Skip to content

AI Transforms

AI-powered video transforms that use face detection for intelligent cropping and tracking. Framing-oriented behavior (headroom / thirds / speed clamp) is implemented on FaceTrackingCrop.

Usage

from videopython.ai import FaceTrackingCrop, SplitScreenComposite
from videopython.base import Video

video = Video.from_path("input.mp4")
video2 = Video.from_path("input_2.mp4")

# Create vertical content from horizontal by tracking faces
crop = FaceTrackingCrop(target_aspect=(9, 16))
vertical_video = crop.apply(video)

# Face-tracking crop with headroom framing and limited camera speed
framing = FaceTrackingCrop(framing_rule="headroom", max_speed=0.1)
framed_video = framing.apply(video)

# Create split-screen with face tracking
composite = SplitScreenComposite(layout="2x1")
split_video = composite.apply(video, video2)

FaceTracker

FaceTracker

Utility for tracking faces across video frames with smoothing.

Provides frame-by-frame face detection with position smoothing using exponential moving average to prevent jitter in the tracked position.

Supports GPU acceleration via YOLOv8-face model for significantly faster detection, with optional frame sampling and interpolation for video.

Example

CPU tracking (default, backward compatible)

tracker = FaceTracker() for i, frame in enumerate(frames): ... pos = tracker.detect_and_track(frame, i)

GPU tracking with frame sampling

tracker = FaceTracker(backend="gpu", sample_rate=5) positions = tracker.track_video(frames)

Source code in src/videopython/ai/transforms.py 
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
class FaceTracker:
    """Utility for tracking faces across video frames with smoothing.

    Provides frame-by-frame face detection with position smoothing using
    exponential moving average to prevent jitter in the tracked position.

    Supports GPU acceleration via YOLOv8-face model for significantly faster
    detection, with optional frame sampling and interpolation for video.

    Example:
        >>> # CPU tracking (default, backward compatible)
        >>> tracker = FaceTracker()
        >>> for i, frame in enumerate(frames):
        ...     pos = tracker.detect_and_track(frame, i)
        >>>
        >>> # GPU tracking with frame sampling
        >>> tracker = FaceTracker(backend="gpu", sample_rate=5)
        >>> positions = tracker.track_video(frames)
    """

    def __init__(
        self,
        selection_strategy: Literal["largest", "centered", "index"] = "largest",
        face_index: int = 0,
        smoothing: float = 0.8,
        detection_interval: int = 3,
        min_face_size: int = 30,
        backend: Literal["cpu", "gpu", "auto"] = "cpu",
        sample_rate: int = 1,
        batch_size: int = 16,
    ):
        """Initialize face tracker.

        Args:
            selection_strategy: How to select which face to track.
                - "largest": Track the face with the largest bounding box.
                - "centered": Track the face closest to frame center.
                - "index": Track the face at a specific index (sorted by area).
            face_index: Index of face to track when using "index" strategy.
            smoothing: Exponential moving average factor (0-1). Higher = smoother.
            detection_interval: Run detection every N frames, interpolate between.
            min_face_size: Minimum face size in pixels for detection.
            backend: Detection backend - "cpu", "gpu", or "auto".
            sample_rate: For GPU backend, detect every Nth frame and interpolate.
                Only used by track_video(). Default 1 (every frame).
            batch_size: Batch size for GPU detection. Default 16.
        """
        self.selection_strategy = selection_strategy
        self.face_index = face_index
        self.smoothing = smoothing
        self.detection_interval = detection_interval
        self.min_face_size = min_face_size
        self.backend: Literal["cpu", "gpu", "auto"] = backend
        self.sample_rate = sample_rate
        self.batch_size = batch_size

        self._detector: FaceDetector | None = None
        self._last_position: tuple[float, float] | None = None
        self._last_size: tuple[float, float] | None = None
        self._smoothed_position: tuple[float, float] | None = None
        self._smoothed_size: tuple[float, float] | None = None

    def _init_detector(self) -> None:
        """Initialize face detector lazily."""
        self._detector = FaceDetector(
            min_face_size=self.min_face_size,
            backend=self.backend,
        )

    def _select_face(
        self,
        faces: list,
        frame_width: int,
        frame_height: int,
    ) -> tuple[float, float, float, float] | None:
        """Select a face based on the configured strategy.

        Args:
            faces: List of DetectedFace objects.
            frame_width: Width of the frame.
            frame_height: Height of the frame.

        Returns:
            Tuple of (center_x, center_y, width, height) in normalized coords, or None.
        """
        if not faces:
            return None

        if self.selection_strategy == "largest":
            # Faces are already sorted by area (largest first)
            face = faces[0]
        elif self.selection_strategy == "centered":
            # Find face closest to center
            frame_center = (0.5, 0.5)
            face = min(
                faces,
                key=lambda f: (
                    (f.bounding_box.center[0] - frame_center[0]) ** 2
                    + (f.bounding_box.center[1] - frame_center[1]) ** 2
                ),
            )
        elif self.selection_strategy == "index":
            if self.face_index < len(faces):
                face = faces[self.face_index]
            else:
                face = faces[0]  # Fall back to largest
        else:
            face = faces[0]

        bbox = face.bounding_box
        return (bbox.center[0], bbox.center[1], bbox.width, bbox.height)

    def detect_and_track(
        self,
        frame: np.ndarray,
        frame_index: int,
    ) -> tuple[float, float, float, float] | None:
        """Detect face in frame and return smoothed position.

        Args:
            frame: Video frame as numpy array (H, W, 3).
            frame_index: Index of current frame.

        Returns:
            Tuple of (center_x, center_y, width, height) in normalized coords,
            or None if no face detected and no fallback available.
        """
        if self._detector is None:
            self._init_detector()
            assert self._detector is not None

        h, w = frame.shape[:2]

        # Only run detection on interval frames
        should_detect = frame_index % self.detection_interval == 0

        if should_detect:
            faces = self._detector.detect(frame)
            face_info = self._select_face(faces, w, h)

            if face_info:
                cx, cy, fw, fh = face_info
                self._last_position = (cx, cy)
                self._last_size = (fw, fh)
        else:
            # Use last detected position
            face_info = None
            if self._last_position and self._last_size:
                face_info = (*self._last_position, *self._last_size)

        if face_info:
            cx, cy, fw, fh = face_info

            # Apply exponential moving average smoothing
            if self._smoothed_position is None:
                self._smoothed_position = (cx, cy)
                self._smoothed_size = (fw, fh)
            else:
                alpha = 1 - self.smoothing
                self._smoothed_position = (
                    self._smoothed_position[0] * self.smoothing + cx * alpha,
                    self._smoothed_position[1] * self.smoothing + cy * alpha,
                )
                assert self._smoothed_size is not None  # Set alongside _smoothed_position
                self._smoothed_size = (
                    self._smoothed_size[0] * self.smoothing + fw * alpha,
                    self._smoothed_size[1] * self.smoothing + fh * alpha,
                )

            return (*self._smoothed_position, *self._smoothed_size)

        # Return last smoothed position as fallback
        if self._smoothed_position and self._smoothed_size:
            return (*self._smoothed_position, *self._smoothed_size)

        return None

    def reset(self) -> None:
        """Reset tracker state for a new video."""
        self._last_position = None
        self._last_size = None
        self._smoothed_position = None
        self._smoothed_size = None

    @staticmethod
    def _interpolate_bbox(
        bbox1: tuple[float, float, float, float],
        bbox2: tuple[float, float, float, float],
        t: float,
    ) -> tuple[float, float, float, float]:
        """Linearly interpolate between two bounding boxes.

