OpenClaw Agent
e350352883
Fix: Gemini 3.1 Pro thinking model needs 32k maxOutputTokens (was 4096 → MAX_TOKENS truncation)
...
Diagnoza:
- Gemini 3.x Pro je thinking model (ima internal reasoning, thoughtsTokenCount)
- Pri velikih transkriptih (60+ segmentov pesmi):
* thoughts ~ 1500-3000 tokens
* output JSON s corrected_segments ~ 3000-7000 tokens
* total ~ 4500-10000 tokens
- Z maxOutputTokens=4096 je bil response prekinjen (finishReason: MAX_TOKENS),
JSON odrezan na pol, _parse_llm_response je threw json.JSONDecodeError
- Rezultat: 'Gemini vrnil prazen string' v logih
Popravki:
1. Gemini maxOutputTokens 4096 → 32768 (dovolj za thinking + dolg JSON)
2. Diagnostika finishReason==MAX_TOKENS in usage tokens v logih
3. Detekcija praznega text-a (ne samo praznega parts array-a)
4. Claude max_tokens 4096 → 8192 (rezerva za dolge pesmi)
5. Claude detekcija stop_reason==max_tokens
Test (60 segmentov, 5631 char prompt):
- 4096 → finishReason=MAX_TOKENS, thoughts=2594, output=1488, JSON odrezan ❌
- 16384 → finishReason=STOP, thoughts=1445, output=3040, JSON popoln ✅
- 32768 → varen default ✅
2026-04-29 09:03:53 +00:00
ec71c54570
Upgrade to Sonnet 4.6 + add Gemini 3.1 Pro support
...
- Refactored analyze_with_claude into shared _build_analysis_prompt + _parse_llm_response helpers
- New analyze_with_gemini() using Gemini 3.1 Pro ($2/M in, MMMLU 92.6% — best multilingual)
- Unified analyze_with_llm(provider) dispatcher with auto-fallback (Claude → Gemini)
- API endpoint accepts llm_provider in StartJobIn (claude/gemini/auto)
- Frontend dropdown to pick LLM
- Default model is now Sonnet 4.6 (was Haiku 4.5) — 3x quality at 3x price (~3 cents/video)
- Gemini support is opt-in: needs GEMINI_API_KEY env var to activate
2026-04-29 08:26:27 +00:00
9faa224885
Upgrade Claude model: Haiku 4.5 → Sonnet 4.6 for better Slavic language transcript correction
2026-04-29 08:22:10 +00:00
69fb2f5ce8
Upgrade default Whisper model: small/medium → large-v3 for much better Slovenian/Slavic transcription accuracy
2026-04-29 08:20:18 +00:00
4bc5ac6756
Major: Claude post-processing of Whisper transcript
...
- Claude now corrects transcription errors (Slavic languages, dialects, mixed langs)
- Returns corrected_segments with same timestamps but cleaner text
- Pipeline generates SRT from Claude-corrected transcript and passes to subtitle.py via --srt
- subtitle.py supports --srt to skip Whisper re-transcription on the trimmed clip
- clip.py propagates --srt through to subtitle.py
- Whisper still runs once (in analyze.py); subtitle.py reuses corrected output instead of re-running
- This means: Whisper's mistakes (mixed langs, hallucinations, wrong words) are fixed by Claude before becoming visible subtitles
2026-04-29 08:13:33 +00:00
af3c933c78
Robust language detection + anti-hallucination
...
