Yakha Amavidiyo E-AI Ekhanda Elikhulumayo Ngesithombe Esisodwa Ngo-Digital Human Koubo

Yini i-Digital Human Koubo?

I-Digital Human Koubo iyi-Agent Skill yomthombo ovulekile yokukhiqiza amavidiyo e-AI anomethuli okhuluma ebheke kwikhamera. Kusukela esithombeni sereferensi esivunyelwe kanye nesicelo sokuqukethwe, i-ejenti ye-AI ingalungisa umbhalo, ikhiqize iziqeshana zomethuli ezihambisana nokunyakaza kwezindebe, ihlole ikhwalithi yazo, ihlanganise izwi, yengeze amagama-ncazo nokupakishwa okubukwayo, bese ikhipha i-MP4 ephelele nge-Remotion.

Ukucushwa komethuli kugcinwa ku-persona.json, okuhlanganisa ireferensi yobuso, izingubo, incazelo yezwi, nethempulethi ye-prompt. Lokhu kwenza kube nokwenzeka ukusebenzisa umuntu ofanayo wedijithali ezitudiyo zokusakaza bukhoma, emakilasini, ezitolo, emigwaqweni nakwezinye izigcawu ezikhiqiziwe.

Sebenzisa kuphela ubuso nezwi lakho, noma izinto onemvume ecacile yokuzisebenzisa. Ungazenzi umuntu wangempela noma umdali, futhi maka amavidiyo aqediwe njengemidlalo yabantu bedijithali eyenziwe yi-AI.

Izici ezibalulekile

• Umuntu wedijithali osebenziseka kabusha: Dala umethuli ongaguquki ngesithombe sereferensi esisodwa kanye nokucushwa komuntu okugcinwe.
• Ukukhiqiza okuhambisana nokunyakaza kwezindebe: Hlukanisa imibhalo ibe izingxenye ezingedluli imizuzwana engu-15, bese uzikhiqiza nge-Grok Video 1.5 usebenzisa i-Renoise MCP.
• Ukusetshenziswa kwemali okusekelwe ekuvunyweni: I-ejenti ibika inani lezingxenye, ubude besikhathi, isilinganiso samakhredithi, nebhalansi ye-akhawunti ngaphambi kokukhiqiza. Ilinda ukuvunywa bese ikhiqiza ingxenye yokuqala ukuze ibuyekezwe ngaphambi kokukhiqiza ezinye ngobuningi.
• Ukulawulwa kwekhwalithi okuzenzakalelayo: Inkulumo ekhiqiziwe ibhalwa phansi nge-whisper bese iqhathaniswa nombhalo. Amafreyimu amane esiqeshini ngasinye nawo ayahlolwa ukuze kutholakale izinkinga ezifana namagama-ncazo ashiselwe ngaphakathi angafuneki.
• Ukufana kwezwi: Ukuqhathanisa i-spectral, ukususa umsindo we-sibilance, i-reverb yegumbi efanayo, nokulinganisa umsindo kwenza iziqeshana ezikhiqizwe ngokuhlukana zizwakale njengokuqoshwa okukodwa.
• Izitayela zokuhlela ezinhlanu: I-ejenti incoma izitayela ngokwesihloko, ifomethi, ubude besikhathi, nezethameli ezihlosiwe.
• Iqoqo lamathuluzi okupakisha: Izingxenye ezingama-48 zinikeza ama-hook, amakhadi obufakazi, izibali, amamephu, amakhadi ezahluko, amagama-ncazo agqwayizayo, ama-punch-in, nezinye izindlela zokuhlela.
• Izilimi eziningi: IsiNgisi, isiJapane, isiKorea, isiFulentshi, isiSpanishi, nesi-Arabhu kuhloliwe, kusetshenziswa i-whisper ukuqinisekisa inkulumo.
• Umbhalo ojulile ongazikhethela: Ku-macOS, i-Apple Vision person segmentation ingabeka umbhalo omkhulu phakathi kwengemuva nomethuli.