        Args:
            bbox1: First bounding box (cx, cy, w, h).
            bbox2: Second bounding box (cx, cy, w, h).
            t: Interpolation factor (0 = bbox1, 1 = bbox2).

        Returns:
            Interpolated bounding box (cx, cy, w, h).
        """
        return (
            bbox1[0] + (bbox2[0] - bbox1[0]) * t,
            bbox1[1] + (bbox2[1] - bbox1[1]) * t,
            bbox1[2] + (bbox2[2] - bbox1[2]) * t,
            bbox1[3] + (bbox2[3] - bbox1[3]) * t,
        )

    def track_video(
        self,
        frames: np.ndarray,
    ) -> list[tuple[float, float, float, float] | None]:
        """Track face through entire video using optimized batch detection.

        This method is optimized for GPU backends with frame sampling and
        interpolation for smooth tracking with reduced computation.

        Args:
            frames: Video frames array of shape (N, H, W, 3).

        Returns:
            List of face positions (cx, cy, w, h) for each frame, or None if
            no face detected and no fallback available.
        """
        if self._detector is None:
            self._init_detector()
            assert self._detector is not None

        n_frames = len(frames)
        if n_frames == 0:
            return []

        h, w = frames[0].shape[:2]

        # Determine which frames to sample
        if self.sample_rate > 1 and self.backend in ("gpu", "auto"):
            sample_indices = list(range(0, n_frames, self.sample_rate))
            # Ensure last frame is included
            if sample_indices[-1] != n_frames - 1:
                sample_indices.append(n_frames - 1)
        else:
            sample_indices = list(range(n_frames))

        # Batch detect on sampled frames
        sampled_frames = [frames[i] for i in sample_indices]

        # Process in batches
        sampled_detections: list[list] = []
        for batch_start in range(0, len(sampled_frames), self.batch_size):
            batch_end = min(batch_start + self.batch_size, len(sampled_frames))
            batch = sampled_frames[batch_start:batch_end]
            batch_results = self._detector.detect_batch(batch)
            sampled_detections.extend(batch_results)

        # Extract face info from detections
        sampled_faces: list[tuple[float, float, float, float] | None] = []
        for faces in sampled_detections:
            face_info = self._select_face(faces, w, h)
            sampled_faces.append(face_info)

        # If no sampling, apply smoothing directly
        if self.sample_rate == 1 or self.backend == "cpu":
            self.reset()
            results: list[tuple[float, float, float, float] | None] = []
            for i, face_info in enumerate(sampled_faces):
                if face_info:
                    cx, cy, fw, fh = face_info
                    self._last_position = (cx, cy)
                    self._last_size = (fw, fh)

                    if self._smoothed_position is None:
                        self._smoothed_position = (cx, cy)
                        self._smoothed_size = (fw, fh)
                    else:
                        alpha = 1 - self.smoothing
                        self._smoothed_position = (
                            self._smoothed_position[0] * self.smoothing + cx * alpha,
                            self._smoothed_position[1] * self.smoothing + cy * alpha,
                        )
                        assert self._smoothed_size is not None
                        self._smoothed_size = (
                            self._smoothed_size[0] * self.smoothing + fw * alpha,
                            self._smoothed_size[1] * self.smoothing + fh * alpha,
                        )

                    results.append((*self._smoothed_position, *self._smoothed_size))
                elif self._smoothed_position and self._smoothed_size:
                    results.append((*self._smoothed_position, *self._smoothed_size))
                else:
                    results.append(None)
            return results

        # Interpolate between sampled frames
        all_positions: list[tuple[float, float, float, float] | None] = [None] * n_frames

        # Fill in sampled positions
        for idx, sample_idx in enumerate(sample_indices):
            all_positions[sample_idx] = sampled_faces[idx]

        # Interpolate gaps
        for i in range(len(sample_indices) - 1):
            start_idx = sample_indices[i]
            end_idx = sample_indices[i + 1]
            start_face = sampled_faces[i]
            end_face = sampled_faces[i + 1]

            if start_face is None and end_face is None:
                continue
            elif start_face is None:
                # Use end face for all
                for j in range(start_idx, end_idx):
                    all_positions[j] = end_face
            elif end_face is None:
                # Use start face for all
                for j in range(start_idx + 1, end_idx + 1):
                    all_positions[j] = start_face
            else:
                # Interpolate
                gap = end_idx - start_idx
                for j in range(start_idx + 1, end_idx):
                    t = (j - start_idx) / gap
                    all_positions[j] = self._interpolate_bbox(start_face, end_face, t)

        # Apply smoothing to interpolated positions
        self.reset()
        results = []
        for face_info in all_positions:
            if face_info:
                cx, cy, fw, fh = face_info

                if self._smoothed_position is None:
                    self._smoothed_position = (cx, cy)
                    self._smoothed_size = (fw, fh)
                else:
                    alpha = 1 - self.smoothing
                    self._smoothed_position = (
                        self._smoothed_position[0] * self.smoothing + cx * alpha,
                        self._smoothed_position[1] * self.smoothing + cy * alpha,
                    )
                    assert self._smoothed_size is not None
                    self._smoothed_size = (
                        self._smoothed_size[0] * self.smoothing + fw * alpha,
                        self._smoothed_size[1] * self.smoothing + fh * alpha,
                    )

                results.append((*self._smoothed_position, *self._smoothed_size))
            elif self._smoothed_position and self._smoothed_size:
                results.append((*self._smoothed_position, *self._smoothed_size))
            else:
                results.append(None)

        return results

__init__ 

__init__(
    selection_strategy: Literal[
        "largest", "centered", "index"
    ] = "largest",
    face_index: int = 0,
    smoothing: float = 0.8,
    detection_interval: int = 3,
    min_face_size: int = 30,
    backend: Literal["cpu", "gpu", "auto"] = "cpu",
    sample_rate: int = 1,
    batch_size: int = 16,
)

Initialize face tracker.

Parameters:

Name Type Description Default
selection_strategy Literal['largest', 'centered', 'index']

How to select which face to track. - "largest": Track the face with the largest bounding box. - "centered": Track the face closest to frame center. - "index": Track the face at a specific index (sorted by area).

 'largest'
face_index int

Index of face to track when using "index" strategy.

 0
smoothing float

Exponential moving average factor (0-1). Higher = smoother.

 0.8
detection_interval int

Run detection every N frames, interpolate between.

 3
min_face_size int

Minimum face size in pixels for detection.

 30
backend Literal['cpu', 'gpu', 'auto']

Detection backend - "cpu", "gpu", or "auto".

 'cpu'
sample_rate int

For GPU backend, detect every Nth frame and interpolate. Only used by track_video(). Default 1 (every frame).

 1
batch_size int

Batch size for GPU detection. Default 16.

 16
Source code in src/videopython/ai/transforms.py 
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
def __init__(
    self,
    selection_strategy: Literal["largest", "centered", "index"] = "largest",
    face_index: int = 0,
    smoothing: float = 0.8,
    detection_interval: int = 3,
    min_face_size: int = 30,
    backend: Literal["cpu", "gpu", "auto"] = "cpu",
    sample_rate: int = 1,
    batch_size: int = 16,
):
    """Initialize face tracker.

