Open-source AI podcast clipper. Generate vertical clips with face tracking and burned-in captions. CLI, MCP server, web app.
Python ai-toolsauto-captionscaptionscliffmpegmcpmcp-serveropusclip-alternativepodcast-clipspodcastsshortform-videovertical-videowhisper
Write the comments I added without em dashes (#176)
Nika Siradzed7bc9d395cc5
diff --git a/backend/services/face_analysis.py b/backend/services/face_analysis.pyindex e855a50..29f5a9a 100644--- a/backend/services/face_analysis.py+++ b/backend/services/face_analysis.py@@ -182,7 +182,7 @@ def analyze_faces(if progress_callback:progress_callback(75, "Clustering face positions...")- # Cluster faces into seats — the fixed positions a person occupies.+ # Cluster faces into seats: the fixed positions a person occupies.## Two seats mean two people on screen at once, and the only evidence for# that is a frame holding both. Splitting a flat list of positions at the@@ -212,8 +212,9 @@ def analyze_faces("crop_x": max(mid_x + seam_margin, min(right_seat - crop_w // 2, width - crop_w)),})else:- # One seat. Not "one speaker" — the guest may be off camera entirely,- # and the render should stay on the person it can actually see.+ # One seat, which is not the same as one speaker. The guest may be+ # off camera entirely, and the render should stay on the person it can+ # actually see.cx = int(np.median(positions))clusters_list.append({"center_x": cx,diff --git a/backend/services/face_track_helpers.py b/backend/services/face_track_helpers.pyindex 3aa5ecc..4c09829 100644--- a/backend/services/face_track_helpers.py+++ b/backend/services/face_track_helpers.py@@ -184,10 +184,10 @@ def crop_center_keeping_faces_visible("""Pick the crop center that leaves the most sampled faces in frame.The median of a run of face positions is only the right answer when- those positions are unimodal. When a run mixes two layouts — a- fullscreen shot at 960 and a split-screen tile at 1440 — the median- lands at 1200, which on a 607-wide crop is the wall between them, and- the render holds on nothing for the length of the run.+ those positions are unimodal. When a run mixes two layouts, a+ fullscreen shot at 960 and a split-screen tile at 1440, the median lands+ at 1200, which on a 607-wide crop is the wall between them, and the+ render holds on nothing for the length of the run.So score candidate centers by how many faces they actually keep insidethe crop, and only fall back to closeness as a tiebreak. On a unimodal@@ -243,8 +243,8 @@ def seats_from_frames(paired = [sorted(xs) for xs in frame_positions if len(xs) >= 2]with_faces = sum(1 for xs in frame_positions if xs)- # A share as well as a count, so a handful of false positives — a face in a- # photo on the shelf behind someone — cannot seat a second person.+ # A share as well as a count, so a handful of false positives, a face in a+ # photo on the shelf behind someone, cannot seat a second person.if len(paired) < max(min_paired_frames, with_faces * min_paired_share):return Nonediff --git a/backend/services/video_processor.py b/backend/services/video_processor.pyindex b7a5451..6ed8c2a 100644--- a/backend/services/video_processor.py+++ b/backend/services/video_processor.py@@ -1496,7 +1496,7 @@ def _track_and_crop_inner(t, cx = face_points[i]run_median = float(median(run_cxs))if abs(cx - run_median) > jump_thresh:- # Layout changed — close current run, start new one.+ # Layout changed, so close this run and start a new one.# Close it where the next one opens, not at the previous# sample: the sampler runs at about 12 Hz, so ending a run# one sample early threw away 80ms of video at every cut.@@ -1645,8 +1645,8 @@ def _track_and_crop_inner(# dissolve and keep the framing: one cut per run, joined end to# end. This used to fall back to a static crop of the longest# run, which is right for that run and wrong for every other- # layout in the clip — on a recording that cuts to a fullscreen- # shot, a hold on the wall for a third of its length.+ # layout in the clip. On a recording that cuts to a fullscreen+ # shot, that is a hold on the wall for a third of its length.for ri, (r_start, r_end, r_cx) in enumerate(runs):seg_start = max(0.0, r_start)seg_end = min(duration, r_end)diff --git a/tests/test_choose_segment_tracks.py b/tests/test_choose_segment_tracks.pyindex 65c5235..bc64242 100644--- a/tests/test_choose_segment_tracks.py+++ b/tests/test_choose_segment_tracks.py@@ -201,7 +201,7 @@ class AnchorOnlySegmentTests(unittest.TestCase):def test_unstood_anchor_is_replaced_by_a_real_face(self):# One person at 1400 all clip. The turn's speaker has no track of- # their own and an anchor at 400 — the invented seat.+ # their own and an anchor at 400, which is the invented seat.targets = self._targets(segment_tracks=[(0.0, 1.2, "SPEAKER_01", 7, 400.0)],detections=self._one_person(1400, 8.0),