Vocaloid Style Mv

Vocaloid-style MV pipeline (手書き / 文字PV)

This skill turns a song + its lyrics into a finished, showreel-quality lyric MV (16:9 master, 9:16 vertical version, share encodes). It packages a pipeline that was used to make 「残光」 (Music / Lyrics / Movie：NikusonP) — a 3:46 MV with an original V家-style singer, a cracked-sky / afterglow concept, JIZURA typography, two Blender 3D sequences and ~70 beat-locked shots, rendered in ~4 minutes on an RTX 3080.

The goal is a film of similar craft, not a copy. Every song gets its own concept, character, palette, motifs and storyboard derived from its lyrics and structure. The scene library and the 残光 case study are a toolbox and a worked example; reusing their look wholesale defeats the point. references/02-creative-direction.md explains how to find a new concept and gives a variation matrix.

What you have

• template/ — a complete project: the engine (engine/), scene library (engine/scenes/*.js), JIZURA adapter (jizura/), all tools (tools/), Blender shot scripts (blender/), palettes/LUTs, a procedural demo (demo/, synthetic song) and a default commented timeline.
• scripts/new_project.py — scaffold a project from the template (--song, --lyrics, --title, --title-sub, --artist, --credit, --setup). scripts/check_env.py — what's installed.
• references/ — the method. Read the ones you need, when you need them (table at the end).
• The 残光 case study (storyboard, timeline, prompts, contact sheets) is not inside this skill folder. In a repo checkout it is examples/zanko/ (two levels above this file). With only the installed skill, read it online: https://github.com/EGSECDA/vocaloid-style-mv-pipeline/tree/main/examples/zanko (raw: https://raw.githubusercontent.com/EGSECDA/vocaloid-style-mv-pipeline/main/examples/zanko/<FILE>). Every examples/zanko/... path in references/ means this folder.

Requirements

Required: Node 20+, Python 3.10+, ffmpeg/ffprobe, Microsoft Edge or Chrome (Playwright drives it headless with the GPU). Optional but used for full quality: NVIDIA GPU (WebGL + NVENC), an image generator (the kit wraps the Codex CLI's native image tool; any model works with the chroma-key method), Blender 5.x (3D shots), the lyric-timing set (faster-whisper, onnxruntime, pykakasi + CPU torch/torchaudio 2.8; only when the lyrics have no timestamps). Run python <skill>/scripts/check_env.py first (<skill> = the folder containing this SKILL.md) and tell the user what is missing and what that costs (e.g. "no Blender → I'll do the 3D moments as 2.5D").

The workflow (phases, deliverables, checkpoints)

Work in phases; each ends with something on disk you can show. Details, commands and acceptance criteria: references/01-workflow.md.

1. Scaffold — python <skill>/scripts/new_project.py <dir> --song <song.wav> --lyrics <lyrics> --title <title> --title-sub <Latin subtitle> --artist <artist> --credit "Music / Lyrics / Movie：<artist>" --setup. The four strings set Z.TITLE, Z.TITLE_SUB, Z.ARTIST and Z.CREDIT in engine/timeline.js (HUD, end card, 9:16 header and footer); ask the user for the exact credit line. --credit without --artist takes the artist from after the credit's colon; --title without --title-sub clears the demo's subtitle. Setup installs deps, fetches fonts, clones JIZURA (pinned to the tested commit) and builds it.
   • Plain-text lyrics (no timestamps — the usual case) need forced alignment for the timings: add --setup-args "-Lyrics" (macOS / Linux: --setup-args "--lyrics"), then install the alignment deps once inside the project: python -m pip install --target vendor/pydeps_torch torch==2.8.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cpu.
   • Smoke test (recommended on a new machine): run the demo in a separate throwaway project — python <skill>/scripts/new_project.py _smoke --setup, then the steps it prints (demo/README.md). Never run it in the film's project: demo/make_demo_song.py writes audio/song.wav (it refuses to replace an existing file unless --force), and the demo steps overwrite analysis/lyrics_mv.lrc and analysis/sections.json. Run it before scaffolding the film and the film's project can reuse its fonts: --setup-args "-FontsFrom ../_smoke/engine/fonts" (bash: --fonts-from; sibling folders; combine as "-Lyrics -FontsFrom …").
2. Listen — python tools/analyze_audio.py → analysis/audio.json (beats, downbeats, bars, sections, stops, impacts) + envelope_30fps.json. Curate section labels. Lyrics: always ask for / use the official lyrics; transcription is only for when nothing else exists, forced alignment gives plain-text lyrics their timings. Produce analysis/lyrics_mv.lrc with JIZURA markup. → references/03-audio-and-lyrics.md
3. Concept — from the lyrics: thesis, logline, emotional arc, colour script, master shapes/motifs, character bible, signature moments. Write docs/CONCEPT.md. Checkpoint: show the user the concept in a few lines before generating art. → references/02-creative-direction.md
4. Character & assets — write the character bible, generate 3 master candidates, let the user (or you) pick REF_master.png, then every pose with the master as reference + a shared style block; backgrounds from the storyboard's locations; chroma-key to RGBA. Checkpoint: contact sheet of masters. → references/04-asset-generation.md
5. Storyboard → timeline — shot list mapped to bars and lyric lines (docs/STORYBOARD.md), then engine/timeline.js (shots with scene/args/lyric plan/post/transition). Times come from audio.json (B(n) bar starts, nb(t) nearest beat) so every cut lands on the music. → references/01-workflow.md, references/05-engine.md
6. Build — reuse/adapt scenes from the library, write new ones for the song's signature moments, set up JIZURA plans + curation (references/06-jizura.md), optional Blender sequences (references/07-blender-3d.md). Parallelise: one agent per scene file (disjoint ownership), each with a test timeline and a visual review loop.
7. Review — stills at shot midpoints → contact sheets → fix; then full render (~4 min) → review sheets every 1.5–2 s → fix → re-render. Use the checklist and known failure modes in references/09-qa-and-lessons.md. This loop is where most of the quality comes from; budget for 2–3 passes.
8. Deliver — 16:9 master (NVENC CQ16), share encode (x264 CRF 19), 9:16 vertical re-layout, previews; credits in the title lockup, outro and end card (ask the user for the exact credit line, e.g. "Music / Lyrics / Movie：NikusonP"). Optional upload (tools/gigafile_upload.mjs). → references/08-render-and-delivery.md

