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 instant cut, undefined cluster_radius, tighter clip selection

Nika Siradze42499673d6bf
diff --git a/backend/cli.py b/backend/cli.py
index 1f20bac..dc26206 100644
--- a/backend/cli.py
+++ b/backend/cli.py
@@ -1076,6 +1076,70 @@ def cmd_thumbnails(args):
print(f" {gray}Edit .podcli/thumbnail-config.json to customize colors, fonts, layout.{reset}\n")
+def cmd_cache(args):
+ """Manage transcription cache."""
+ cache_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", ".podcli", "cache")
+
+ accent = "\033[38;2;212;135;74m"
+ gray = "\033[38;5;245m"
+ green = "\033[38;2;74;222;128m"
+ bold = "\033[1m"
+ reset = "\033[0m"
+
+ action = getattr(args, "cache_action", None) or "status"
+
+ if action == "clear":
+ if os.path.exists(cache_dir):
+ import shutil
+ count = len([f for f in os.listdir(cache_dir) if f.endswith(".json")])
+ shutil.rmtree(cache_dir)
+ print(f"\n {green}✓{reset} Cleared {count} cached transcription(s)")
+ else:
+ print(f"\n {gray}Cache is already empty{reset}")
+ print()
+ return
+
+ # Status (default)
+ print(f"\n {bold}Transcription Cache{reset}")
+ print(f" {'─' * 35}")
+
+ if not os.path.exists(cache_dir):
+ print(f" {gray}Empty — no cached transcriptions{reset}\n")
+ return
+
+ files = [f for f in os.listdir(cache_dir) if f.endswith(".json")]
+ if not files:
+ print(f" {gray}Empty — no cached transcriptions{reset}\n")
+ return
+
+ total_size = 0
+ for fname in files:
+ fpath = os.path.join(cache_dir, fname)
+ size = os.path.getsize(fpath)
+ total_size += size
+
+ # Try to read the cached file to show what video it's for
+ try:
+ with open(fpath) as f:
+ data = json.load(f)
+ n_words = len(data.get("words", []))
+ n_segs = len(data.get("segments", []))
+ lang = data.get("language", "?")
+ mtime = os.path.getmtime(fpath)
+ import datetime
+ age = datetime.datetime.fromtimestamp(mtime).strftime("%b %d %H:%M")
+ print(f" {accent}•{reset} {n_segs} segments, {n_words} words, {lang} {gray}({size/1024:.0f}KB, {age}){reset}")
+ except Exception:
+ print(f" {accent}•{reset} {fname} {gray}({size/1024:.0f}KB){reset}")
+
+ print(f" {'─' * 35}")
+ if total_size > 1024 * 1024:
+ print(f" Total: {bold}{total_size / (1024*1024):.1f}MB{reset} ({len(files)} file{'s' if len(files) != 1 else ''})")
+ else:
+ print(f" Total: {bold}{total_size / 1024:.0f}KB{reset} ({len(files)} file{'s' if len(files) != 1 else ''})")
+ print(f" {gray}Run {accent}podcli cache clear{reset} {gray}to delete all{reset}\n")
+
+
def cmd_info(args):
"""Show system info."""