    Args:
        selection_strategy: How to select which face to track.
            - "largest": Track the face with the largest bounding box.
            - "centered": Track the face closest to frame center.
            - "index": Track the face at a specific index (sorted by area).
        face_index: Index of face to track when using "index" strategy.
        smoothing: Exponential moving average factor (0-1). Higher = smoother.
        detection_interval: Run detection every N frames, interpolate between.
        min_face_size: Minimum face size in pixels for detection.
        backend: Detection backend - "cpu", "gpu", or "auto".
        sample_rate: For GPU backend, detect every Nth frame and interpolate.
            Only used by track_video(). Default 1 (every frame).
        batch_size: Batch size for GPU detection. Default 16.
    """
    self.selection_strategy = selection_strategy
    self.face_index = face_index
    self.smoothing = smoothing
    self.detection_interval = detection_interval
    self.min_face_size = min_face_size
    self.backend: Literal["cpu", "gpu", "auto"] = backend
    self.sample_rate = sample_rate
    self.batch_size = batch_size

    self._detector: FaceDetector | None = None
    self._last_position: tuple[float, float] | None = None
    self._last_size: tuple[float, float] | None = None
    self._smoothed_position: tuple[float, float] | None = None
    self._smoothed_size: tuple[float, float] | None = None

detect_and_track 

detect_and_track(
    frame: ndarray, frame_index: int
) -> tuple[float, float, float, float] | None

Detect face in frame and return smoothed position.

Parameters:

Name Type Description Default
frame ndarray

Video frame as numpy array (H, W, 3).

 required
frame_index int

Index of current frame.

 required

Returns:

Type Description
tuple[float, float, float, float] | None

Tuple of (center_x, center_y, width, height) in normalized coords,
tuple[float, float, float, float] | None

or None if no face detected and no fallback available.
Source code in src/videopython/ai/transforms.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
def detect_and_track(
    self,
    frame: np.ndarray,
    frame_index: int,
) -> tuple[float, float, float, float] | None:
    """Detect face in frame and return smoothed position.

    Args:
        frame: Video frame as numpy array (H, W, 3).
        frame_index: Index of current frame.

    Returns:
        Tuple of (center_x, center_y, width, height) in normalized coords,
        or None if no face detected and no fallback available.
    """
    if self._detector is None:
        self._init_detector()
        assert self._detector is not None

    h, w = frame.shape[:2]

    # Only run detection on interval frames
    should_detect = frame_index % self.detection_interval == 0

    if should_detect:
        faces = self._detector.detect(frame)
        face_info = self._select_face(faces, w, h)

        if face_info:
            cx, cy, fw, fh = face_info
            self._last_position = (cx, cy)
            self._last_size = (fw, fh)
    else:
        # Use last detected position
        face_info = None
        if self._last_position and self._last_size:
            face_info = (*self._last_position, *self._last_size)

    if face_info:
        cx, cy, fw, fh = face_info

        # Apply exponential moving average smoothing
        if self._smoothed_position is None:
            self._smoothed_position = (cx, cy)
            self._smoothed_size = (fw, fh)
        else:
            alpha = 1 - self.smoothing
            self._smoothed_position = (
                self._smoothed_position[0] * self.smoothing + cx * alpha,
                self._smoothed_position[1] * self.smoothing + cy * alpha,
            )
            assert self._smoothed_size is not None  # Set alongside _smoothed_position
            self._smoothed_size = (
                self._smoothed_size[0] * self.smoothing + fw * alpha,
                self._smoothed_size[1] * self.smoothing + fh * alpha,
            )

        return (*self._smoothed_position, *self._smoothed_size)

    # Return last smoothed position as fallback
    if self._smoothed_position and self._smoothed_size:
        return (*self._smoothed_position, *self._smoothed_size)

    return None

reset 

reset() -> None

Reset tracker state for a new video.

Source code in src/videopython/ai/transforms.py 
208
209
210
211
212
213
def reset(self) -> None:
    """Reset tracker state for a new video."""
    self._last_position = None
    self._last_size = None
    self._smoothed_position = None
    self._smoothed_size = None

track_video 

track_video(
    frames: ndarray,
) -> list[tuple[float, float, float, float] | None]

Track face through entire video using optimized batch detection.

This method is optimized for GPU backends with frame sampling and interpolation for smooth tracking with reduced computation.

Parameters:

Name Type Description Default
frames ndarray

Video frames array of shape (N, H, W, 3).

 required

Returns:

Type Description
list[tuple[float, float, float, float] | None]

List of face positions (cx, cy, w, h) for each frame, or None if
list[tuple[float, float, float, float] | None]

no face detected and no fallback available.
Source code in src/videopython/ai/transforms.py 
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
def track_video(
    self,
    frames: np.ndarray,
) -> list[tuple[float, float, float, float] | None]:
    """Track face through entire video using optimized batch detection.

    This method is optimized for GPU backends with frame sampling and
    interpolation for smooth tracking with reduced computation.

    Args:
        frames: Video frames array of shape (N, H, W, 3).

    Returns:
        List of face positions (cx, cy, w, h) for each frame, or None if
        no face detected and no fallback available.
    """
    if self._detector is None:
        self._init_detector()
        assert self._detector is not None

    n_frames = len(frames)
    if n_frames == 0:
        return []

    h, w = frames[0].shape[:2]

    # Determine which frames to sample
    if self.sample_rate > 1 and self.backend in ("gpu", "auto"):
        sample_indices = list(range(0, n_frames, self.sample_rate))
        # Ensure last frame is included
        if sample_indices[-1] != n_frames - 1:
            sample_indices.append(n_frames - 1)
    else:
        sample_indices = list(range(n_frames))

    # Batch detect on sampled frames
    sampled_frames = [frames[i] for i in sample_indices]

    # Process in batches
    sampled_detections: list[list] = []
    for batch_start in range(0, len(sampled_frames), self.batch_size):
        batch_end = min(batch_start + self.batch_size, len(sampled_frames))
        batch = sampled_frames[batch_start:batch_end]
        batch_results = self._detector.detect_batch(batch)
        sampled_detections.extend(batch_results)

    # Extract face info from detections
    sampled_faces: list[tuple[float, float, float, float] | None] = []
    for faces in sampled_detections:
        face_info = self._select_face(faces, w, h)
        sampled_faces.append(face_info)

    # If no sampling, apply smoothing directly
    if self.sample_rate == 1 or self.backend == "cpu":
        self.reset()
        results: list[tuple[float, float, float, float] | None] = []
        for i, face_info in enumerate(sampled_faces):
            if face_info:
                cx, cy, fw, fh = face_info
                self._last_position = (cx, cy)
                self._last_size = (fw, fh)

                if self._smoothed_position is None:
                    self._smoothed_position = (cx, cy)
                    self._smoothed_size = (fw, fh)
                else:
                    alpha = 1 - self.smoothing
                    self._smoothed_position = (
                        self._smoothed_position[0] * self.smoothing + cx * alpha,
                        self._smoothed_position[1] * self.smoothing + cy * alpha,
                    )
                    assert self._smoothed_size is not None
                    self._smoothed_size = (
                        self._smoothed_size[0] * self.smoothing + fw * alpha,
                        self._smoothed_size[1] * self.smoothing + fh * alpha,
                    )

                results.append((*self._smoothed_position, *self._smoothed_size))
            elif self._smoothed_position and self._smoothed_size:
                results.append((*self._smoothed_position, *self._smoothed_size))
            else:
                results.append(None)
        return results

