Ukuphupha ku-Cubes | Mayelana neSayensi Yedatha

lokho kuyathandeka kimina (nakwabanye abaningi) ngoba kuye, ngandlela thile, kwangibona ngikhula ngisuka ekubeni ngumfundi wesikole samabanga aphansi, ngaze ngayofika kophothule (maduze!) ekolishi. Ingxenye engenakuphikwa yobuhle begeyimu ukuphinda idlaleke okungapheli okususelwe esizukulwaneni sayo somhlaba. Kuzinhlelo zamanje zegeyimu, i-Minecraft isebenzisa izinhlobonhlobo zemisebenzi yomsindo ngokuhlangana ukuze ikhiqize ngokwenqubo. [1] imihlaba yayo ibe yizicucu, okungukuthi, $16 izikhathi 16 izikhathi 384$ amabhulokhi, ngendlela evamise (ngaphezulu noma ngaphansi) ukwakha indawo ebukeka 'yemvelo', ehlinzeka ngokucwiliswa okukhulu kwegeyimu.

Umgomo wami ngale phrojekthi kwakuwukubona ukuthi ngingakwazi yini ukwedlula umsindo onekhodi eqinile futhi esikhundleni salokho ngifundise imodeli 'ukuphupha' ngama-voxel. Ngokusebenzisa intuthuko yakamuva kuma-Vector Quantized Variational Autoencoder (VQ-VAE) kanye nama-Transformers, ngakhe ipayipi ukuze ngikhiqize izingcezu zomhlaba ze-3D ezithwebula umongo wesakhiwo sezindawo zegeyimu. Njengokukhipha ukhonkolo, ngangifuna ikhono lokukhiqiza $4$ izingcezu (ezihlelwe ngo-a $2 izikhathi 2$ grid) eyayibukeka njengendawo ye-Minecraft.

Njengenothi eseceleni, lona akuwona umbono wenoveli ngokuphelele, okungukuthi, iChunkGAN [2] inikeza enye indlela yokubhekana nomgomo ofanayo.

Inselele ye-3D Generative Modelling

Kuvidiyo [3] kusukela ngoJanuwari 2026, i-Computerphile ibonise u-Lewis Stuart owagqamisa izingqinamba eziyinhloko ngesizukulwane se-3D futhi ngingakhuthaza abafundi ukuthi banikeze iwashi nokho, ukufingqa amaphuzu abalulekile, isizukulwane se-3D sinzima ngoba amasethi edatha e-3D anzima ukuwathola noma awakho nje futhi ukwengeza ubukhulu benkululeko kwenza izinto zibe nzima kakhulu (cabanga ngenkinga yakudala ye-Three-body [4]). Kufanele kuqashelwe ukuthi ividiyo ikhuluma ngokucacile ngamamodeli okusabalalisa (okudinga idatha enelebula) nakuba ukukhathazeka okuningi kungase kudluliselwe embonweni ovamile wokwenziwa kwe-3D. Enye inkinga imane nje isikali; a $512 izikhathi 512$ isithombe ($2^{18}$ pixels) cishe cishe izobhekwa njengesixazululo esiphansi ngamazinga esimanje kodwa imodeli ye-3D ngokwethembeka okufanayo ingadinga $2^{27}$ amazwi. Amaphuzu engeziwe ngokushesha asikisela izidingo eziphezulu zokubala futhi angenza ngokushesha imisebenzi enjalo ingenzeki.

Ukuze nginqobe ukushoda kwedatha ye-3D okukhulunywe ngayo ngu-Stuart, ngaphendukela ku-Minecraft, okuyinto, ngokubona kwami, iwumthombo ongcono kakhulu wedatha ye-voxel etholakalayo yokukhiqiza indawo. Ngokusebenzisa iskripthi ukuze ngithumele nge-teleport ngomhlaba owenziwe ngaphambili, ngiphoqe injini yegeyimu ukuthi ilayishe futhi inikeze izinkulungwane zezigaxa ezihlukile. Ngisebenzisa umbhalo ohlukile wokukhipha, ngidonse lezi zingcezu ngokuqondile kumafayela esifunda somdlalo. Lokhu kunginike idathasethi enokulingana okuphezulu kwe-semantic; ngokungafani neqoqo lezinto ze-3D ezingahleliwe, lezi zingcezu zimelela indawo eqhubekayo, egelezayo lapho 'ingqondo' yendawo (indlela umbhede womfula ocwila ngayo noma ukuthi intaba iphakama kanjani) ilondolozwe phezu kwemingcele eyinqwaba.