Qonda ukuhamba komsebenzi wokukhiqiza

Le skill iqondisa i-ejenti ehambisanayo kule pipeline elandelayo:

1. Hlola i-Renoise MCP, i-ffmpeg, i-Python, i-Node.js, i-Remotion, ne-whisper.
2. Hlaziya okuqukethwe, ncoma izitayela ezimbili zokuhlela, bese uqoqe ukukhetha okudingekayo emzuliswaneni owodwa wemibuzo.
3. Dala umbhalo ohlukaniswe izingxenye ku-script.json, ugcine iziqeshana zokukhuluma ezikhiqiziwe zingaphakathi komkhawulo wemizuzwana engu-15.
4. Caphuna izindleko zamakhredithi e-Renoise ezilindelekile bese ulinda ukuqinisekiswa.
5. Khiqiza ingxenye yokuqala, cela ukuvunywa, bese ukhiqiza izingxenye ezisele.
6. Landa iziqeshana ngokuphepha, bhala inkulumo phansi, bese uhlola amafreyimu amelela okuqukethwe.
7. Sebenzisa ukuhlanganisa izwi bese udala umugqa wesikhathi osezingeni lamagama ku-tl.json.
8. Hlela ngethempulethi yesitayela se-Remotion kanye nezingxenye zokupakisha ezikhethiwe.
9. Khipha ividiyo, hlola izinguquko, bese ulinganisa umsindo wokugcina.

Ngaphambi kokufaka

Ukuze wenze wonke umsebenzi uzenzekele, i-ejenti yakho kufanele ikwazi ukulayisha ama-Agent Skills, ixhume kuseva ye-MCP ekude, futhi isebenzise imiyalo yasendaweni. I-Claude Code yasetshenziswa ukwakha nokuhlola le repository. I-Codex kulindeleke ukuthi isebenze nemibhalo efanayo, kanti i-OpenClaw, i-WorkBuddy, neminye imikhiqizo ingasebenza uma inikeza lawa makhono amathathu.

I-ejenti ekwazi kuphela i-MCP nengxoxo isengakwazi ukuhlela ividiyo, icaphune izindleko zokukhiqiza, futhi idale iziqeshana nge-Renoise. Ukulanda kwasendaweni, ukuhlolwa kwekhwalithi, ukuhlela, nokukhipha ividiyo kufanele kuqedwe nge-ejenti ekwazi ukusebenzisa i-shell noma ngokusebenzisa imiyalo ebhaliwe mathupha.

Faka i-Digital Human Koubo

1. Clone i-repository

git clone https://github.com/andyhuo520/digital-human-koubo.git
cd digital-human-koubo

2. Faka womabili ama-Agent Skills

I-repository iqukethe i-skill eyinhloko ye-digital-human-koubo kanye nenjini ye-finance-explainer esetshenziswa yi-Style A. Abasebenzisi be-Claude Code bangawafaka womabili ngale ndlela:

mkdir -p ~/.claude/skills
cp -R skills/digital-human-koubo skills/finance-explainer ~/.claude/skills/

Kwi-Codex, awakopishe kufolda yayo yama-skills:

mkdir -p ~/.codex/skills
cp -R skills/digital-human-koubo skills/finance-explainer ~/.codex/skills/

Uma i-ejenti yakho ingenisa ama-skills ngesixhumi esibonakalayo, dala i-ZIP archive equkethe womabili amafolda:

cd skills
zip -r ../digital-human-skills.zip digital-human-koubo finance-explainer
cd ..

3. Xhuma i-Renoise MCP

Dala i-akhawunti ku-Renoise. Ukuqinisekisa kwenzeka esipheqululini, ngakho awudingi ukunamathisela ukhiye we-API kuphrojekthi.

