Chii chaitika
Chikwata revatsvagiri vanomwe vakaendesa bepa rearXiv richikurudzira mubairo-wakavakirwa velocity matching (RVM), nzira yekunyatso-tunisisa mamodheru ekufambiswa kwezvido zvevanhu kana zvinangwa zvakanangana nebasa. RVM inogadziridza iyo modhi yevelocity ndima yakananga pachinzvimbo chekuvakazve mikana kubva kune denoising trajectories kana kufungidzira magumo ekupedzisira.
Kunobva kwatumirwa arXiv yemusi waNyamavhuvhu 24, 2026, ine musoro unoti "Kuwedzera Kusimbisa Kudzidza Kwemamodhiyoni eDiffusion kuburikidza neVelocity Matching." Vanyori vanopa mubairo-tuning senzira yekuchinjisa mamodheru ekupararira kune zvinodiwa nevanhu uye nezvinangwa zvakanangana nebasa. Chirevo chavo chepakati ndechekuti nzira dziripo dzinokwereta mutemo-gradient muchina wakagadzirirwa autoregressive modhi, zvichigadzira kuwedzera kuomarara nekuti mamodheru ekuparadzira haape mikana inofambika yemasampuli akagadzirwa.
Iyo yakarongwa nzira, mubairo-based velocity matching, inotsanangurwa senge trajectory-isina. Panzvimbo pekugadzira mikana kubva kune stochastic denoising shanduko kana kufungidzira magumo ekupedzisira ane humbowo hwakaderera, RVM inobata zvakananga pandima yevelocity. Zvinoenderana neiyo bepa's abstract, iyo yekuvandudza inosimbisa mafambiro ane hukama neane mibairo mibairo mikuru uye inodzvanya mafambiro ane chekuita nezvizvarwa zvine mubairo wakaderera.
RVM inosanganisira inosarudzika anchor izwi rakagadzirirwa kudzora kudonha kubva kune referensi velocity. Vanyori vanotiwo chimiro chinogona kudzoreredza nzira dzemazuva ano dzekugadzirisa, kusanganisira RAM uye DiffusionNFT, semakesi akakosha. Izvi zvinoita kuti chikumbiro chive chimiro chekubatanidza pamwe nemutemo mutsva wekuvandudza, kunyangwe kwaunobva kwacho kusingatsanangure mabviro chaiwo kana mamiriro ane izvo zvakaenzana.
Vatsvaguri vanoshuma bvunzo pane dzakasiyana siyana-yakakura-yakakura-modhi mubairo-zvakanaka-tuning mabasa. Vanoti RVM yaikwikwidza kana kuti yakapfuura nzira-yakavakirwa mutemo-gradient nzira uku ichida mari shoma yekudzidzisa. Kune chizvarwa chevhidhiyo, vanoshuma kuti mibairo yakajairika yekuda inogona kuburitsa zvinooneka zvakachena asi zvakapotsa zvabuda; vanounza mubairo wekutevera-simba uye vanoti yakavandudza kufamba uku ichivandudza mashandiro eVBench. Hapana zvibodzwa zvenhamba, mazita emhando, mabhajeti ekudzidziswa kana matafura ekuenzanisa anosanganisirwa mune yakapihwa sosi.
Nei zvichikosha
Iro bepa rinopokana kuti yakananga velocity inogadziridza inogona kurerutsa uye kuderedza mutengo wemubairo wekugadzirisa zvakanaka kune yakakura-mapikicha uye vhidhiyo diffusion masisitimu. Vhidhiyo yekuedza kwayo inozivisawo kutengeserana-kusiyana mune yakajairwa mibairo yekuda: vanogona kufarira zvinoonekwa zvakachena zvinobuda zvine zvishoma kufamba.
Kana zvichemo zvebepa zvakazara, kunyatso kwenenzvera vhezheni yemodhi yemuno inomiririra inogona kuita kuti mubairo-kugadzirisa kuve nyore kuyera. Kukosha kunoshanda hakusi kungoita kutsva kwekurasikirwa basa: iyo nzira inonangana neiyo computational uye algorithmic pamusoro yakabatana neiyo mukana-based policy optimization. Mutengo wakaderera wekudzidzira unogona kuita kuti kuda kana kuti basa-rakanangana naro riwanikwe kuzvikwata zvinoshanda nemifananidzo mikuru uye majenareta evhidhiyo.
Pepa racho rinotarisawo pakugadzirwa kwemubairo pachawo. Mukuedza kwevhidhiyo yevanyori, mubairo unosimbisa kuchena kwekuona unonzi wakafarirwa zvinobuda nekufamba kushoma. Mubairo wavo wekutsvaga-simba unounzwa senzira yekuyera kufamba zvinobudirira, vachiri kuvandudza bepa rose reVBench mhedzisiro. Izvi zvinoratidza kuti kuvandudza dhizaini system haibvi pane mutemo wekuvandudza chete asiwo pane izvo maitiro ekugadzirisa anokumbirwa kukoshesa.