Ukuvala igebe phakathi kwenkimbinkimbi yamavokseli e-3D kanye nemikhawulo yezingxenyekazi zekhompuyutha zesimanje, angikwazanga ukumane ngiphakele izingcezu ezingavuthiwe zibe imodeli nethemba lokungcono kakhulu. Ngangidinga indlela yokufinyeza 'umsindo' wezigidi zamabhuloki ngiwenze ulimi olunengqondo, olucindezelwe. Lokhu kungiholela enhliziyweni yephrojekthi: ipayipi elikhiqizayo elinezigaba ezimbili elifunda kuqala 'ukwenza i-tokenize' isikhala se-3D, bese ifunda 'ukuyikhuluma'.

Ukucutshungulwa Kwedatha

Ukubuka okubalulekile nokho okungekho sobala ukuthi ingxenye ebalulekile yezingcezu zeMinecraft igcwele amabhulokhi 'womoya'. Ukuqaphela okungasho lutho ikakhulukazi ngoba umoya awusona ibhulokhi ngokobuchwepheshe, awukwazi ukuwubeka noma ukuwususa ngendlela ongenza ngayo ngawo wonke amanye amabhulokhi kugeyimu kodwa kunalokho, wukungabi khona kwebhulokhi ngaleso sikhathi. Ku-Minecraft yesimanje, isikhathi esiningi esiqondile siwumoya futhi kanjalo, esikhundleni sokucabangela ngokugcwele $384$ amazinga obude, ngiwukhawulele kuwo $y nge [0, 128]$. Labo abajwayelene kakhulu nesizukulwane somhlaba seMinecraft bazokwazi ukuthi amabhlogo anegethivu $y$-amanani, kuze kube $-64$ futhi ngalesi sikhathi, kufanele ngixolise ngoba ngesikhathi ngisebenzisa lesi sakhiwo, lokhu kwase kuwusulile umqondo wami ngokuphelele. Imodeli engiyethula kulesi sihloko izosebenza kahle uma ucabangela ubude besikhathi esithe thwi esikhudlwana kodwa ngenxa yokwengamela kwami ​​ngeshwa, imiphumela engiyethulayo izophuma esikhathini esikhawulelwe samabhulokhi.

Enothi lamabhulokhi akhawulayo, izingcezu zinamabhulokhi amaningi angabonisi njalo futhi angafaki sandla kumumo ojwayelekile wendawo kodwa adingekayo ukuze kugcinwe ukucwiliswa komdlali. Okungenani kule phrojekthi, ngikhetha ukukhawulela amabhulokhi kumabhulokhi aphezulu angama-30 akha izingcezu ngobuningi.

Ukuthena amagama ngomqondo ongokomfanekiso, kuyasiza, kodwa kuyingxenye yempi. Njengoba kushiwo ngaphambili, ngenxa yokuthi imihlaba ye-Minecraft yakhiwe ikakhulukazi 'umoya' kanye 'netshe,' idathasethi ihlushwa ukungalingani kwesigaba okuhle kakhulu. Ukuvimbela imodeli ekuthatheni 'indlela yokumelana okuncane,' okungukuthi, ukubikezela nje isikhala esingenalutho ukuze ngithole ukulahlekelwa okuphansi, ngisebenzise ukulahlekelwa kwe-Weighted Cross-Entropy. Ngokukala ukulahlekelwa ngokusekelwe ku-log-frequency ephambene yebhulokhi ngayinye, ngiphoqe i-VQ-VAE ukuthi ibeke phambili 'okuncane' kwesakhiwo njengotshani, amanzi, neqhwa.

$umbhalo{isisindo[block]} = frac{1}{log(1.1 + umbhalo{okungenzeka[block]})}$

Ngamagama asobala: lapho uhlobo lwebhulokhi lungavamile kakhulu kudathasethi, kulapho imodeli ihlawuliswa kakhulu ngokuhluleka ukuyibikezela, ukusunduza inethiwekhi ukuze iphathe isiqeshana seqhwa noma umbhede womfula njengokubaluleke kakhulu njengobubanzi betshe nomoya obusa izingcezu eziningi.

Ukubuka konke Kwezakhiwo

Lo mdwebo wokulandelana kwe-mermaid [6] inikeza umbono wenyoni wesakhiwo.