Kwi-Claude Code, faka i-plugin esemthethweni:

/plugin marketplace add ArcoCodes/renoise-plugins-official
/plugin install renoise@renoise-plugins-official

Kwi-Codex, engeza bese uqinisekisa iseva ye-MCP ekude:

codex mcp add renoise --url https://renoise.ai/api/public/mcp
codex mcp login renoise

Ezinye izinhlelo ezihambisanayo zingasebenzisa https://renoise.ai/api/public/mcp njenge-URL ye-MCP yangokwezifiso ekude. Ngemva kokuxhuma, cela i-agent ukuthi ihlole ibhalansi yakho ye-Renoise. Lokhu kuqinisekisa kokubili ukuqinisekiswa kobunikazi kanye nokufinyelela kumathuluzi ngaphambi kokuba uqale ukukhiqiza izimpahla ezikhokhelwayo.

4. Faka amathuluzi okukhiqiza endaweni

Uhlelo lokusebenza lwendawo ludinga i-ffmpeg, i-Python 3 ene-NumPy ne-Pillow, i-Node.js 18 noma entsha, i-whisper.cpp, kanye nemodeli ye-whisper. Ku-macOS, ukusethwa okujwayelekile kungaba:

brew install ffmpeg python node whisper-cpp
python3 -m pip install numpy pillow

Landa ggml-medium-q5_0.bin eqoqweni lamamodeli e-whisper.cpp bese uyibeka ku-~/models/. Noma, qondisa uhlelo endaweni ehlukile:

export WHISPER_MODEL=/path/to/model.bin

Abasebenzisi be-Claude Code bangasebenzisa iskripthi sokuxilonga esifakiwe. Siyaziba izinc依 ezisefakiwe bese sibala izinketho zanoma yini engekho:

bash ~/.claude/skills/digital-human-koubo/scripts/doctor.sh

Ku-Windows, faka i-ffmpeg, i-Python, ne-Node.js nge-Winget, bese uthola whisper-cli kokukhishiwe kwe-whisper.cpp. Ku-Linux, faka i-ffmpeg ne-Python usebenzisa umphathi wamaphakheji wesistimu, sebenzisa inguqulo ye-LTS ye-Node.js, bese wakha i-whisper.cpp kusuka kukhodi yomthombo uma kudingeka.

5. Khiqiza izimpahla zomsindo

Indawo yokugcina ikhodi ayifaki umsindo ovela kubantu besithathu. Khiqiza iqoqo lokuqala lemiphumela yomsindo engu-30 kanye nomculo wangemuva wesikhashana usebenzisa:

python3 ~/.claude/skills/digital-human-koubo/scripts/make_assets.py

Iskripthi se-new_edit.sh naso senza lesi sinyathelo kumaphrojekthi amasha. Ukuze usebenzise umculo wekhwalithi ephezulu, beka umsindo onelayisensi yokuwusebenzisa ku-<project>/public/music/bgm.mp3.

Dala umuntu wakho wedijithali

1. Khetha isithombe esibheke phambili esinokukhanya okulinganayo, umlomo ovaliwe, nezinwele ezisuswe ebusweni.
2. Cela i-agent ukuthi ikhiqize ishidi lereferensi elinemibono emine elibonisa ingaphambili, izinhlangothi zombili, kanye nengemuva. I-README incoma i-GPT Image 2.5 ngokulungiswa kwe-2K.
3. Hlola isithombe esifanayo nomugqa okhulunyiwe ngamamodeli evidiyo atholakalayo, njenge-Grok Video 1.5 ne-Wan 3.0, bese ukhetha umphumela owuthandayo.
4. Layisha ireferensi egunyaziwe ku-Renoise.
5. Kopisha inani elibuyisiwe le-material:<id> uye ku-skills/digital-human-koubo/persona.json ngaphansi kwe-renoise.reference_material.

Omazisi bezinto abafakwe endaweni yokugcina ikhodi bangabaka-Renoise account yombhali futhi ngeke basebenze kwenye i-account. Bamiselele ngereferensi yakho oyilayishile. Umethuli wesibili angalungiswa ngenkambu ebhalwe phansi ethi other_hosts.me.