    # Interpolate between sampled frames
    all_positions: list[tuple[float, float, float, float] | None] = [None] * n_frames

    # Fill in sampled positions
    for idx, sample_idx in enumerate(sample_indices):
        all_positions[sample_idx] = sampled_faces[idx]

    # Interpolate gaps
    for i in range(len(sample_indices) - 1):
        start_idx = sample_indices[i]
        end_idx = sample_indices[i + 1]
        start_face = sampled_faces[i]
        end_face = sampled_faces[i + 1]

        if start_face is None and end_face is None:
            continue
        elif start_face is None:
            # Use end face for all
            for j in range(start_idx, end_idx):
                all_positions[j] = end_face
        elif end_face is None:
            # Use start face for all
            for j in range(start_idx + 1, end_idx + 1):
                all_positions[j] = start_face
        else:
            # Interpolate
            gap = end_idx - start_idx
            for j in range(start_idx + 1, end_idx):
                t = (j - start_idx) / gap
                all_positions[j] = self._interpolate_bbox(start_face, end_face, t)

    # Apply smoothing to interpolated positions
    self.reset()
    results = []
    for face_info in all_positions:
        if face_info:
            cx, cy, fw, fh = face_info

            if self._smoothed_position is None:
                self._smoothed_position = (cx, cy)
                self._smoothed_size = (fw, fh)
            else:
                alpha = 1 - self.smoothing
                self._smoothed_position = (
                    self._smoothed_position[0] * self.smoothing + cx * alpha,
                    self._smoothed_position[1] * self.smoothing + cy * alpha,
                )
                assert self._smoothed_size is not None
                self._smoothed_size = (
                    self._smoothed_size[0] * self.smoothing + fw * alpha,
                    self._smoothed_size[1] * self.smoothing + fh * alpha,
                )

            results.append((*self._smoothed_position, *self._smoothed_size))
        elif self._smoothed_position and self._smoothed_size:
            results.append((*self._smoothed_position, *self._smoothed_size))
        else:
            results.append(None)

    return results

FaceTrackingCrop

FaceTrackingCrop

Bases: Transformation

Crops video to follow detected faces.

Useful for creating vertical (9:16) content from horizontal (16:9) video by tracking the speaker's face and keeping it framed.

Supports GPU acceleration for faster processing with optional frame sampling. Also supports simple cinematographic framing rules (headroom / thirds) and optional movement speed clamping.

Example

CPU (default, backward compatible)

video = FaceTrackingCrop().apply(video)

GPU with frame sampling for speed

video = FaceTrackingCrop(backend="gpu", sample_rate=5).apply(video)

Source code in src/videopython/ai/transforms.py 
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
class FaceTrackingCrop(Transformation):
    """Crops video to follow detected faces.

    Useful for creating vertical (9:16) content from horizontal (16:9) video
    by tracking the speaker's face and keeping it framed.

    Supports GPU acceleration for faster processing with optional frame sampling.
    Also supports simple cinematographic framing rules (headroom / thirds) and
    optional movement speed clamping.

    Example:
        >>> # CPU (default, backward compatible)
        >>> video = FaceTrackingCrop().apply(video)
        >>>
        >>> # GPU with frame sampling for speed
        >>> video = FaceTrackingCrop(backend="gpu", sample_rate=5).apply(video)
    """

    def __init__(
        self,
        target_aspect: tuple[int, int] = (9, 16),
        face_selection: Literal["largest", "centered", "index"] = "largest",
        face_index: int | None = None,
        padding: float = 0.3,
        vertical_offset: float = -0.1,
        framing_rule: Literal["offset", "center", "headroom", "thirds", "dynamic"] = "offset",
        headroom: float = 0.15,
        lead_room: float = 0.1,
        smoothing: float = 0.8,
        max_speed: float | None = None,
        fallback: Literal["center", "last_position", "full_frame"] = "last_position",
        detection_interval: int = 3,
        backend: Literal["cpu", "gpu", "auto"] = "cpu",
        sample_rate: int = 1,
    ):
        """Initialize face tracking crop.

        Args:
            target_aspect: Output aspect ratio as (width, height).
            face_selection: Strategy for selecting which face to track.
            face_index: Index of face to track when using "index" selection.
            padding: Extra space around face (0.3 = 30% padding on each side).
            vertical_offset: Legacy vertical position offset used by ``framing_rule="offset"``.
            framing_rule: Subject framing strategy.
                - "offset": Use legacy ``vertical_offset`` behavior.
                - "center": Keep face centered.
                - "headroom": Keep extra room above the face.
                - "thirds": Place face near the upper-third line.
                - "dynamic": Currently same as "headroom".
            headroom: Headroom amount for framing rules that use it.
            lead_room: Reserved for future motion/look-direction framing.
            smoothing: Position smoothing factor (0-1, higher = smoother).
            max_speed: Optional max camera movement per frame (normalized).
            fallback: Behavior when no face detected.
            detection_interval: Frames between face detections.
            backend: Detection backend - "cpu", "gpu", or "auto".
            sample_rate: For GPU backend, detect every Nth frame and interpolate.
        """
        self.target_aspect = target_aspect
        self.face_selection = face_selection
        self.face_index = face_index if face_index is not None else 0
        self.padding = padding
        self.vertical_offset = vertical_offset
        self.framing_rule = framing_rule
        self.headroom = headroom
        self.lead_room = lead_room
        self.smoothing = smoothing
        self.max_speed = max_speed
        self.fallback = fallback
        self.detection_interval = detection_interval
        self.backend: Literal["cpu", "gpu", "auto"] = backend
        self.sample_rate = sample_rate

    def _apply_framing_offset(
        self,
        face_cx: float,
        face_cy: float,
        face_h: float,
    ) -> tuple[float, float]:
        """Apply framing rule to get desired crop center in normalized coords."""
        if self.framing_rule == "offset":
            return (face_cx, face_cy + self.vertical_offset)
        if self.framing_rule == "center":
            return (face_cx, face_cy)
        if self.framing_rule == "headroom":
            return (face_cx, face_cy - self.headroom)
        if self.framing_rule == "thirds":
            return (face_cx, face_cy - (1 / 3 - 0.5))
        if self.framing_rule == "dynamic":
            # Placeholder until motion/look-direction framing is implemented.
            return (face_cx, face_cy - self.headroom)
        return (face_cx, face_cy)

    def _clamp_speed(
        self,
        current: tuple[float, float],
        target: tuple[float, float],
    ) -> tuple[float, float]:
        """Clamp crop-center movement speed if ``max_speed`` is configured."""
        if self.max_speed is None:
            return target

        dx = target[0] - current[0]
        dy = target[1] - current[1]
        distance = (dx**2 + dy**2) ** 0.5

        if distance <= self.max_speed:
            return target
        if distance == 0:
            return target

        scale = self.max_speed / distance
        return (current[0] + dx * scale, current[1] + dy * scale)

    def _calculate_crop_region(
        self,
        face_cx: float,
        face_cy: float,
        face_w: float,
        face_h: float,
        frame_w: int,
        frame_h: int,
        center_position: tuple[float, float] | None = None,
    ) -> tuple[int, int, int, int]:
        """Calculate crop region centered on face with padding and framing.