Iyo yekusarudzika anchor term ine basa nekuti mubairo optimization inogona kutamisa modhi kure nemaitiro ereferenzi. Tsime rinoti anchor inodzora kudonha kubva pareferensi velocity, asi haitaure kuti kutonga ikoko kunokanganisa sei kunaka, kusiyana, kugadzikana kana njodzi yekuwedzeredza kumubairo. Iwo mameseji ndiwo anozoona kana RVM ichibatsira pakudzorwa kuchinjika kwete kungoita benchmark optimization.
Mhedzisiro yacho inoramba iri yekutsvagisa tsvakiridzo kubva kune rimwe bepa rearXiv, kwete humbowo hwekuti diffusion-modhi kudzidziswa kwachinja zvakanyanya. Iyo yakapihwa abstract hairatidze mamodheru akaedzwa, dhataseti, mibairo magnitudes, compute savings, statistical kusiyanisa kana kukundikana nyaya. Izvo zvakare hazviratidze kana iyo nzira inoshanda zvakaenzana kumifananidzo, vhidhiyo uye mamwe madhizaini maapplication, kana kuti mabhenefiti ayo anoenderana nemamwe mibairo dhizaini.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mibvunzo mikuru ndeyekuti RVM yakataurwa mutengo uye kunaka kwakanakira kunobata pane yakazvimiririra yakadzokororwa kuyedza, mibairo mibairo, saizi yemhando uye mabasa echizvarwa, uye kana nzira yacho ichigona kudzivirira optimize kune nhete kana mibairo inorasisa. Nzvimbo yakapihwa haipe mhedzisiro yenhamba, ruzivo rwekuita, compute mabhajeti kana humbowo hwekusimbisa kwekunze.
Kubereka kunofanirwa kutanga kutarisisa pakuenzanisa kwebepa nemaitiro-akavakirwa mutemo-gradient nzira. Cheki dzinobatsira dzinosanganisira modhi imwechete uye basa riri pasi pemabhajeti akaenzanirana, nekuti "yakanyanya kudzikiswa mutengo wekudzidzira" unoreva chete kana accounting ichisanganisira ese akakodzera kudzidziswa uye basa rekuongorora. Iyo yakapihwa kwainopa haipi nhamba yemutengo wechiyero, saka saizi yeyainzi mukana hauzivikanwe.
Iyo dhizaini yemubairo inofanirwa kuongororwa kwakasiyana kubva kune velocity update. Vanyori vanoti, kana iyo velocity yekuvandudza yarerutswa, iyo chaiyo kurasikirwa musiyano inobata zvishoma pane mubairo uye anchor dhizaini. Chirevo ichocho chinoratidza kuti kuvandudzwa kunogona kutsamira zvakanyanya pakuti zvizvarwa zvine mibairo yepamusoro uye yakaderera zvinonyorwa kana zvibodzwa. Miedzo yakazvimirira inofanirwa kusiyanisa mabasa emubairo uye kuyera kana mibairo ichienderera mberi kupfuura zvinangwa zvebepa zvakasarudzwa.
Kune chizvarwa chevhidhiyo, vacherechedzi vanofanira kutarisa kana mibairo-yekutevera inovandudza mafambiro ane musoro kana kungowedzera shanduko inooneka. Iyo sosi inoshuma kufamba kuri nani uye yakagadziridzwa yakazara VBench mhedzisiro, asi haipe ari pasi metrics kana kutsanangura kutadza modes. Ongororo inofanirwa kuongorora mhando yekuona, kuenderana kwechinguvana uye kusiyana-siyana pamwe chete, kusanganisira zviitiko umo kufamba kusingafadzi kana kupokana nezviri kurongwa.
Kukosha kwebepa kwakakura kunoenderana nekuitwa uye kusimbiswa kweruzivo rwusipo mune yakapihwa sosi. Zvinokosha zvisingazivikanwe zvinosanganisira iyo chaiyo velocity parameterization, anchor marongero, modhi uye dhatabheti kufukidzwa, nguva yekudzidziswa, kuberekana kwezvakafananidzwa kuenzanisa, uye kana RVM inoramba yakagadzikana sezvo mibairo zvinangwa zvinoshanduka. Mapepa ekutevera, kodhi yakaburitswa uye zvidzoreso zvakazvimiririra zvaizobatsira kuona kuti chikumbiro inzira yakajairika yekuyera kana mhedzisiro inosungirirwa kune zvimwe zviedzo.