Dala ividiyo yakho yokuqala

Uma ikhono namathuluzi sekutholakala, chaza ividiyo ngolimi olujwayelekile. Faka isihloko, ubude obufunwayo, imininingwane eyiqiniso yesevisi, kanye nezicelo ezibonakalayo.

Use my digital human skill to make a 30-second livestream ad:
WorkBuddy course, 10 lessons, ¥99, previously ¥999.

Ungacela futhi incazelo ende noma ukupakishwa okuqondile:

Make a 3-minute explainer about why Silicon Valley Bank collapsed.
Use the finance and technology explainer style.
Add a map route when the countries are mentioned.
Put “¥99” behind the presenter and begin with a teaser cold open.

I-agent kufanele iphendule ngokuxilonga okufushane kokuqukethwe, izitayela ezimbili ezinconyiwe, kanye neqoqo elilodwa lemibuzo ehlanganisa isitayela, isilinganiso, ubude, ubukhulu bamagama-ncazo, nomethuli. Bese ilungisa umbhalo nomklamo wezigcawu ngaphambi kokubonisa isilinganiso samakhredithi.

Ungagunyazi ngokuzenzakalelayo. Hlola inani lezigaba, imizuzwana ephelele, ibhalansi yamanje, namakhredithi alindelekile. Ngemva kokugunyaza, hlola ingxenye yokuqala ukuze uqinisekise ukuhambisana kobuwena, ukuphinyiselwa, ukuvumelanisa kwezindebe, ukubiywa kwesithombe, kanye nokungabikho kwemibhalo engezansi engafuneki. Gunyaza iqoqo elisele kuphela lapho isampula yamukelekile.

Khetha isitayela sokuhlela

• Isitayela A, incazelo yezezimali noma yezobuchwepheshe: Ifomethi engu-16:9 yamavidiyo acishe abe yimizuzu engu-3–20 mayelana nezehlakalo, ubufakazi, idatha, nemibuzo ethi “kungani”.
• Isitayela B, umgqugquzeli okhuluma ngqo: Ifomethi engu-9:16 yamavidiyo emizuzwana engu-30–120 ancoma imikhiqizo, amavidiyo okuthengisa, nemibono. Amagama-ncazo angaba mancane, ajwayelekile, noma aphilayo.
• Isitayela C, ukuhlaziywa kokwethulwa: Ifomethi engu-16:9 yezimemezelo zemikhiqizo noma zamamodeli zemizuzwana engu-30–90, ngokuvamile nomethuli eseduze kwezithombe ezisekelayo.
• Isitayela D, okokufundisa kwesinyathelo ngesinyathelo: Ifomethi engu-16:9 yemihlahlandlela yokusetha nokuqukethwe okufundisayo okuthatha imizuzu engu-3–8.
• Isitayela E, abethuli abaningi nezigcawu eziningi: Ifomethi engu-16:9 yemiboniso yezilimi eziningi ethatha imizuzu engu-3–7, enabethuli noma izindawo eziningi.

Isitayela A siboleka kuphela izimiso zesakhiwo nezokuhlela ezivela kumafomethi adumile ezincazelo. Asisebenzisi ubuso, izwi, indawo, noma amagama omdali.

Yenza ukupakishwa kube ngokwakho

Uhlelo lokupakisha luqukethe izingxenye ezingu-48 ezinamagama. Zihlelwe ngokugxila ekuvuleni okuhehayo, ebufakazini nasengedatheni, ezichasiselweni, kubantu, kusigqi, nasekusebenzelaneni. Izibonelo zifaka i-TeaserColdOpen, NumberSpike, EvidenceCard, MapRoute, LowerThird, PunchInHost, ChapterCard, kanye ne-TextBehind.