Inkinga ye-Voxel eluhlaza kanye ne-Tokenizing 3D Space

Indlela engenangqondo yokwakha isakhiwo esinjalo singabandakanya ukufunda nokwakha izingcezu ngebhulokhi. Kunenqwaba yezizathu zokuthi kungani lokhu kungeke kulunge kodwa inkinga ebaluleke kakhulu ukuthi kungenzeka kube yinto engenakwenzeka ngokushesha ngaphandle kokuhlinzeka ngesakhiwo se-semantic. Cabanga ukuhlanganisa isethi ye-LEGO nezinkulungwane $1 izikhathi 1$ izitini. Nakuba kungenzeka, izohamba kancane kakhulu futhi ngeke ibe nobuqotho besakhiwo: izingcezu ezincikene ngokuvundlile ngeke zixhunywe futhi empeleni uzobe wakha isethi yemibhoshongo ehlukene. Indlela i-LEGO ebhekana ngayo nalokhu iwukuba namabhulokhi amakhulu, njengesithonjana $2 izikhathi 4$ isitini, esithatha isikhala esivame ukudinga okuningi $1 izikhathi 1$ izingcezu. Kanjalo, ugcwalisa isikhala ngokushesha futhi kunobuqotho obuningi besakhiwo.

Okwesistimu, amagama ekhodi yi- $2 izikhathi 4$ Izitini ze-LEGO. Kusetshenziswa i-VQ-VAE (Vector Quantized Variational AutoEncoder), umgomo uwukwakha i-codebook, okungukuthi, isethi yamasiginesha esakhiwo engawasebenzisa ukuze akhe kabusha izingxenye ezigcwele. Cabanga ngezakhiwo ezifana nengxenye eyisicaba yotshani noma i-blob ye-diorite. Ekusebenziseni kwami, ngivumele i-codebook nge $512$ amakhodi ahlukile.

Ukwenza lokhu, ngisebenzise i-3D Convolutions. Ngenkathi ama-convolutions e-2D eyisinkwa nebhotela lokucubungula isithombe, ama-convolutions e-3D avumela imodeli ukuthi ifunde izinhlamvu ezislayida zinqamula izimbazo zika-X, Y, kanye no-Z ngesikhathi esisodwa. Lokhu kubalulekile ku-Minecraft, lapho ubudlelwano phakathi kwebhulokhi naleli elingaphansi kwayo (amandla adonsela phansi/ukwesekwa) bubaluleke njengobudlelwano bayo nalena eseduze kwayo.

Eminye Imininingwane

Ingxenye ebaluleke kakhulu yalesi sigaba `VectorQuantizer`. Lesi sendlalelo sihlala 'enqenqemeni' yenethiwekhi, okuphoqa amasiginali we-neural aqhubekayo ukuthi adlulele 'kulwazimagama' olungaguquki lwamamo e-3D afundiwe angama-512.

Esinye sezithiyo zami ezinkulu ekuqeqeshweni kwe-VQ-VAE ukushumeka 'okufile', okungukuthi, amagama ekhodi isifaki khodi esingalokothi siwakhethe, okumosha ngempumelelo umthamo wemodeli. Ukuxazulula lokhu, ngengeze indlela 'yokusetha kabusha' amagama afile. Uma ukusetshenziswa kwe-codeword kwehla kakhulu, imodeli iphinda iliqalise ngamandla 'ngokuntshontsha' i-vector eqoqweni lamanje lokokufaka:

Brick by Brick

I-assortment ehlukahlukene yamabhulokhi yinhle kodwa ayisho okuningi ngaphandle kokuthi ehlanganiswe kahle. Ngakho-ke, ukuze ngisebenzise kahle la magama ekhodi, ngisebenzise i-GPT. Ukuze ngenze lo msebenzi, ngithathe igridi efihlekile ekhiqizwe i-VQ-VAE ngenza isethi yamathokheni, empeleni, umhlaba we-3D uyathotshiswa ube olimini lwe-1D. Khona-ke, i-GPT ibona amathokheni angu-8 ekufanele afunde uhlelo lolimi lwendawo, ngomqondo ongokomfanekiso, we-Minecraft ukuze kuzuzwe ukuvumelana kwe-semantic okukhulunywe ngakho ngenhla.

Ukufeza lokhu, ngisebenzise Ukuzinaka Okuvamile:

Ekugcineni, ngesikhathi sokubikezela, imodeli isebenzisa isampula ephezulu-k, kanye nezinga lokushisa elithile ukuze lilawule isizukulwane esiguquguqukayo ubuciko ku-loop yesizukulwane esilandelayo:

Ekupheleni kwalokhu kulandelana, i-GPT 'isibhale' ipulani yesakhiwo engamathokheni angu-256 ubude. Isinyathelo esilandelayo siwukudlulisa lokhu kusikhikhoda se-VQ-VAE ukuze ubonise a $2 izikhathi 2$ igridi yendawo ebonakalayo ye-Minecraft.