from services.encoder import get_encoder_info
@@ -1152,10 +1216,17 @@ def print_banner():
print(f" {bold}podcli{reset} v{VERSION}")
+ # Cache info
+ cache_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", ".podcli", "cache")
+ cache_count = 0
+ if os.path.isdir(cache_dir):
+ cache_count = len([f for f in os.listdir(cache_dir) if f.endswith(".json")])
+
# Status — one line
claude_tag = f"{green}✓{reset}" if claude_path else f"{yellow}✗{reset}"
speaker_tag = f"{green}✓{reset}" if speakers_ok else f"{yellow}✗{reset}"
- print(f" {gray}Encoder {green}{encoder_label}{reset} {gray}· Claude {claude_tag} {gray}· Speakers {speaker_tag} {gray}· Knowledge {green}{kb_count}{reset}")
+ cache_tag = f"{green}{cache_count}{reset}" if cache_count else f"{gray}0{reset}"
+ print(f" {gray}Encoder {green}{encoder_label}{reset} {gray}· Claude {claude_tag} {gray}· Speakers {speaker_tag} {gray}· Cache {cache_tag}{reset}")
# Assets — one line if any
try:
@@ -1201,6 +1272,7 @@ def print_help():
print(f" {accent}assets{reset} {gray}<action>{reset} Manage logos, intros, outros")
print(f" {accent}presets{reset} {gray}<action>{reset} Save/load rendering presets")
print(f" {accent}thumbnails{reset} {gray}<title>{reset} Generate thumbnail variations")
+ print(f" {accent}cache{reset} {gray}[clear]{reset} Show/clear transcription cache")
print(f" {accent}info{reset} Show system info (encoder, codecs)")
print()
print(f" {bold}Process options:{reset}")
@@ -1312,6 +1384,12 @@ def main():
thumb.add_argument("--logo", help="Logo (asset name or path)")
thumb.add_argument("-n", "--variations", type=int, default=3, help="Number of variations")
+ # ── cache ──
+ cache_p = sub.add_parser("cache", help="Manage transcription cache")
+ cache_sub = cache_p.add_subparsers(dest="cache_action")
+ cache_sub.add_parser("status", help="Show cache size and contents")
+ cache_sub.add_parser("clear", help="Delete all cached transcriptions")
+
# ── info ──
sub.add_parser("info", help="Show system info (encoder, etc.)")
@@ -1331,6 +1409,8 @@ def main():
cmd_presets(args)
elif args.command == "assets":
cmd_assets(args)
+ elif args.command == "cache":
+ cmd_cache(args)
elif args.command == "info":
cmd_info(args)
else:
@@ -1354,6 +1434,7 @@ def interactive_menu():
print(f" {accent}1{reset} Process a video → shorts + content package")
print(f" {accent}2{reset} Open Web UI")
print(f" {accent}3{reset} Manage assets")
+ print(f" {accent}4{reset} Cache {gray}(view/clear transcription cache){reset}")
print(f" {accent}q{reset} Quit")
print()
@@ -1371,12 +1452,16 @@ def interactive_menu():
print(f"\n {gray}Starting Web UI...{reset}\n")
import subprocess as sp
sp.run(["npm", "run", "ui"], cwd=os.path.join(os.path.dirname(os.path.abspath(__file__)), ".."))
+ interactive_menu()
elif choice == "3":
_interactive_assets()
- elif choice in ("q", "Q", ""):
+ interactive_menu()
+ elif choice == "4":
+ _interactive_cache()
+ elif choice in ("q", "Q"):
return
else:
- print(f"\n {gray}Unknown option.{reset}\n")
+ interactive_menu()
def _clean_path(val):
@@ -1592,6 +1677,28 @@ def _interactive_process():
sys.exit(exit_code)
+def _interactive_cache():
+ """Interactive cache management."""
+ import argparse as _ap
+ # Show status first
+ cmd_cache(_ap.Namespace(cache_action=None))
+
+ accent = "\033[38;2;212;135;74m"
+ gray = "\033[38;5;245m"
+ reset = "\033[0m"
+
+ try:
+ choice = input(f" {gray}Clear cache? (y/n/enter=back){reset} {accent}▸{reset} ").strip().lower()
+ except (EOFError, KeyboardInterrupt):
+ print()
+ return interactive_menu()
+
+ if choice in ("y", "yes"):
+ cmd_cache(_ap.Namespace(cache_action="clear"))
+
+ interactive_menu()
+
+
def _interactive_assets():
"""Interactive asset management."""
from services.asset_store import register, unregister, list_assets
diff --git a/backend/services/claude_suggest.py b/backend/services/claude_suggest.py
index 3afa50e..06b739c 100644
--- a/backend/services/claude_suggest.py
+++ b/backend/services/claude_suggest.py
@@ -69,6 +69,15 @@ DURATION RULES (CRITICAL):
- Minimum: 20 seconds (too short = no value)
- If a moment runs longer than 50 seconds, find a tighter cut point
+CUTTING RULES (CRITICAL):
+- Cut TIGHT. Every second must earn its place.
