Skip to content

Open-source AI podcast clipper. Generate vertical clips with face tracking and burned-in captions. CLI, MCP server, web app.

fix: split-screen clustering, thumbnail styling, and outro-to-thumb transition

Nika Siradze7958c1744e3a
diff --git a/backend/cli.py b/backend/cli.py
index 5d7fa1c..31defd4 100644
--- a/backend/cli.py
+++ b/backend/cli.py
@@ -401,10 +401,10 @@ def cmd_process(args):
)
if paths:
thumb_video = os.path.join(clip_thumb_dir, "thumb_frame.mp4")
- _thumb_to_video(paths[0], thumb_video)
+ _thumb_to_video(paths[0], thumb_video, duration=1.5)
from services.video_processor import concat_outro
final_with_thumb = result["output_path"].replace(".mp4", "_with_thumb.mp4")
- concat_outro(result["output_path"], thumb_video, final_with_thumb, crossfade_duration=0.3)
+ concat_outro(result["output_path"], thumb_video, final_with_thumb, crossfade_duration=0.15)
os.replace(final_with_thumb, result["output_path"])
print(f" + thumbnail appended ({len(paths)} variations in {os.path.basename(clip_thumb_dir)}/)")
except Exception as e:
diff --git a/backend/services/thumbnail_generator.py b/backend/services/thumbnail_generator.py
index f8e61c4..891d18b 100644
--- a/backend/services/thumbnail_generator.py
+++ b/backend/services/thumbnail_generator.py
@@ -578,16 +578,10 @@ def generate_thumbnail(
x = (w - tw) // 2
if line2_style == "highlight":
- try:
- draw.rounded_rectangle(
- [x - hl_pad_x, y - hl_pad_y, x + tw + hl_pad_x, y + th + hl_pad_y],
- radius=12, fill=(*accent_color, 255),
- )
- except AttributeError:
- draw.rectangle(
- [x - hl_pad_x, y - hl_pad_y, x + tw + hl_pad_x, y + th + hl_pad_y],
- fill=(*accent_color, 255),
- )
+ draw.rectangle(
+ [x - hl_pad_x, y - hl_pad_y, x + tw + hl_pad_x, y + th + hl_pad_y],
+ fill=(*accent_color, 255),
+ )
draw.text((x, y), line, fill=(*dark_text, 255), font=font2)
y += th + hl_pad_y * 2
else:
diff --git a/backend/services/thumbnail_html.py b/backend/services/thumbnail_html.py
index c5150fb..f75be79 100644
--- a/backend/services/thumbnail_html.py
+++ b/backend/services/thumbnail_html.py
@@ -36,29 +36,29 @@ def _load_config() -> dict:
"frame_border_color": None, # None = accent
# Text box
- "box_x": "110px", # CSS left
+ "box_x": "40px", # CSS left — near edge with safe area
"box_y": "73%", # CSS top (no photo)
"box_y_with_photo": "76%",
- "box_width": "860px", # CSS width
- "box_min_height": "220px",
+ "box_width": "1000px", # CSS width — max width with safe areas
+ "box_min_height": "180px",
"box_border_width": 3,
"box_border_color": None, # None = accent
"box_fill_color": "rgba(26,26,46,0.90)",
- "box_padding": "28px 36px",
+ "box_padding": "24px 32px",
# Line 1 (top line — typically white text)
- "line1_font_size": "64px",
+ "line1_font_size": "56px",
"line1_font_weight": "600",
"line1_letter_spacing": "1px",
"line1_line_height": 1.15,
- "line1_margin_bottom": "10px",
+ "line1_margin_bottom": "8px",
"line1_uppercase": True,
"line1_color": "#FFFFFF",
"line1_nowrap": True,
# Line 2 (bottom line — highlighted)
- "line2_font_size": "60px",
- "line2_font_weight": "500",
+ "line2_font_size": "52px",
+ "line2_font_weight": "400",
"line2_font_style": "italic",
"line2_letter_spacing": "1px",
"line2_line_height": 1.15,
@@ -364,7 +364,7 @@ body {{
background: {hl_color};
color: {l2_text_color};
padding: {l2_hl_pad};
- border-radius: 6px;
+ border-radius: 0;
display: inline-block;
box-decoration-break: clone;
-webkit-box-decoration-break: clone;
diff --git a/backend/services/video_processor.py b/backend/services/video_processor.py
index 1e4b972..3498845 100644
--- a/backend/services/video_processor.py
+++ b/backend/services/video_processor.py
@@ -411,31 +411,38 @@ def _build_speaker_aware_crop(
if len(face_observations) < 5:
return None
- # Cluster faces by position (left vs right)
+ # Cluster faces by position (left vs right).
