HejuruUbuyobozi bukurikira
Gusuzuma Sisitemu Zitanga Ibyifuzo: NDCG, Kanda Igipimo nibindi
Tekiniki
Ubuyobozi bwa tekiniki
VBench ni urutonde rw'ibipimo byifashishwa mu gusuzuma amamodeli ahindura inyandiko akazigira amashusho (text-to-video).
Umva iki gitaboAI-byavuzwe · ~3 min
Igabanya ubwiza bw'amashusho mu bipimo 16 bitandukanye, nko guhuza ibintu (subject consistency), ubwiza bw'imigendere (motion smoothness), uburyo amashusho agenda (dynamic degree) n'ubwiza bw'ishusho (aesthetic quality), kandi igaha buri kimwe amanota ikoresheje uburyo bwikora bwo gusuzuma. Ibi ni ingenzi kuko amanota amwe adashobora kugaragaza niba amashusho ari mazima, agenda neza, ahuye n'ibyo yasabwe kandi akaba agenda koko, kandi VBench ituma ibyo bintu byose bigaragara.
VBench was introduced in late 2023 by researchers from Nanyang Technological University's S-Lab and Shanghai AI Laboratory, and presented at CVPR 2024. Its core idea is decomposition. Instead of one score such as FVD (Frechet Video Distance), it measures 16 dimensions grouped into two families. Video quality dimensions judge the clip itself, regardless of the prompt: subject consistency, background consistency, temporal flickering, motion smoothness, dynamic degree, aesthetic quality and imaging quality. Video-condition consistency dimensions judge how well the video follows its prompt: object class, multiple objects, human action, color, spatial relationship, scene, appearance style, temporal style and overall consistency. Each dimension has its own prompt suite and a specialized evaluator, usually an existing vision model. Subject consistency compares DINO features across frames. Background consistency uses CLIP features. Dynamic degree uses optical flow. Aesthetic and imaging quality use learned quality predictors, and object and color checks use detection or captioning models. The authors collected human preference annotations for each dimension and reported that the automatic scores agree well with them. The breakdown exposes trade-offs that a total hides. A nearly frozen video can score very well on consistency and smoothness while scoring poorly on dynamic degree, so reading only the average can reward timid models. The VBench total is not a complete ranking of video models; it reflects the chosen prompts, evaluators and weights. Later work extended the suite. VBench++ added image-to-video and trustworthiness evaluation, and VBench-2.0 targets deeper properties such as physics, commonsense and human anatomy. Public arenas, where people vote between two anonymous clips from the same prompt, capture overall preference, including qualities such as physically plausible motion that automated evaluators do not yet measure reliably.
Laboratwari isohora amamodeli y'amashusho afunguye (open video model) itangaza amanota yayo ya VBench yose hamwe n'ay'ibipimo byayo bitandukanye. Iyi mbonerahamwe igaragaza ko iyoboye mu bwiza bw'ishusho ariko ikaba inyuma mu bijyanye n'uburyo amashusho agenda (dynamic degree).
Umuteza imbere w'ikoranabuhanga agereranya imiterere ibiri y'imashini (sampler settings) akabona ko gukoresha intambwe nkeya bigabanya amanota y'ubwiza bw'imigendere (motion smoothness) n'ay'imihindagurikire y'igihe (temporal flickering). Ibi bimufasha guhitamo hagati y'umuvuduko n'ubwiza.
Itsinda rishinzwe ibicuruzwa ririmo guhitamo amashusho ya API rihuza amanota ya VBench n'ibisubizo byavuye mu matora y'abantu ku mugaragaro, kuko abakoresha ibicuruzwa byabo bitaho cyane ku bintu bisa n'ukuri n'abantu.
Umushakashatsi abona model ifite amanota yo hejuru mu bijyanye no guhuza ibintu n'inyuma y'ishusho (subject and background consistency) ariko ikora amashusho adahinduka. Areba igipimo cy'uburyo amashusho agenda (dynamic degree) kugira ngo abyemeze.
Isuzuma ry'amashusho ririmo guhinduka riva ku bwiza bw'inyuma rigana ku kureba niba isi ikozwe n'ikoranabuhanga ikora mu buryo bwumvikana: ibintu biguma ari bikomeye, imbaraga zikurura ibintu (gravity) zikora neza, n'intoki zikomeza kugira umubare w'intoki zikwiye. Ibipimo nka VBench-2.0 n'ibindi byibanda ku mategeko ya fiziki bigamije ibi, akenshi bikoresha amamodeli y'indimi nyinshi (multimodal language models) nk'abacamanza, ibyo bikaba bizamura ibibazo bishya ku bijyanye n'ubwizere bw'abo bacamanza. Amashusho maremare, amajwi, n'igenzura rya kamera na byo bikeneye ibipimo ibikoresho byo ubu bitarageraho byose. Amatora y'abantu azakomeza kuba ingenzi kuko agaragaza ibyo abantu bakunda muri rusange, ariko ahenze, yoroshye kugira uwo abogamira bitewe n'uko ibintu byerekanwe, kandi bigoye kuyasubiramo, bityo birashoboka ko urwego ruzakomeza gukoresha uburyo bwombi.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
VBench ni urutonde rw'ibipimo byifashishwa mu gusuzuma amamodeli ahindura inyandiko akazigira amashusho (text-to-video). Igabanya ubwiza bw'amashusho mu bipimo 16 bitandukanye, nko guhuza ibintu (subject consistency), ubwiza bw'imigendere (motion smoothness), uburyo amashusho agenda (dynamic degree) n'ubwiza bw'ishusho (aesthetic quality), kandi igaha buri kimwe amanota ikoresheje uburyo bwikora bwo gusuzuma. Ibi ni ingenzi kuko amanota amwe adashobora kugaragaza niba amashusho ari mazima, agenda neza, ahuye n'ibyo yasabwe kandi akaba agenda koko, kandi VBench ituma ibyo bintu byose bigaragara.
VBench igabanya ubwiza bw'amashusho mu bipimo 16, byashyizwe mu matsinda y'ubwiza bw'amashusho no guhuza amashusho n'ibyo yasabwe.
Igipimo cy'uburyo amashusho agenda (dynamic degree) gikoresha uburyo bwo gupima imigendere (optical flow) kugira ngo kimenye niba amashusho arimo kugenda. Ishusho idahinduka ihabwa amanota make hano niyo yaba ifite amanota meza mu bijyanye no guhuza ibintu.
Guhuza ibintu (subject consistency) bigereranya ibiranga DINO by'ikintu mu mashusho atandukanye kugira ngo barebe niba gikomeza kugira ishusho imwe.
Ubumenyi bw'ubwiza bw'amavideo buhanura ikiganiro (clip) ariko nta kintu cyerekanwa mu bisabwa (prompt), naho ubumenyi bw'ubwiza bw'amavideo bukurikiza amabwiriza buhanura neza neza uko iva yubahiriza ibisabwa byayo.
Amanota yose ni impuzandengo y'ibice byandikanye n'ibindi. Ubwiza bwo gukurikiza amabwiriza bushobora guhisha ubwiza bwo gukora imyitozo, bityo amanota y'ikice cyose ni ngombwa.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Gusuzuma Sisitemu Zitanga Ibyifuzo: NDCG, Kanda Igipimo nibindi
Tekiniki