        Args:
            face_cx, face_cy: Face center in normalized coords.
            face_w, face_h: Face dimensions in normalized coords.
            frame_w, frame_h: Frame dimensions in pixels.
            center_position: Optional crop center override in normalized coords.

        Returns:
            Tuple of (x, y, width, height) for crop region in pixels.
        """
        target_ratio = self.target_aspect[0] / self.target_aspect[1]
        frame_ratio = frame_w / frame_h

        # Calculate crop size to achieve target aspect ratio
        # Use _make_even to ensure H.264 compatibility
        if target_ratio < frame_ratio:
            # Target is taller (e.g., 9:16) - height limited
            crop_h = _make_even(frame_h)
            crop_w = _make_even(int(crop_h * target_ratio))
        else:
            # Target is wider - width limited
            crop_w = _make_even(frame_w)
            crop_h = _make_even(int(crop_w / target_ratio))

        # Calculate minimum crop size based on face + padding
        min_face_dim = max(face_w * frame_w, face_h * frame_h)
        min_crop_dim = min_face_dim * (1 + 2 * self.padding)

        # Ensure crop is at least large enough for face with padding
        if crop_w < min_crop_dim * target_ratio:
            crop_w = _make_even(min(int(min_crop_dim * target_ratio), frame_w))
            crop_h = _make_even(min(int(crop_w / target_ratio), frame_h))

        if center_position is None:
            center_position = self._apply_framing_offset(face_cx, face_cy, face_h)

        center_x = center_position[0] * frame_w
        center_y = center_position[1] * frame_h

        x = int(center_x - crop_w / 2)
        y = int(center_y - crop_h / 2)

        # Clamp to frame bounds
        x = max(0, min(x, frame_w - crop_w))
        y = max(0, min(y, frame_h - crop_h))

        return (x, y, crop_w, crop_h)

    def apply(self, video: Video) -> Video:
        """Apply face tracking crop to video.

        Args:
            video: Input video.

        Returns:
            Video cropped to follow faces.
        """
        tracker = FaceTracker(
            selection_strategy=self.face_selection,
            face_index=self.face_index,
            smoothing=self.smoothing,
            detection_interval=self.detection_interval,
            backend=self.backend,
            sample_rate=self.sample_rate,
        )

        h, w = video.frame_shape[:2]
        target_ratio = self.target_aspect[0] / self.target_aspect[1]

        # Calculate output dimensions maintaining target aspect ratio
        # Use _make_even to ensure H.264 compatibility (requires even dimensions)
        if target_ratio < w / h:
            out_h = _make_even(h)
            out_w = _make_even(int(out_h * target_ratio))
        else:
            out_w = _make_even(w)
            out_h = _make_even(int(out_w / target_ratio))

        # Default crop region (center)
        default_x = (w - out_w) // 2
        default_y = (h - out_h) // 2
        last_crop = (default_x, default_y, out_w, out_h)
        current_position = (0.5, 0.5)

        framing_label = self.framing_rule if self.framing_rule != "offset" else "legacy-offset"
        print(
            "Face tracking crop: "
            f"{w}x{h} -> {out_w}x{out_h} "
            f"({self.target_aspect[0]}:{self.target_aspect[1]}, framing={framing_label})"
        )

        new_frames = []
        for i in tqdm(range(len(video.frames)), desc="Face tracking crop"):
            frame = video.frames[i]
            face_info = tracker.detect_and_track(frame, i)

            if face_info:
                cx, cy, fw, fh = face_info
                target_position = self._apply_framing_offset(cx, cy, fh)
                current_position = self._clamp_speed(current_position, target_position)
                crop = self._calculate_crop_region(
                    cx,
                    cy,
                    fw,
                    fh,
                    w,
                    h,
                    center_position=current_position,
                )
                last_crop = crop
            else:
                # Fallback behavior
                if self.fallback == "center":
                    crop = (default_x, default_y, out_w, out_h)
                elif self.fallback == "last_position":
                    crop = last_crop
                else:  # full_frame
                    crop = (0, 0, w, h)

            x, y, cw, ch = crop
            cropped = frame[y : y + ch, x : x + cw]

            # Resize to output dimensions if needed
            if cropped.shape[1] != out_w or cropped.shape[0] != out_h:
                cropped = cv2.resize(cropped, (out_w, out_h), interpolation=cv2.INTER_AREA)

            new_frames.append(cropped)

        video.frames = np.array(new_frames, dtype=np.uint8)
        return video

__init__ 

__init__(
    target_aspect: tuple[int, int] = (9, 16),
    face_selection: Literal[
        "largest", "centered", "index"
    ] = "largest",
    face_index: int | None = None,
    padding: float = 0.3,
    vertical_offset: float = -0.1,
    framing_rule: Literal[
        "offset", "center", "headroom", "thirds", "dynamic"
    ] = "offset",
    headroom: float = 0.15,
    lead_room: float = 0.1,
    smoothing: float = 0.8,
    max_speed: float | None = None,
    fallback: Literal[
        "center", "last_position", "full_frame"
    ] = "last_position",
    detection_interval: int = 3,
    backend: Literal["cpu", "gpu", "auto"] = "cpu",
    sample_rate: int = 1,
)

Initialize face tracking crop.

Parameters:

Name Type Description Default
target_aspect tuple[int, int]

Output aspect ratio as (width, height).

 (9, 16)
face_selection Literal['largest', 'centered', 'index']

Strategy for selecting which face to track.

 'largest'
face_index int | None

Index of face to track when using "index" selection.

 None
padding float

Extra space around face (0.3 = 30% padding on each side).

 0.3
vertical_offset float

Legacy vertical position offset used by framing_rule="offset".

 -0.1
framing_rule Literal['offset', 'center', 'headroom', 'thirds', 'dynamic']

Subject framing strategy. - "offset": Use legacy vertical_offset behavior. - "center": Keep face centered. - "headroom": Keep extra room above the face. - "thirds": Place face near the upper-third line. - "dynamic": Currently same as "headroom".

 'offset'
headroom float

Headroom amount for framing rules that use it.

 0.15
lead_room float

Reserved for future motion/look-direction framing.

 0.1
smoothing float

Position smoothing factor (0-1, higher = smoother).

 0.8
max_speed float | None

Optional max camera movement per frame (normalized).

 None
fallback Literal['center', 'last_position', 'full_frame']

Behavior when no face detected.

 'last_position'
detection_interval int

Frames between face detections.

 3
backend Literal['cpu', 'gpu', 'auto']

Detection backend - "cpu", "gpu", or "auto".

 'cpu'
sample_rate int

For GPU backend, detect every Nth frame and interpolate.

 1
Source code in src/videopython/ai/transforms.py 
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
def __init__(
    self,
    target_aspect: tuple[int, int] = (9, 16),
    face_selection: Literal["largest", "centered", "index"] = "largest",
    face_index: int | None = None,
    padding: float = 0.3,
    vertical_offset: float = -0.1,
    framing_rule: Literal["offset", "center", "headroom", "thirds", "dynamic"] = "offset",
    headroom: float = 0.15,
    lead_room: float = 0.1,
    smoothing: float = 0.8,
    max_speed: float | None = None,
    fallback: Literal["center", "last_position", "full_frame"] = "last_position",
    detection_interval: int = 3,
    backend: Literal["cpu", "gpu", "auto"] = "cpu",
    sample_rate: int = 1,
):
    """Initialize face tracking crop.

