ad-rebrand

Ad rebrand

Turn a batch of existing ads into the new brand: every spoken mention re-voiced in the original speaker's cloned voice, every caption re-lettered in the video's own caption font, every on-screen logo/tag/label re-lettered or cleanly blanked — with the rest of each ad untouched (same frames, timing, music).

Parallelise with subagents (Agent tool, run_in_background: true, one message launching all independent agents). Prompt templates: references/agent-prompts.md. Read references/lessons.md before starting — it records what broke last time.

Inputs to pin down first

• Old names (product and brand) and the new name — and how whisper might mishear the old ones (e.g. "physio rest", "artivate").
• The files (videos, images). Work in a project folder the user can keep (e.g. ~/Downloads/ Video Edits); never edit the originals.

Workflow

1. Set up. scripts/setup_jobs.sh PROJECT START_N video1 video2 … → job-NN/{assets/source.mp4,analysis,project,renders}. Copy images to PROJECT/images/source/. Environment: ffmpeg/whisper-cli via Homebrew, Python venv ~/.cache/video-studio-venv (numpy, opencv-contrib, pillow, soundfile, demucs, torch; pyworld for sung words), whisper model ~/.cache/whisper/ggml-small.en.bin. If missing, install (Homebrew needs the user's password in a terminal tab).

2. Triage (parallel, one agent per video). Template §1. Each writes analysis/triage.json|md with refined speech spans, speakers + clean clone-reference ranges, every visual/caption occurrence with frame ranges, and oddities. Check images yourself (zoom into tags/labels at full res) — many need nothing.

3. Consolidate and ask once (AskUserQuestion):

   • Voice rights: list every distinct voice that must be cloned (group videos that share narration — compare vocal-stem envelopes; compare median F0 to tell speakers apart). Cloning requires the user's explicit yes per voice.
   • Anything that is not the old brand but looks like a problem: other brands' marks on the product, retail/endorsement settings (store staff, store logos), inconsistent claims, script redundancy from the rename. Third-party watermarks: flag, never remove.
   • Sung mentions: explain they need re-pitching and will need a human listen.

4. Start the visual team immediately (template §2) — it doesn't depend on voices. Group videos that are the same edit into one agent. Image agent: template §5.

5. Voices. Export clone references from each job's vocals stem (25–35 s, no brand names) into PROJECT/voice_refs/. Prefer the Fish API (examples/voice-splice/fish_tts_api.py, key in ~/.config/video-studio/.env) if it has API credit; otherwise create private clones in the Fish web app in the user's Chrome — references/fish-audio-web.md. Generate each voice's lines in one script, both takes, saved to voice_refs/takes/.

6. Audio team (parallel, one agent per voice/job group). Template §3 for speech, §4 for sung words. Output: analysis/stems/vocals_nookwave.wav per job, verified by transcription.

7. Assemble each job as soon as its picture and voice are both ready: scripts/mux_final.py JOB OUT_NAME.mp4 (rebranded picture + new vocals + original music stem). If a visual agent hands back unfinished, run its project/rebrand_job.py yourself.

8. QC every final yourself — agents' reports are claims, not proof: scripts/check_final.py source.mp4 final.mp4 --old "physio|fisio|artuvate|artivate" --new "nook" --frames <one frame per fixed item> → frame-count match, clean decode, loudness vs source, zero spoken old names, and a source|final sheet you must look at. Spot-transcribe 5 s windows around each mention if totals look off (full-file whisper merges sentences). For sung lines expect whisper to struggle on the original too.

9. Deliver. Finals at the project root with descriptive names, a zip if asked (zip -0, videos are already compressed), and a summary: what changed per file, what needs a human listen (sung words, pause lengths after a shorter name), known imperfect frames, and the flagged non-brand issues.

Optional: engagement cuts

For "make it more engaging" requests on top of the rebrand, examples/engagement-edit/ holds the plan builder (sentence-level keep lists snapped to quiet points, callouts, punch-ins) and renderer (zoom alternation on jump cuts, branded callouts, progress bar, synthesized SFX, caption-hold fixes at cuts, loudnorm −14 LUFS). Reuse rather than rewrite.

Bundled files

• scripts/setup_jobs.sh, scripts/mux_final.py, scripts/check_final.py — generic.
• examples/ — working, job-specific code from the first run (paths and frame numbers are hard-coded; adapt, don't run blind):
  • caption-and-tag-library/ — caption redraw, rotating leather tags, curved plates, bag eraser, multi-scale tracking (rebrand_job01.py); glyph-cut captions, engraved/debossed tags, tracked bag re-print (caption_lib.py, text_lib.py, track_lib.py, rebrand_job05.py).
  • web-and-sign-library/ — own-glyph captions (captions.py), building-sign plate, web copy re-set, tag pressing, hand-keyed tumbling bag (rbk.py, rebrand_job02.py), laptop-screen page (rebrand_job06_laptop.py), review-page scroll, end-card logo, third-party emboss removal (rebrand_job07.py), QC sheets (visual_qc.py).
  • voice-splice/ — measured-span word replacement with take overrides; Fish API client.
  • sung-word/ — WORLD-based re-singing onto the original melody.
  • still-image/ — homography text replacement on packaging.