- 3-sample voting for auto-detect (start/middle/end of song) prevents lang switching mid-song
- Lock detected language for full transcription
- Anti-hallucination: condition_on_previous_text=False, temperature=0.0
- compression_ratio_threshold=2.4 (rejects repetitive hallucinations)
- log_prob_threshold=-1.0 (rejects low-confidence segments)
- no_speech_threshold=0.6 (more aggressive silence detection)
- Default Whisper model changed: small → medium (better for all langs incl. Slavic)
2026-04-29 07:59:20 +00:00
c870d80726
Fix: extend clip if ends mid-vocal (no chorus cut-off), DejaVu Sans font (supports SLO/HR/BS chars), auto-upgrade to medium Whisper model for Slavic languages
2026-04-29 07:35:00 +00:00
5d5e169f9d
Disable Whisper VAD filter — was dropping vocal segments in songs creating gaps in subtitles
2026-04-29 07:07:29 +00:00
a04811bdc9
Add Claude LLM analysis: sends full transcript to Claude API for true song structure understanding (refrain detection across all repetitions, not just local heuristic)
2026-04-29 06:55:41 +00:00
e072eec362
Fix: handle Whisper transcribe failure for instrumental-only audio (fallback to empty transcript)
2026-04-29 06:33:52 +00:00
33a138af9e
Fix: force native Python bool/float for JSON serialization (numpy types)
2026-04-29 06:23:41 +00:00
8512076b91
Major: smart selection pipeline (analyze.py) + audio fade + multi-lang auto-detect
...
- New analyze.py: full transcript + energy + structural analysis
- Smart clip range: includes pre-chorus, can exceed 30s up to max_duration (default 45s)
- Audio fade in/out: auto-detected from vocal boundaries
- Instrumental detection: auto-disables subs if vocals < 10% of duration
- Multi-language: auto-detect via Whisper or explicit (DE/SL/HR/BS/SR/EN/IT/ES/FR)
- Frontend: cleaner UX, added bs language, smart selection description
- reframe.py: --fade-in --fade-out args
- clip.py: propagates fade params
- app/main.py: replaces find_chorus.py call with analyze.py
2026-04-29 06:21:35 +00:00
81edd24ca3
Subtitles: smaller font 56px (was 84), higher position MarginV=400, side margins 80px for safe zone
2026-04-29 06:09:26 +00:00
ba787744a6
Subtitles: cap chunk duration at 2.5s, split long lines into multiple time slices for faster reels pacing
2026-04-29 05:59:36 +00:00
e001387a89
Subtitles: convert SRT to ASS directly with PlayResY=1920 for predictable scaling instead of unreliable force_style
2026-04-28 18:09:53 +00:00
28d933c916
Subtitles: UPPERCASE + position lower (MarginV=320 for 1080x1920) + bigger font
2026-04-28 17:40:48 +00:00
15ef4888a1
Debug: log exact clip.py cmd in job + clip.py logs run_clip args
2026-04-28 17:28:10 +00:00
bc3fe1f9d4
Add explicit FFmpeg trim command logging + duration verification
2026-04-28 17:17:11 +00:00
8eaef029e2
Find chorus: weight repetitive short phrases (like 'Ohne dich x5') as strong chorus signal
2026-04-28 16:57:45 +00:00
c17578521a
Fix find_chorus: RMS energy parser was broken (no pts_time available), now syntheses timestamps; energy weight x10 (refren je glasnejši)
2026-04-28 16:55:51 +00:00
64e8854cea
Track mode: more sensitive face detection + longer smoothing window
2026-04-28 16:45:13 +00:00
400f6dbb6d
Fix: limit FFmpeg crop expression to 20 sample points (was overflowing 4KB limit)
2026-04-28 16:32:26 +00:00
2e337ff079
Fix: shutil import was inside finally block, causing NameError when shutil.move was called
2026-04-28 16:22:39 +00:00
6e2a13d8a3
Fix cross-device link error: use shutil.move instead of os.replace
2026-04-28 16:15:20 +00:00
47509b4f06
Add cookies support to yt_download.py for YouTube bot detection bypass
2026-04-28 15:47:59 +00:00
30b969e4b8
Initial: reels clipper app
...
- FastAPI backend (auth, jobs, SSE, download)
- Frontend: drag&drop + YouTube URL + jobs panel
- Pipeline: yt_download → find_chorus → reframe → subtitle
- Modes: track (face follow), center, blur
- Whisper for SI/DE/EN subtitles
- Auto-chorus detection via Whisper + RMS energy
- Docker + Coolify ready
2026-04-28 15:28:22 +00:00