    Args:
        target_aspect: Output aspect ratio as (width, height).
        face_selection: Strategy for selecting which face to track.
        face_index: Index of face to track when using "index" selection.
        padding: Extra space around face (0.3 = 30% padding on each side).
        vertical_offset: Legacy vertical position offset used by ``framing_rule="offset"``.
        framing_rule: Subject framing strategy.
            - "offset": Use legacy ``vertical_offset`` behavior.
            - "center": Keep face centered.
            - "headroom": Keep extra room above the face.
            - "thirds": Place face near the upper-third line.
            - "dynamic": Currently same as "headroom".
        headroom: Headroom amount for framing rules that use it.
        lead_room: Reserved for future motion/look-direction framing.
        smoothing: Position smoothing factor (0-1, higher = smoother).
        max_speed: Optional max camera movement per frame (normalized).
        fallback: Behavior when no face detected.
        detection_interval: Frames between face detections.
        backend: Detection backend - "cpu", "gpu", or "auto".
        sample_rate: For GPU backend, detect every Nth frame and interpolate.
    """
    self.target_aspect = target_aspect
    self.face_selection = face_selection
    self.face_index = face_index if face_index is not None else 0
    self.padding = padding
    self.vertical_offset = vertical_offset
    self.framing_rule = framing_rule
    self.headroom = headroom
    self.lead_room = lead_room
    self.smoothing = smoothing
    self.max_speed = max_speed
    self.fallback = fallback
    self.detection_interval = detection_interval
    self.backend: Literal["cpu", "gpu", "auto"] = backend
    self.sample_rate = sample_rate

apply 

apply(video: Video) -> Video

Apply face tracking crop to video.

Parameters:

Name Type Description Default
video Video

Input video.

 required

Returns:

Type Description
Video

Video cropped to follow faces.
Source code in src/videopython/ai/transforms.py 
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
def apply(self, video: Video) -> Video:
    """Apply face tracking crop to video.

    Args:
        video: Input video.

    Returns:
        Video cropped to follow faces.
    """
    tracker = FaceTracker(
        selection_strategy=self.face_selection,
        face_index=self.face_index,
        smoothing=self.smoothing,
        detection_interval=self.detection_interval,
        backend=self.backend,
        sample_rate=self.sample_rate,
    )

    h, w = video.frame_shape[:2]
    target_ratio = self.target_aspect[0] / self.target_aspect[1]

    # Calculate output dimensions maintaining target aspect ratio
    # Use _make_even to ensure H.264 compatibility (requires even dimensions)
    if target_ratio < w / h:
        out_h = _make_even(h)
        out_w = _make_even(int(out_h * target_ratio))
    else:
        out_w = _make_even(w)
        out_h = _make_even(int(out_w / target_ratio))

    # Default crop region (center)
    default_x = (w - out_w) // 2
    default_y = (h - out_h) // 2
    last_crop = (default_x, default_y, out_w, out_h)
    current_position = (0.5, 0.5)

    framing_label = self.framing_rule if self.framing_rule != "offset" else "legacy-offset"
    print(
        "Face tracking crop: "
        f"{w}x{h} -> {out_w}x{out_h} "
        f"({self.target_aspect[0]}:{self.target_aspect[1]}, framing={framing_label})"
    )

    new_frames = []
    for i in tqdm(range(len(video.frames)), desc="Face tracking crop"):
        frame = video.frames[i]
        face_info = tracker.detect_and_track(frame, i)

        if face_info:
            cx, cy, fw, fh = face_info
            target_position = self._apply_framing_offset(cx, cy, fh)
            current_position = self._clamp_speed(current_position, target_position)
            crop = self._calculate_crop_region(
                cx,
                cy,
                fw,
                fh,
                w,
                h,
                center_position=current_position,
            )
            last_crop = crop
        else:
            # Fallback behavior
            if self.fallback == "center":
                crop = (default_x, default_y, out_w, out_h)
            elif self.fallback == "last_position":
                crop = last_crop
            else:  # full_frame
                crop = (0, 0, w, h)

        x, y, cw, ch = crop
        cropped = frame[y : y + ch, x : x + cw]

        # Resize to output dimensions if needed
        if cropped.shape[1] != out_w or cropped.shape[0] != out_h:
            cropped = cv2.resize(cropped, (out_w, out_h), interpolation=cv2.INTER_AREA)

        new_frames.append(cropped)

    video.frames = np.array(new_frames, dtype=np.uint8)
    return video

SplitScreenComposite

SplitScreenComposite

Bases: Transformation

Arranges multiple face-tracked crops in a grid layout.

Useful for interview-style videos, reaction videos, or showing multiple perspectives simultaneously.

Supports GPU acceleration for faster face tracking.

Source code in src/videopython/ai/transforms.py 
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
class SplitScreenComposite(Transformation):
    """Arranges multiple face-tracked crops in a grid layout.

    Useful for interview-style videos, reaction videos, or showing
    multiple perspectives simultaneously.

    Supports GPU acceleration for faster face tracking.
    """

    def __init__(
        self,
        layout: Literal["2x1", "1x2", "2x2", "1+2", "2+1"] = "2x1",
        output_size: tuple[int, int] | None = None,
        gap: int = 4,
        gap_color: tuple[int, int, int] = (0, 0, 0),
        border_width: int = 0,
        border_color: tuple[int, int, int] = (255, 255, 255),
        face_padding: float = 0.2,
        smoothing: float = 0.8,
        detection_interval: int = 3,
        audio_source: Literal["main", "loudest", "mix"] = "main",
        backend: Literal["cpu", "gpu", "auto"] = "cpu",
        sample_rate: int = 1,
    ):
        """Initialize split screen composite.

        Args:
            layout: Grid layout for cells.
                - "2x1": Two cells side by side (horizontal)
                - "1x2": Two cells stacked (vertical)
                - "2x2": Four cells in 2x2 grid
                - "1+2": One large cell on left, two small on right
                - "2+1": Two small cells on left, one large on right
            output_size: Output dimensions (width, height). If None, use source size.
            gap: Gap between cells in pixels.
            gap_color: Color of gap between cells (RGB).
            border_width: Border width around each cell.
            border_color: Border color (RGB).
            face_padding: Extra space around face in each cell.
            smoothing: Position smoothing factor.
            detection_interval: Frames between face detections.
            audio_source: Audio handling ("main" uses first source).
            backend: Detection backend - "cpu", "gpu", or "auto".
            sample_rate: For GPU backend, detect every Nth frame and interpolate.
        """
        self.layout = layout
        self.output_size = output_size
        self.gap = gap
        self.gap_color = gap_color
        self.border_width = border_width
        self.border_color = border_color
        self.face_padding = face_padding
        self.smoothing = smoothing
        self.detection_interval = detection_interval
        self.audio_source = audio_source
        self.backend: Literal["cpu", "gpu", "auto"] = backend
        self.sample_rate = sample_rate

    def _get_cell_rects(self, width: int, height: int) -> list[tuple[int, int, int, int]]:
        """Calculate cell rectangles for the layout.