+ # Simple split: faces in the left half vs right half of the frame.
+ # This is more robust than radius-based clustering which can merge
+ # faces that are close to the center line.
positions = np.array([obs[1] for obs in face_observations])
- cluster_radius = width * 0.20 # 20% of frame width — groups face positions with natural head movement
+ mid_x = width // 2
- clusters = []
- used = np.zeros(len(positions), dtype=bool)
- sorted_idx = np.argsort(positions)
+ left_pos = positions[positions < mid_x]
+ right_pos = positions[positions >= mid_x]
- for idx in sorted_idx:
- if used[idx]:
- continue
- mask = np.abs(positions - positions[idx]) < cluster_radius
- mask &= ~used
- if np.any(mask):
- cluster_center = int(np.median(positions[mask]))
- clusters.append({
- "center_x": cluster_center,
- "count": int(np.sum(mask)),
- "crop_x": max(0, min(cluster_center - crop_w // 2, width - crop_w)),
- })
- used |= mask
+ clusters = []
+ if len(left_pos) >= 3:
+ cx = int(np.median(left_pos))
+ clusters.append({
+ "center_x": cx,
+ "count": len(left_pos),
+ "crop_x": max(0, min(cx - crop_w // 2, mid_x - crop_w)),
+ })
+ if len(right_pos) >= 3:
+ cx = int(np.median(right_pos))
+ clusters.append({
+ "center_x": cx,
+ "count": len(right_pos),
+ "crop_x": max(mid_x, min(cx - crop_w // 2, width - crop_w)),
+ })
if len(clusters) < 2:
- if clusters:
- return str(clusters[0]["crop_x"])
+ # Not a clear split-screen — try single cluster from all positions
+ if len(positions) >= 3:
+ cx = int(np.median(positions))
+ crop_x = max(0, min(cx - crop_w // 2, width - crop_w))
+ return str(crop_x)
return None
# Sort clusters left to right
@@ -986,52 +993,24 @@ def _detect_face_center(
if not face_positions:
return None
- # Cluster face positions to detect split-screen layouts.
- # Use a tighter radius (15%) so faces at 25% and 75% don't merge.
+ # Detect split-screen: check if faces appear in both halves
positions = np.array(face_positions)
- cluster_radius = width * 0.15
-
- clusters = []
- used = np.zeros(len(positions), dtype=bool)
- sorted_idx = np.argsort(positions)
-
- for idx in sorted_idx:
- if used[idx]:
- continue
- mask = np.abs(positions - positions[idx]) < cluster_radius
- mask &= ~used
- if np.any(mask):
- cluster_center = int(np.median(positions[mask]))
- clusters.append({
- "center_x": cluster_center,
- "count": int(np.sum(mask)),
- })
- used |= mask
-
- if len(clusters) >= 2:
- # Split-screen: faces in distinct horizontal groups.
- # Pick the cluster with the most detections, then crop to
- # that HALF of the frame (not just center on face — centering
- # still shows the seam between cameras).
- clusters.sort(key=lambda c: c["count"], reverse=True)
- best_cx = clusters[0]["center_x"]
-
- # Snap to the half containing this face
- if best_cx < width // 2:
- # Left half: crop starts at 0
- half_w = width // 2
- face_x = best_cx # within the left half
- else:
- # Right half: crop starts at midpoint
- half_w = width - width // 2
- face_x = best_cx
-
- crop_x = face_x - crop_w // 2
- # Clamp to stay within the chosen half
- if best_cx < width // 2:
- crop_x = max(0, min(crop_x, half_w - crop_w))
+ mid_x = width // 2
+
+ left_count = np.sum(positions < mid_x)
+ right_count = np.sum(positions >= mid_x)
+
+ if left_count >= 3 and right_count >= 3:
+ # Split-screen — pick the side with the most detections,
+ # clamp crop to that half so we never show the seam.
+ if left_count >= right_count:
+ side_pos = positions[positions < mid_x]
+ face_x = int(np.median(side_pos))
+ crop_x = max(0, min(face_x - crop_w // 2, mid_x - crop_w))
else:
- crop_x = max(width // 2, min(crop_x, width - crop_w))
+ side_pos = positions[positions >= mid_x]
+ face_x = int(np.median(side_pos))
+ crop_x = max(mid_x, min(face_x - crop_w // 2, width - crop_w))
else:
face_x = int(np.median(face_positions))
crop_x = face_x - crop_w // 2