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
fix: keep frame seam outside split-screen thumbnail and per-frame crops
Nika Siradze61e1be1970c2
diff --git a/backend/services/thumbnail_generator.py b/backend/services/thumbnail_generator.pyindex 64539c4..b2799fc 100644--- a/backend/services/thumbnail_generator.py+++ b/backend/services/thumbnail_generator.py@@ -274,6 +274,13 @@ def extract_face_frame(best_face_w = 0best_face_h = 0+ # Track per-frame leftmost/rightmost face centers so we can later+ # tell a TRUE split-screen (two video tiles with a gap between+ # subjects, like Riverside) from an in-person podcast with two+ # heads sharing one camera close to center.+ split_left_cxs = []+ split_right_cxs = []+# Sample frames from the clip's time range, or middle 80% of videoif start_second is not None and end_second is not None and end_second > start_second:start_t = max(0.0, float(start_second))@@ -292,6 +299,7 @@ def extract_face_frame(h, w = frame.shape[:2]faces_detected = detect_faces(detector, frame, src_w, src_h)+ qualifying = []for f in faces_detected:conf = f["confidence"]face_w = f["fw"]@@ -300,6 +308,8 @@ def extract_face_frame(if face_w < w * 0.08:continue+ qualifying.append(f)+# Score: confidence × face size (prefer larger/closer faces)score = conf * (face_w * face_h)if score > best_score:@@ -310,26 +320,77 @@ def extract_face_frame(best_face_w = face_wbest_face_h = face_h+ if len(qualifying) >= 2:+ qualifying.sort(key=lambda f: f["cx"])+ split_left_cxs.append(qualifying[0]["cx"])+ split_right_cxs.append(qualifying[-1]["cx"])+cap.release()if best_frame is None:return Noneh, w = best_frame.shape[:2]-- # Crop a vertical (9:16) slice centered on the face.- # The HTML template shows the photo in the top ~78% (1498px of 1920px),- # so we need the face well-centered in that visible zone.+ mid_x = w // 2+ seam_margin = 20++ # TRUE split-screen detection requires BOTH conditions:+ # 1. ≥3 sampled frames had ≥2 qualifying faces+ # 2. Median horizontal gap between leftmost and rightmost face+ # across those frames exceeds 20% of source width+ # Pure presence on both halves (#1) is not enough — an in-person+ # podcast with two heads sharing one camera will satisfy #1 but+ # not #2, and applying the seam clamp there would push a centered+ # face to the crop edge.+ is_split_screen = False+ winner_side = "left" if best_face_cx < mid_x else "right"+ if len(split_left_cxs) >= 3:+ gap = int(np.median(split_right_cxs)) - int(np.median(split_left_cxs))+ if gap > w * 0.20:+ is_split_screen = True++ # Compute the natural 9:16 crop window.target_ratio = target_width / target_height # 0.5625-crop_h = hcrop_w = int(crop_h * target_ratio)if crop_w > w:crop_w = wcrop_h = int(crop_w / target_ratio)- # Center horizontally on face exactly+ # Seam-exclusion strategy for split-screen sources:+ # a) Natural crop already fits inside the winner's half → no-op+ # and clamp will be a no-op too+ # b) Crop is too wide for one half (square/near-square sources)+ # → shrink crop window to fit, then the resize step below+ # stretches it back to 1080×1920 (tighter zoom, no bleed)+ # c) Shrinking would produce a hyper-zoomed image (< 30% of+ # source width or < 200px) → bail on seam exclusion; some+ # bleed is better than a degenerate thumbnail+ MIN_CROP_W = max(200, int(w * 0.30))+ if is_split_screen:+ half_room = (mid_x - seam_margin) if winner_side == "left" else (w - mid_x - seam_margin)+ if half_room < MIN_CROP_W:+ is_split_screen = False+ elif crop_w > half_room:+ shrunk_w = half_room+ shrunk_h = int(shrunk_w / target_ratio)+ if shrunk_h <= h