CommunityImagemgithub.com

pixel-cinema-mv

Shows an agent how to direct or produce "pixel cinema mv" so the concrete output is an edit with timed captions. The procedure moves through Core Engineering & Aesthetic Rules (Non-Negotiable), 7-Phase Production Pipeline, Phase 1: Audio Inspection & Whisper Exact Sync. Use it when a new assignment calls for this craft.

O que é pixel-cinema-mv?

pixel-cinema-mv is a Claude Code agent skill that shows an agent how to direct or produce "pixel cinema mv" so the concrete output is an edit with timed captions. The procedure moves through Core Engineering & Aesthetic Rules (Non-Negotiable), 7-Phase Production Pipeline, Phase 1: Audio Inspection & Whisper Exact Sync. Use it when a new assignment calls for this craft.

Funciona com~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/L22000011/pixel-cinema-mv/tree/HEAD

Installed? Explore more Imagem skills: steipete/songsee, affaan-m/frontend-design, mattpocock/improve-codebase-architecture · View all 6 →

Perguntar na sua IA favorita

Abre um novo chat com esta habilidade de agente já pré-carregada.

Documentação

Pixel Cinema MV v2.0 — Software Pixel Framebuffer & Whisper-Synced MV Workflow

Create cinematic, HUD-free 16-bit/32-bit indie pixel-art narrative music videos synced to exact Whisper vocal onsets and audio beats, rendered via a deterministic Uint32Array software pixel framebuffer (270×480 → ×4 nearest-neighbor 1080×1920, 30 fps, BT.709).


Core Engineering & Aesthetic Rules (Non-Negotiable)

  1. Pure Uint32Array Software Pixel Framebuffer (Zero Mixels / Zero CSS Blur)

    • Draw every background, light cone, weather particle, prop, and character onto an integer logical framebuffer (270×480 for 9:16 portrait, 480×270 for 16:9 landscape) via templates/pixel_fb_engine/src/engine/PixelFB.ts.
    • Upscale strictly 4× via imageRendering: "pixelated" + --image-format=png lossless intermediate frames.
    • Use ordered 4×4 / 8×8 Bayer dithering for night skies, streetlamp cones, and sunrise halos; use beginSpr()/endSpr() for automatic 1px selective-hue pixel outlines and punch(2) for crisp 2× integer close-ups.
  2. Whisper Exact Vocal-Onset Sync (Never Trust Unverified Internet LRC)

    • Always run scripts/transcribe_whisper_sync.py directly on the user's local audio track (ffmpeg float32 PCM decode → local faster-whisper with word_timestamps=True).
    • Map every lyric cue and every individual Chinese character's at timestamp directly to the real sung onset in the extracted clip so Zpix character bounce highlights fire on the exact syllable.
  3. Causal Physical Storytelling & Zero Game HUD

    • Every shot follows a strict causal chain (Trigger → Action → Contact/Feedback → Result → Transition), never static slideshow/Ken Burns panning.
    • Zero HUD clutter (STAGE, BPM, SCORE, MEMORY LOG are strictly forbidden).
    • Keep main action in the vertical 14%–72% safe zone and place Zpix binary-edged pixel subtitles in 74%–84% (clearing Douyin/TikTok bottom caption & right icon zones).
  4. Anti-Dead-Gap Integer Camera Rule

    • Because camAt() snaps to integer logical pixels (Math.round), never leave a slow scene camera static or stalled at a sine extremum for > 1.0s.
    • Combine steady linear camera drift (>= 0.8 px/s) with continuous ambient motion (window rain streaks, swaying curtains, lamp dust motes, drifting leaves/snow) so film_gate.py clears dead_gap <= 1.0s and moving_share >= 0.90.
  5. Standardized Naming & BT.709 Delivery

    • Tag final H.264 MP4 with -color_primaries bt709 -color_trc bt709 -colorspace bt709.
    • Deliver final video named <歌曲名字>-<时长>s.mp4 (e.g., 我落泪情绪零碎-96s.mp4) alongside .srt, .lrc, _contact_sheet.png, and _qa_report.json.

7-Phase Production Pipeline

Phase 1: Audio Inspection & Whisper Exact Sync

  1. Check audio/ for the user's licensed track (if missing, synthesize a tempo-matched temporary OST and provide scripts/swap_audio.py).
  2. Slice the target emotional highlight (60s–96s) with ffmpeg (-ar 44100 -ac 2, smooth fade-in/out).
  3. Run scripts/transcribe_whisper_sync.py on the exact clip to extract word/character timestamps:
    python scripts/transcribe_whisper_sync.py \
      --audio audio/song.mp3 --start 43.0 --duration 96.0 \
      --model small --out docs/whisper_exact.json
    
  4. Extract BPM, beats, and RMS energy via song_facts.py.

Phase 2: Storyboard, Palette Constitution & Timeline

  1. Author TASTE.md (5-act color palettes, character canon, anti-slop rules) and STORYBOARD.md (4–6 fixed main scenes + causal shot table).
  2. Generate src/data/lyrics_timeline.json, .srt, and .lrc with exact Whisper character at timestamps and shot boundaries aligned to real vocal transitions.

Phase 3: Visual Design & Still Inspection

  1. Scaffold from templates/pixel_fb_engine/ (PixelFB.ts, Palettes.ts, CastEngine.ts, Scenes.ts, fonts.ts).
  2. Run npm run typecheck and render CharacterDesignSheet + key scene stills (remotion still) and visually inspect character proportions, bangs/eyes clarity, and subtitle safe zones.

Phase 4: 10-Second Pilot Sample Check

Render frames=0-299 (10s) first to verify Zpix font loading, audio sync, and integer pixel crispness before committing to a full render:

npx remotion render src/index.ts LuoleiPixelMV-9x16 qc/pilot_10s.mp4 \
  --frames=0-299 --image-format=png --codec=h264 --audio-codec=aac --crf=16 --overwrite

Phase 5: Full Render & BT.709 Packaging

npx remotion render src/index.ts LuoleiPixelMV-9x16 output/raw.mp4 \
  --image-format=png --codec=h264 --audio-codec=aac --crf=16 --overwrite

ffmpeg -y -i output/raw.mp4 -c:v copy -c:a copy \
  -color_primaries bt709 -color_trc bt709 -colorspace bt709 \
  "output/<歌曲名字>-<时长>s.mp4"

Phase 6: Automated QA Gate & Contact Sheet Sign-Off

Run scripts/verify_mv_v2.py to enforce:

  • dead_gap <= 1.0s (zero frozen spans)
  • flicker (0 luminance spikes)
  • slideshow (moving_share >= 0.60, target 1.0)
  • delivery (bt709=True, CFR 30fps, av_offset_ms=0.0)
  • Visually inspect qc/final_contact_sheet.png across 12 sampled keyframes before delivery.

Habilidades Relacionadas