Imiphumela

Kulokhu kuhumusha [6]imodeli ihlanganisa ngempumelelo amabhlogo weqabunga, ilingisa izakhiwo zesihlahla somdlalo.

Kulena [6]imodeli isebenzisa amabhlokhi eqhwa ukuze ivale itshe notshani, okubonisa ukuphakama okuphezulu noma izingcezu ze-tundra ezitholakala kudatha yokuqeqeshwa. Ukwengeza, lokhu kuhumusha kubonisa ukuthi imodeli yafunda indlela yokwenza imigede.

Kulesi sithombe [6]imodeli ibeka amanzi ekucindezelekeni futhi iwavale ngesihlabathi, okubonisa ukuthi ifake ngaphakathi ingqondo yendawo yogu, kunokuba ihlakaze amabhlokhi amanzi ngokuzenzakalelayo endaweni.

Mhlawumbe umphumela ohlaba umxhwele kakhulu ukwakheka kwangaphakathi kwezingcezu. Ngenxa yokuthi ukusetshenziswa kusetshenziswe ama-convolutions e-3D nomsebenzi wokulahlekelwa isisindo, imodeli empeleni ikhiqiza izici ezingaphansi komhlaba njengemigede ehlangene, ama-overhangs, namawa.

Nakuba imiphumela ibonakala, ayiwona ama-clones aphelele we-Minecraft. Ukuminyanisa kwe-VQ-VAE 'kuyalahleka,' ngezinye izikhathi kubangela 'ukufiphala' kancane kwemingcele yebhulokhi noma ibhulokhi elintantayo ngezikhathi ezithile. Nokho, kumodeli esebenza endaweni ecashile ecindezelwe kakhulu, ikhono lokulondoloza ubuqotho besakhiwo kuyo yonke indawo $2 izikhathi 2$ I-chunk grid, ngikholwa ukuthi, iyimpumelelo enkulu.

Ukuzindla kanye Nomsebenzi Wakusasa

Ngenkathi imodeli 'iphupha' ngempumelelo kuma-voxel, sikhona isikhala esibalulekile sokunwetshwa. Iziphindaphindo zesikhathi esizayo zingaphinde zivakashele ubude obuqondile obuphelele be- $y nge [−64,320]$ ukwamukela iziqongo ezimangelengele kanye nemigede ejulile 'yoshizi' ewuphawu lwezinguqulo zesimanje zeMinecraft. Ngaphezu kwalokho, ukukala i-codebook ngaphezu kokufakiwe okungu-512 kuzovumela uhlelo ukuthi lufake uphawu lwezakhiwo eziyinkimbinkimbi, ezinjengamadolobhana noma amathempeli asogwadule. Mhlawumbe okujabulisa kakhulu amandla okukhiqiza okunemibandela, noma 'i-biomerizing' i-GPT, engavumela abasebenzisi ukuqondisa inqubo yezakhiwo ngemiyalelo ethize njengokuthi 'Intaba' noma 'Ulwandle,' ukuguqula iphupho elingahleliwe libe ithuluzi lokudala eliqondisiwe.

Siyabonga ngokufunda! Uma unentshisekelo ekusetshenzisweni okugcwele noma ufuna ukuzama izisindo ngokwakho, zizwe ukhululekile ukuhlola indawo yokugcina [5].

Izingcaphuno kanye nezixhumanisi

[1] Abahleli be-Minecraft Wiki, isizukulwane somhlaba (2026),

[2] x3voo, ChunkGAN (2024),

[3] U-Lewis Stuart weComputerphile, Ekhiqiza Amamodeli e-3D ane-Diffusion – Computerphile (2026),

[4] Abahleli Be-Wikipedia, Inkinga Yemizimba Emithathu (2026),

[5] isinkwa sesikhala, irobhothi elikhazimulayo (2026),

[6] Isithombe ngombhali.

Inothi ku-Dataset

Yonke idatha yokuqeqeshwa yenziwe umbhali esebenzisa isibonelo esiqhutshwa endaweni se-Minecraft Java Edition. Ama-Chunks akhishwe kumafayela omhlaba akhiqizwe ngokwenqubo kusetshenziswa iskripthi sokukhipha ngokwezifiso. Awekho amasethi edatha ezinkampani zangaphandle asetshenzisiwe. Njengoba idatha yenziwe futhi yakhishwa umbhali esibonelweni sakhe segeyimu, ayikho imikhawulo yangaphandle yelayisense esebenza ekusetshenzisweni kwayo kulo mongo wocwaningo.