Izingxenye zengezwa ohlwini lwe-pack. Isikhathi singabhekisa esigabeni nasegameni elikhulunyiwe, okuvumela imiphumela ukuthi ilandele inkulumo esikhundleni sokuthembela kuphela kumasekhondi aluhlaza:

{
  "pack": [
    {
      "use": "NumberSpike",
      "at": "1:不到十块钱",
      "until": "1:end",
      "text": "¥9.9/节",
      "label": "一节课"
    },
    {
      "use": "Danmaku",
      "items": [
        { "t": "真的假的？", "at": "1:" },
        { "t": "AI 还能帮我干活？", "at": "1:教你", "hot": true }
      ]
    }
  ]
}

Qinisekisa ukucushwa kokuhlela ngaphambi kokunikezela:

python3 scripts/check_edit.py <project>/src

Dala umbhalo ngemuva komethuli

Umphumela wombhalo ojulile usebenzisa ungqimba lwangemuva, umbhalo omkhulu, kanye nomumo womethuli osikiwe obonisa obala. Ukuhlukaniswa komuntu kuncike ku-Apple Vision, ngakho lo mphumela othile udinga i-macOS nakuba yonke enye ingxenye yenqubo isebenza kuzo zonke izinkundla.

python3 scripts/cutout_video.py \
  public/main.mp4 \
  public/main_cut_s4.webm \
  --from 17.2 \
  --to 22.4

Yenza kuphela imizuzwana edinga lo mphumela. Usebenza kahle kakhulu ngesithombe esiphakathi nendawo kanye nesikhala esibonakalayo esizungeze umuntu. Uma umethuli egcwalisa ngaphezu kwamaphesenti angaba ngu-60 ohlaka, iskripthi siyaxwayisa ngokuthi umbhalo ungase ungabonakali. Maka izingxenye ezifanele ngo-"shot": "depth" ukuze zenziwe ngohlaka olufanele kakhulu.

Izeluleko ezithuthukile zokukhiqiza

Gcina ukuqala kushesha

Qala ngesimangalo, isenzo, noma umbuzo oyinhloko phakathi kwemizuzwana emithathu yokuqala. Gwema ukuqala ngokubingelela noma ngelogo. Emizuzwaneni engu-30 yokuqala, wethula ushintsho olubonakalayo noma lokuhlela cishe njalo ngemizuzwana emithathu kuya kwemihlanu.

Sebenzisa ukunyakaza ngokonga

Kumuntu wedijithali osebenzisa ikhamera eyodwa, ukusondeza okungazelelwe emikhawulweni yemisho kunikeza isigqi esibonakalayo esingabizi. Iphethini ebhalwe phansi ishintsha phakathi kohlaka olucishe lube ngu-1.0×, 1.18×, kanye no-1.32×.

Sebenzisa imisindo ekhethekile ngokukhetha

Gcina umculo wangemuva uqhubeka, kodwa gcina imisindo ekhethekile ezenzakalweni ezibalulekile. Ukuthula okufushane okungaba yimizuzwana engu-0.6–1.0 ngaphambi kokudalulwa kungenza lokho kudalulwa kusebenze kangcono.

Buyekeza amaphutha okubhalwa kwenkulumo ngobuhlakani

I-Whisper ingaguqula amagama ezinombolo zesiShayina akhulunyiwe abe izinombolo. Lokhu kunganciphisa amaphuzu abaliwe okufana kombhalo nakuba umsindo ulungile. Lalela izingxenye ezingabazekayo ngaphambi kokukhokha ukuze ziphinde zenziwe.

Gcina isilinganiso sohlaka lomthombo

Umphumela we-Grok ongu-9:16 720p ungu-720×1280. Ungaweluli uwufake kwesinye isilinganiso sohlaka. Sebenzisa ithempulethi efanele, sika ngenhloso, noma uwenze kabusha ngefomethi edingekayo.

Gwema umsindo ophindwe kabili womethuli

Thulisa ithrekhi yevidiyo yomethuli lapho usebenzisa umsindo ohlanziwe. Ukudlala amathrekhi okuqala nacutshunguliwe ndawonye ngokubambezeleka okuncane kudala ukuhlunga okubukhali okuhlanganisa amaza.