and shrunk_w >= MIN_CROP_W:+ crop_w = shrunk_w+ crop_h = shrunk_h+ else:+ is_split_screen = False++ # Center horizontally on face, then clamp to keep the seam outside+ # the crop window on true split-screen sources. When the winner+ # face sits very close to the seam, the clamp will push it toward+ # the crop's inner edge — that's intentional and a strictly better+ # failure mode than showing the other speaker's tile bleed.crop_x = best_face_cx - crop_w // 2+ if is_split_screen:+ if winner_side == "left":+ crop_x = min(crop_x, mid_x - crop_w - seam_margin)+ else:+ crop_x = max(crop_x, mid_x + seam_margin)crop_x = max(0, min(crop_x, w - crop_w))# Place face at ~25% from top so it's centered in the visible photodiff --git a/backend/services/video_processor.py b/backend/services/video_processor.pyindex 0d8744f..9780576 100644--- a/backend/services/video_processor.py+++ b/backend/services/video_processor.py@@ -2013,6 +2013,19 @@ def _build_speaker_aware_crop(has_single = len(single_frames) >= 3is_mixed_layout = has_split and has_single+ # TRUE split-screen requires presence on both halves AND a real+ # horizontal gap between the leftmost and rightmost face (median+ # gap > 20% of source width). This filters out in-person podcasts+ # where two heads share one camera near the middle — applying the+ # seam clamp there would push centered crops to the edge.+ true_split = False+ if has_split:+ sorted_frames = [sorted(faces, key=lambda f: f[0]) for _, faces in split_frames]+ lefts = [sf[0][0] for sf in sorted_frames]+ rights = [sf[-1][0] for sf in sorted_frames]+ gap = int(np.median(rights)) - int(np.median(lefts))+ true_split = gap > width * 0.20+# Get unique speakers and build speaker segmentsspeakers = sorted(set(w.get("speaker") for w in transcript_words if w.get("speaker")))@@ -2122,9 +2135,33 @@ def _build_speaker_aware_crop(# Single face or no side mapping: use the best (most confident) facereturn max(faces, key=lambda f: f[3])- def _crop_xy_for_face(cx, cy):- """Compute crop_x and crop_y for a face at (cx, cy)."""+ def _crop_xy_for_face(cx, cy, side=None):+ """Compute crop_x and crop_y for a face at (cx, cy).++ On TRUE split-screen sources (faces on both halves with a real+ gap between them — Riverside/Squadcast tiles, not in-person+ shots), clamp crop_x so the frame seam stays outside the 9:16+ window. Otherwise the other speaker bleeds into the crop and+ the median across frames can land on the gap.++ For square/near-square sources where the 9:16 crop is wider+ than the winner's half, the clamp falls through (seam_limit < 0+ or seam_floor + adj_w > width) and the crop will still include+ the seam. That case requires shrinking adj_w upstream, which+ this per-frame helper can't do — leave it to the thumbnail+ path, which can resize.+ """crop_x = cx - adj_w // 2+ if true_split and side in ("left", "right"):+ seam_margin = 20+ if side == "left":+ seam_limit = mid_x - adj_w - seam_margin+ if seam_limit >= 0:+ crop_x = min(crop_x, seam_limit)+ else:+ seam_floor = mid_x + seam_margin+ if seam_floor + adj_w <= width:+ crop_x = max(crop_x, seam_floor)crop_x = max(0, min(crop_x, width - adj_w))crop_y = max(0, (height - adjusted_h) // 2)return crop_x, crop_y@@ -2139,7 +2176,8 @@ def _build_speaker_aware_crop(if face is None:continuecx, cy, fw, conf = face- crop_x, crop_y = _crop_xy_for_face(cx, cy)+ side_hint = speaker_side.get(speaker) if len(faces) >= 2 else None+ crop_x, crop_y = _crop_xy_for_face(cx, cy, side_hint)reliable = len(faces) >= 2 or speaker not in speaker_sidetimed_crops.append((t, crop_x, crop_y, speaker, reliable))