        Returns:
            List of (x, y, width, height) tuples for each cell.
        """
        gap = self.gap

        if self.layout == "2x1":
            cell_w = (width - gap) // 2
            return [
                (0, 0, cell_w, height),
                (cell_w + gap, 0, width - cell_w - gap, height),
            ]
        elif self.layout == "1x2":
            cell_h = (height - gap) // 2
            return [
                (0, 0, width, cell_h),
                (0, cell_h + gap, width, height - cell_h - gap),
            ]
        elif self.layout == "2x2":
            cell_w = (width - gap) // 2
            cell_h = (height - gap) // 2
            return [
                (0, 0, cell_w, cell_h),
                (cell_w + gap, 0, width - cell_w - gap, cell_h),
                (0, cell_h + gap, cell_w, height - cell_h - gap),
                (cell_w + gap, cell_h + gap, width - cell_w - gap, height - cell_h - gap),
            ]
        elif self.layout == "1+2":
            # Large cell on left (2/3 width), two small on right (1/3 width)
            large_w = (width - gap) * 2 // 3
            small_w = width - large_w - gap
            small_h = (height - gap) // 2
            return [
                (0, 0, large_w, height),
                (large_w + gap, 0, small_w, small_h),
                (large_w + gap, small_h + gap, small_w, height - small_h - gap),
            ]
        elif self.layout == "2+1":
            # Two small cells on left, large on right
            large_w = (width - gap) * 2 // 3
            small_w = width - large_w - gap
            small_h = (height - gap) // 2
            return [
                (0, 0, small_w, small_h),
                (0, small_h + gap, small_w, height - small_h - gap),
                (small_w + gap, 0, large_w, height),
            ]
        else:
            raise ValueError(f"Unknown layout: {self.layout}")

    def _get_required_sources(self) -> int:
        """Get number of video sources required for the layout."""
        if self.layout in ("2x1", "1x2"):
            return 2
        elif self.layout == "2x2":
            return 4
        else:  # 1+2, 2+1
            return 3

    def apply(self, video: Video, *additional_videos: Video) -> Video:
        """Apply split screen composite to videos.

        Args:
            video: Primary video (used for timing and audio).
            *additional_videos: Additional videos to include in grid.

        Returns:
            Composite video with all sources in grid layout.
        """
        all_videos = [video] + list(additional_videos)
        required = self._get_required_sources()

        if len(all_videos) < required:
            raise ValueError(f"Layout '{self.layout}' requires {required} videos, got {len(all_videos)}")

        # Use first video for timing
        n_frames = len(video.frames)

        # Determine output size
        if self.output_size:
            out_w, out_h = self.output_size
        else:
            out_w, out_h = video.frame_shape[1], video.frame_shape[0]

        # Keep final composite dimensions encoder-friendly by default.
        out_w = _make_even(out_w)
        out_h = _make_even(out_h)

        cell_rects = self._get_cell_rects(out_w, out_h)

        # Create face trackers for each cell
        trackers = [
            FaceTracker(
                selection_strategy="largest",
                smoothing=self.smoothing,
                detection_interval=self.detection_interval,
                backend=self.backend,
                sample_rate=self.sample_rate,
            )
            for _ in range(len(cell_rects))
        ]

        print(f"Creating {self.layout} split screen: {out_w}x{out_h}")

        new_frames = []
        for i in tqdm(range(n_frames), desc="Split screen composite"):
            # Create output frame with gap color
            output = np.full((out_h, out_w, 3), self.gap_color, dtype=np.uint8)

            for cell_idx, (cx, cy, cw, ch) in enumerate(cell_rects):
                if cell_idx >= len(all_videos):
                    break

                src_video = all_videos[cell_idx]
                src_idx = i % len(src_video.frames)
                src_frame = src_video.frames[src_idx]
                src_h, src_w = src_frame.shape[:2]

                # Track face in source
                face_info = trackers[cell_idx].detect_and_track(src_frame, i)

                if face_info:
                    fcx, fcy, fw, fh = face_info
                    # Calculate crop to fit face in cell
                    cell_aspect = cw / ch
                    src_aspect = src_w / src_h

                    if cell_aspect < src_aspect:
                        # Cell is taller - crop width
                        crop_h = _make_even(src_h)
                        crop_w = _make_even(int(crop_h * cell_aspect))
                    else:
                        # Cell is wider - crop height
                        crop_w = _make_even(src_w)
                        crop_h = _make_even(int(crop_w / cell_aspect))

                    # Center on face with padding consideration
                    center_x = int(fcx * src_w)
                    center_y = int(fcy * src_h)

                    crop_x = max(0, min(center_x - crop_w // 2, src_w - crop_w))
                    crop_y = max(0, min(center_y - crop_h // 2, src_h - crop_h))

                    cropped = src_frame[crop_y : crop_y + crop_h, crop_x : crop_x + crop_w]
                else:
                    # Center crop as fallback
                    cell_aspect = cw / ch
                    src_aspect = src_w / src_h

                    if cell_aspect < src_aspect:
                        crop_h = _make_even(src_h)
                        crop_w = _make_even(int(crop_h * cell_aspect))
                        crop_x = (src_w - crop_w) // 2
                        crop_y = 0
                    else:
                        crop_w = _make_even(src_w)
                        crop_h = _make_even(int(crop_w / cell_aspect))
                        crop_x = 0
                        crop_y = (src_h - crop_h) // 2

                    cropped = src_frame[crop_y : crop_y + crop_h, crop_x : crop_x + crop_w]

                # Resize to cell size
                resized = cv2.resize(cropped, (cw, ch), interpolation=cv2.INTER_AREA)

                # Apply border if specified
                if self.border_width > 0:
                    bw = self.border_width
                    resized[:bw, :] = self.border_color
                    resized[-bw:, :] = self.border_color
                    resized[:, :bw] = self.border_color
                    resized[:, -bw:] = self.border_color

                # Place in output
                output[cy : cy + ch, cx : cx + cw] = resized

            new_frames.append(output)

        video.frames = np.array(new_frames, dtype=np.uint8)

        # Audio handling - keep main video audio
        # (mixing multiple audio tracks would require more complex handling)
        return video

__init__ 

__init__(
    layout: Literal[
        "2x1", "1x2", "2x2", "1+2", "2+1"
    ] = "2x1",
    output_size: tuple[int, int] | None = None,
    gap: int = 4,
    gap_color: tuple[int, int, int] = (0, 0, 0),
    border_width: int = 0,
    border_color: tuple[int, int, int] = (255, 255, 255),
    face_padding: float = 0.2,
    smoothing: float = 0.8,
    detection_interval: int = 3,
    audio_source: Literal[
        "main", "loudest", "mix"
    ] = "main",
    backend: Literal["cpu", "gpu", "auto"] = "cpu",
    sample_rate: int = 1,
)

Initialize split screen composite.

Parameters:

Name Type Description Default
layout Literal['2x1', '1x2', '2x2', '1+2', '2+1']

Grid layout for cells. - "2x1": Two cells side by side (horizontal) - "1x2": Two cells stacked (vertical) - "2x2": Four cells in 2x2 grid - "1+2": One large cell on left, two small on right - "2+1": Two small cells on left, one large on right

 '2x1'
output_size tuple[int, int] | None

Output dimensions (width, height). If None, use source size.