Sebenzisa amazwi ahlukile ngokucophelela

Izwi lendabuko le-Grok ngokuvamile linikeza ukuvumelanisa okuhle kakhulu ngoba imodeli idala ukunyakaza kwezindebe ngalelo zwi. I-Xiaomi MiMo TTS inganikeza izwi eliklanywe ngokukhethekile, izwi elikopishiwe eligunyaziwe, noma ukulandisa, kodwa ukufaka elinye esikhundleni sezwi elenziwe kungashiya umehluko ongaba ama-millisecond angu-100. Lokhu akubonakali kakhulu ezithombeni ze-picture-in-picture kunasekuvalweni kwesikrini esigcwele. Ukuze uthole ukufana okuqondile nokuqoshiwe kwakho, bheka imibhalo yekhono ngemodeli eqhutshwa umsindo efana ne-OmniHuman.

Isu lezindleko nokugunyaza

Ukwenza okuqukethwe kusebenzisa amakhredithi e-Renoise, futhi amanani angashintsha. I-README ibhala isikhangiso sokusakaza bukhoma esime mpo semizuzwana engu-27 ngamakhredithi angu-482, kanti izifundo zemizuzu eminingi zazidinga izinkulungwane eziningana zamakhredithi. Phatha lezo zinombolo njengezibonelo, hhayi njengamanani angaguquki.

Hlala uthembele enanini elibukhoma le-ejenti. Buyekeza nokulungiswa okuhleliwe: amanani ayisithenjwa abhalwe phansi ayengamakhredithi angu-17 ngomzuzwana we-Grok Video 1.5 ku-720p, amakhredithi angu-10 ngomzuzwana ku-480p, kanye namakhredithi angu-23 okwenziwa kwe-2K GPT Image 2.5 ngesikhathi kubhalwa i-README.

Uhlu lokuhlola lokushicilela ngokuzibophezela

• Sebenzisa kuphela ubuso namazwi okungewakho noma onemvume ecacile yokuwasebenzisa.
• Ungalokothi ulingise umuntu wangempela noma usho ngokusobala ukuthi umethuli we-AI uwukugunyazwa komuntu wangempela.
• Bhala ukuthi ividiyo yenziwe umuntu wedijithali we-AI, futhi ubonise izigcawu noma izilwane ezenziwe njengezidalwe yi-AI.
• Ungaqambi amanani, ukubuyekezwa, okunikezwayo, izibalo, noma izimangalo zamakhasimende. Maka ngokucacile idatha eyisibonelo.
• Ungafaki izithombe zamafilimu ezinamalungelo obunikazi, amafreyimu kamabonakude, izithombe zosaziwayo, noma ama-meme ewebhu angenayo ilayisensi.
• Yisho u-Google Noto lapho usebenzisa izimpahla zama-emoji ezinyakazayo ezifakiwe.
• Hlola izimfuneko zelayisensi ehlukile ze-Remotion, ikakhulukazi ezinkampanini ezinabantu abangaphezu kwabathathu.

Isiphetho

U-Digital Human Koubo uhlanganisa ukuhlela okuqhutshwa yi-ejenti, ukwenziwa kwe-Renoise, ukuqinisekiswa okusekelwe ku-whisper, ukucutshungulwa komsindo, nokuhlela kwe-Remotion kube yinqubo yokukhiqiza elawulwayo. Indlela ephephe kakhulu yokuqala ukufaka womabili amakhono, uqinisekise konke okudingekayo, ulungiselele ubuntu bakho obugunyaziwe, bese wenza ividiyo yokuhlola emfushane. Gunyaza inani kanye nengxenye yokuqala ngokucophelela ngaphambi kokwandisela kumaphrojekthi amade, anezilimi eziningi, noma ahlanganiswe kakhulu.