.conf = conf
+
+    class _Result:
+        def __init__(self, boxes):
+            self.boxes = boxes
+
+    def __init__(self, model_path, device="GPU", imgsz=1280):
+        import openvino as ov
+
+        core = ov.Core()
+        if device != "CPU" and device not in core.available_devices:
+            raise RuntimeError(
+                f"OpenVINO device {device!r} not available "
+                f"(have {core.available_devices}). Is the render-group fix "
+                f"active (sg render / re-login) and intel-level-zero-gpu "
+                f"installed?")
+        self._ov = ov
+        self._req = core.compile_model(model_path, device).create_infer_request()
+        self.imgsz = imgsz
+        self.names = {0: "face"}
+
+    def predict(self, frame, conf=0.25, iou=0.45, imgsz=None, device=None,
+                verbose=False):
+        imgsz = imgsz or self.imgsz
+        blob, scale, (pad_l, pad_t) = letterbox(frame, imgsz)
+        out = self._req.infer([self._ov.Tensor(blob)])
+        pred = list(out.values())[0]  # [1, 5, N]: cx cy w h conf cls
+        pred = pred[0].T  # [N, 5]
+        scores = pred[:, 4]
+        cand = pred[scores >= conf]
+        if len(cand) == 0:
+            return [self._Result(self._Boxes(
+                np.zeros((0, 4), np.float32), np.zeros((0,), np.float32)))]
+        cx, cy, w, h = cand[:, 0], cand[:, 1], cand[:, 2], cand[:, 3]
+        boxes = np.stack([cx - w / 2, cy - h / 2, cx + w / 2, cy + h / 2], 1)
+        keep = nms_numpy(boxes, cand[:, 4], iou)
+        boxes, scores = boxes[keep], cand[keep, 4]
+        # map letterbox coords back to original frame pixels
+        boxes[:, [0, 2]] -= pad_l
+        boxes[:, [1, 3]] -= pad_t
+        boxes = boxes / scale
+        h0, w0 = frame.shape[:2]
+        boxes[:, [0, 2]] = boxes[:, [0, 2]].clip(0, w0)
+        boxes[:, [1, 3]] = boxes[:, [1, 3]].clip(0, h0)
+        return [self._Result(self._Boxes(boxes.astype(np.float32),
+                                         scores.astype(np.float32)))]

validate_ov.py

@@ -0,0 +1,41 @@
+"""Cross-validate OVFaceDetector vs ultralytics CPU baseline on real stills."""
+import glob, sys, time
+import numpy as np, cv2
+
+imgs = sorted(glob.glob('/home/alex/Documents/ov-bench/imgs/*.png'))
+sys.path.insert(0, '.')
+from ov_detector import OVFaceDetector
+
+# baseline: ultralytics torch CPU (deterministic on CPU)
+from ultralytics import YOLO
+base = YOLO('models/yolov8x6_animeface.pt')
+
+ovd = OVFaceDetector('models/yolov8x6_animeface.onnx', device='GPU', imgsz=1280)
+
+def iou(a, b):
+    xx1, yy1 = max(a[0], b[0]), max(a[1], b[1])
+    xx2, yy2 = min(a[2], b[2]), min(a[3], b[3])
+    inter = max(0, xx2-xx1) * max(0, yy2-yy1)
+    ua = (a[2]-a[0])*(a[3]-a[1]) + (b[2]-b[0])*(b[3]-b[1]) - inter
+    return inter / max(ua, 1e-9)
+
+ious, n_match, n_only_base, n_only_ov = [], 0, 0, 0
+t0=time.time()
+for p in imgs:
+    fr = cv2.imread(p)
+    rb = base.predict(fr, conf=0.25, imgsz=1280, device='cpu', verbose=False)[0]
+    ro = ovd.predict(fr, conf=0.25, imgsz=1280, verbose=False)[0]
+    bb = rb.boxes.xyxy.cpu().numpy() if rb.boxes is not None else np.zeros((0,4))
+    bo = ro.boxes.xyxy.numpy()
+    used = set()
+    for box in bb:
+        best, bi = 0, -1
+        for j, ob in enumerate(bo):
+            v = iou(box, ob)
+            if v > best: best, bi = v, j
+        if best > 0.5:
+            ious.append(best); n_match += 1; used.add(bi)
+        else: n_only_base += 1
+    n_only_ov += len(bo) - len(used)
+dt = time.time()-t0
+print(f"imgs={len(imgs)} matched={n_match} (mean IoU {np.mean(ious):.3f}) only_base={n_only_base} only_ov={n_only_ov} [{dt:.0f}s total]")