 None
gap int

Gap between cells in pixels.

 4
gap_color tuple[int, int, int]

Color of gap between cells (RGB).

 (0, 0, 0)
border_width int

Border width around each cell.

 0
border_color tuple[int, int, int]

Border color (RGB).

 (255, 255, 255)
face_padding float

Extra space around face in each cell.

 0.2
smoothing float

Position smoothing factor.

 0.8
detection_interval int

Frames between face detections.

 3
audio_source Literal['main', 'loudest', 'mix']

Audio handling ("main" uses first source).

 'main'
backend Literal['cpu', 'gpu', 'auto']

Detection backend - "cpu", "gpu", or "auto".

 'cpu'
sample_rate int

For GPU backend, detect every Nth frame and interpolate.

 1
Source code in src/videopython/ai/transforms.py 
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
def __init__(
    self,
    layout: Literal["2x1", "1x2", "2x2", "1+2", "2+1"] = "2x1",
    output_size: tuple[int, int] | None = None,
    gap: int = 4,
    gap_color: tuple[int, int, int] = (0, 0, 0),
    border_width: int = 0,
    border_color: tuple[int, int, int] = (255, 255, 255),
    face_padding: float = 0.2,
    smoothing: float = 0.8,
    detection_interval: int = 3,
    audio_source: Literal["main", "loudest", "mix"] = "main",
    backend: Literal["cpu", "gpu", "auto"] = "cpu",
    sample_rate: int = 1,
):
    """Initialize split screen composite.

    Args:
        layout: Grid layout for cells.
            - "2x1": Two cells side by side (horizontal)
            - "1x2": Two cells stacked (vertical)
            - "2x2": Four cells in 2x2 grid
            - "1+2": One large cell on left, two small on right
            - "2+1": Two small cells on left, one large on right
        output_size: Output dimensions (width, height). If None, use source size.
        gap: Gap between cells in pixels.
        gap_color: Color of gap between cells (RGB).
        border_width: Border width around each cell.
        border_color: Border color (RGB).
        face_padding: Extra space around face in each cell.
        smoothing: Position smoothing factor.
        detection_interval: Frames between face detections.
        audio_source: Audio handling ("main" uses first source).
        backend: Detection backend - "cpu", "gpu", or "auto".
        sample_rate: For GPU backend, detect every Nth frame and interpolate.
    """
    self.layout = layout
    self.output_size = output_size
    self.gap = gap
    self.gap_color = gap_color
    self.border_width = border_width
    self.border_color = border_color
    self.face_padding = face_padding
    self.smoothing = smoothing
    self.detection_interval = detection_interval
    self.audio_source = audio_source
    self.backend: Literal["cpu", "gpu", "auto"] = backend
    self.sample_rate = sample_rate

apply 

apply(video: Video, *additional_videos: Video) -> Video

Apply split screen composite to videos.

Parameters:

Name Type Description Default
video Video

Primary video (used for timing and audio).

 required
*additional_videos Video

Additional videos to include in grid.

 ()

Returns:

Type Description
Video

Composite video with all sources in grid layout.
Source code in src/videopython/ai/transforms.py 
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
def apply(self, video: Video, *additional_videos: Video) -> Video:
    """Apply split screen composite to videos.

    Args:
        video: Primary video (used for timing and audio).
        *additional_videos: Additional videos to include in grid.

    Returns:
        Composite video with all sources in grid layout.
    """
    all_videos = [video] + list(additional_videos)
    required = self._get_required_sources()

    if len(all_videos) < required:
        raise ValueError(f"Layout '{self.layout}' requires {required} videos, got {len(all_videos)}")

    # Use first video for timing
    n_frames = len(video.frames)

    # Determine output size
    if self.output_size:
        out_w, out_h = self.output_size
    else:
        out_w, out_h = video.frame_shape[1], video.frame_shape[0]

    # Keep final composite dimensions encoder-friendly by default.
    out_w = _make_even(out_w)
    out_h = _make_even(out_h)

    cell_rects = self._get_cell_rects(out_w, out_h)

    # Create face trackers for each cell
    trackers = [
        FaceTracker(
            selection_strategy="largest",
            smoothing=self.smoothing,
            detection_interval=self.detection_interval,
            backend=self.backend,
            sample_rate=self.sample_rate,
        )
        for _ in range(len(cell_rects))
    ]

    print(f"Creating {self.layout} split screen: {out_w}x{out_h}")

    new_frames = []
    for i in tqdm(range(n_frames), desc="Split screen composite"):
        # Create output frame with gap color
        output = np.full((out_h, out_w, 3), self.gap_color, dtype=np.uint8)

        for cell_idx, (cx, cy, cw, ch) in enumerate(cell_rects):
            if cell_idx >= len(all_videos):
                break

            src_video = all_videos[cell_idx]
            src_idx = i % len(src_video.frames)
            src_frame = src_video.frames[src_idx]
            src_h, src_w = src_frame.shape[:2]

            # Track face in source
            face_info = trackers[cell_idx].detect_and_track(src_frame, i)

            if face_info:
                fcx, fcy, fw, fh = face_info
                # Calculate crop to fit face in cell
                cell_aspect = cw / ch
                src_aspect = src_w / src_h

                if cell_aspect < src_aspect:
                    # Cell is taller - crop width
                    crop_h = _make_even(src_h)
                    crop_w = _make_even(int(crop_h * cell_aspect))
                else:
                    # Cell is wider - crop height
                    crop_w = _make_even(src_w)
                    crop_h = _make_even(int(crop_w / cell_aspect))

                # Center on face with padding consideration
                center_x = int(fcx * src_w)
                center_y = int(fcy * src_h)

                crop_x = max(0, min(center_x - crop_w // 2, src_w - crop_w))
                crop_y = max(0, min(center_y - crop_h // 2, src_h - crop_h))

                cropped = src_frame[crop_y : crop_y + crop_h, crop_x : crop_x + crop_w]
            else:
                # Center crop as fallback
                cell_aspect = cw / ch
                src_aspect = src_w / src_h

                if cell_aspect < src_aspect:
                    crop_h = _make_even(src_h)
                    crop_w = _make_even(int(crop_h * cell_aspect))
                    crop_x = (src_w - crop_w) // 2
                    crop_y = 0
                else:
                    crop_w = _make_even(src_w)
                    crop_h = _make_even(int(crop_w / cell_aspect))
                    crop_x = 0
                    crop_y = (src_h - crop_h) // 2

                cropped = src_frame[crop_y : crop_y + crop_h, crop_x : crop_x + crop_w]

            # Resize to cell size
            resized = cv2.resize(cropped, (cw, ch), interpolation=cv2.INTER_AREA)

            # Apply border if specified
            if self.border_width > 0:
                bw = self.border_width
                resized[:bw, :] = self.border_color
                resized[-bw:, :] = self.border_color
                resized[:, :bw] = self.border_color
                resized[:, -bw:] = self.border_color

            # Place in output
            output[cy : cy + ch, cx : cx + cw] = resized

        new_frames.append(output)

    video.frames = np.array(new_frames, dtype=np.uint8)

    # Audio handling - keep main video audio
    # (mixing multiple audio tracks would require more complex handling)
    return video