Principles that make it look like a showreel (why they matter)

• Everything is a pure function of song time t. Parallel render workers seek arbitrary frames; any wall-clock or unseeded randomness makes frames differ between workers and flicker at segment joins. Particles are closed-form; noise is seeded (Z.rnd, Z.rng, Z.noise1).
• The beat grid is law. Tempo drifts (128→132 BPM in 残光), so cuts use real beat/bar times from audio.json, never n * 60/bpm. Lyrics appear ~0.2 s before the voice; cuts land on downbeats; big moments on chorus downbeats, stops and post-chorus hits.
• Layers, not one canvas. Scene layer (illustration, boil + palette grade) → text layer (JIZURA lyrics, custom type, HUD; stays crisp) → foreground layer (a character drawn in front of the lyrics — text passing behind the singer is one of the strongest looks) → WebGL post (line boil on 2s, gradient-map LUTs, bloom, CA, grain, flashes).
• Restraint, then bursts. Quiet verses with long holds and one effect at a time make choruses hit. Escalate per section; the last chorus brings back earlier ideas "on fire" plus one new move. End on an idea (the 残光 ending is a hard cut on the song's dead stop and a cyan retinal afterimage).
• Typography is half the film. JIZURA gives a huge vocabulary, but it must be curated (some parts draw opaque cards, the wrong colour schemes vanish — dark text at night, ivory text on a daylight sky — gimmick layouts break the mood). Custom typographic scenes carry the キメ lines.
• Match the film's key. The template, the scene library and the 残光 notes are tuned for low-key dusk/night frames: dark-scheme lyrics (light text), bloom from a 0.72 threshold, amber rims and glows, black type cards, the dusk palettes P1–P5. For a bright film (summer, noon, sea, snow) change those defaults before the first still: references/02-creative-direction.md §13.
• Original character, consistent. Never imitate an existing character (no Hatsune Miku likeness). One master reference image anchors every pose.
• Look at every result. Render stills, view them (images are readable), critique like an art director, fix, repeat. Don't trust that code "should" look right.

Working style

• Keep the user in the loop at the checkpoints (concept, character master, first stills, first full render) — a 3–6 line summary plus an image is enough.
• Prefer parallel subagents for independent scene files / asset batches / Blender shots, each owning disjoint files and saving work to disk early (long runs can be interrupted by usage limits; on-disk work survives and can be resumed).
• Measure, don't guess: per-frame render cost (tools/stills.mjs prints ms), render throughput, asset counts.
• Credit the artist on screen and in README files; ask for the credit line if not given.

References (read when the phase needs it)

Author of the kit and of 「残光」: NikusonP. JIZURA 字面: github.com/852wa/JIZURA, MIT licence (Copyright (c) 2026 hakoniwa).