+- Start at the exact moment the hook hits — no preamble, no "so", no "well"
+- End the INSTANT the point lands — don't let it trail off
+- If the speaker rambles before making a point, start AFTER the ramble
+- If there's filler/tangent in the middle, pick the tighter segment that preserves the core idea
+- The transcript timestamps are your scalpel — use them precisely
+- A 30s clip with zero dead weight beats a 50s clip with 20s of fluff
+
MOMENT SELECTION:
- Must start with a strong hook (first 3 seconds grab attention)
- Must make complete sense standalone — no "as I mentioned" or "going back to"
@@ -76,7 +85,6 @@ MOMENT SELECTION:
- Single focused idea per clip — one concept, fully delivered
- Prioritize: surprising facts, bold claims, founder stories, counterintuitive insights, "aha" moments
- Variety across content types
-- In conversations, include moments where both speakers contribute (back-and-forth exchanges can be powerful shorts)
{f"KNOWLEDGE BASE:{kb_context}" if kb_context else ""}
diff --git a/backend/services/video_processor.py b/backend/services/video_processor.py
index 3498845..15695cf 100644
--- a/backend/services/video_processor.py
+++ b/backend/services/video_processor.py
@@ -488,7 +488,7 @@ def _build_speaker_aware_crop(
for obs_t, obs_cx, obs_fw in face_observations:
if abs(obs_t - t) < 1.0:
for ci, cl in enumerate(clusters):
- if abs(obs_cx - cl["center_x"]) < cluster_radius:
+ if abs(obs_cx - cl["center_x"]) < width * 0.15:
cluster_votes[ci] += 1
if any(cluster_votes):
best = max(range(len(cluster_votes)), key=lambda i: cluster_votes[i])
@@ -537,34 +537,25 @@ def _build_speaker_aware_crop(
default_cluster = speaker_to_cluster.get(dominant_speaker) or clusters[0]
default_x = default_cluster["crop_x"]
- # Build a timeline: list of (time, target_x) keyframes with smooth transitions.
- # Instead of nested if/between (breaks with many segments), use a linear
- # interpolation chain: lerp between keyframes.
- pan_duration = 0.4 # 400ms smooth pan (was 300ms, felt jerky)
+ # Build a timeline: list of (time, target_x) keyframes.
+ # Split-screen: instant cut (panning through the center seam looks bad).
+ # Non-split-screen: smooth 400ms pan.
+ pan_duration = 0.0 if is_split_screen else 0.4
- # Build keyframe list: (time, crop_x)
- # In split-screen, verify target face is visible before panning.
- # If face isn't detected at that time, stay at current position.
keyframes = []
prev_x = default_x
for start_t, end_t, speaker in segments:
cl = speaker_to_cluster.get(speaker)
target_x = cl["crop_x"] if cl else default_x
- # Verify target face is visible at this time (avoid panning to empty space)
- if is_split_screen and cl:
- # Check if any face detection near the target cluster exists at segment start
- face_near_target = any(
- abs(obs_t - start_t) < 1.5 and abs(obs_cx - cl["center_x"]) < cluster_radius
- for obs_t, obs_cx, obs_fw in face_observations
- )
- if not face_near_target:
- target_x = prev_x # Stay at current position
-
if target_x != prev_x:
- # Ease into the new position over pan_duration
- keyframes.append((start_t, prev_x))
- keyframes.append((start_t + pan_duration, target_x))
+ if pan_duration > 0:
+ keyframes.append((start_t, prev_x))
+ keyframes.append((start_t + pan_duration, target_x))
+ else:
+ # Instant cut — jump at speaker change
+ keyframes.append((start_t, prev_x))
+ keyframes.append((start_t + 0.01, target_x))
prev_x = target_x
if not keyframes: