Pinakamahusay na AI video generator: 45 clip, blind na sinuri

• Sinubukan noong 1 Nobyembre 2026
• 5 modelo
• 3 trabaho, tig-3 take
• 45 clip, 0 ang pumalya
• Kasama ang raw data

I-play ang lima, naka-mute

Iisang prompt para sa lahat ng lima. Ito ang take 1 sa tatlo ng bawat modelo, kaya hindi namin pinili ang mga maganda ang kinalabasan. Pindutin ang play sa kahit anong clip.

Ang maikling sagot: kung gusto mo ng iisang modelo para sa lahat, gamitin ang Kling 4.0, dahil ang pinakamahina nitong take (3.6) ay mas mataas kaysa sa pinakamahina ng ibang modelo. Kung mas mahalaga ang presyo, gamitin ang Wan 3.0, na katumbas ng Kling sa kabuuan sa $0.60 bawat clip. Nasa likod nila nang kaunti ang Seedance 2.5 at pantay sila sa mga clip na nagsasalita. Nangunguna ang MiniMax H3 sa mga eksenang may gumagalaw na kamera. Paalala: Ibinebenta ng Hermoso ang access sa lahat ng lima, at ang Seedance 2.5 ang default namin. Hindi ito nanalo nang tahasan sa kahit anong trabaho.

Sino ang dapat gumamit ng alin

Isang modelo para sa lahat

Kling 4.0

Ang pinakamatatag na resulta ng test. Ang pinakamababa nitong take ay 3.6, at walang trabahong naging mahinang bahagi. Kasama ng Wan 3.0 ang nanguna sa kabuuan.

Pinakamagandang halaga

Wan 3.0

Katumbas ng Kling sa kabuuan sa $0.60 para sa 6 na segundo, ang pinakamatalas na detalye sa test, at ang pinakamagandang natural na tunog. Mas mura bawat segundo ang MiniMax H3, pero mas mababa ang iskor nito sa labas ng mga eksena.

Taong nagsasalita sa harap ng kamera

Tabla: Wan 3.0, Seedance 2.5, Kling 4.0

Sinabi ng tatlo ang linya ng eksaktong salita sa salita. Ang Wan ang may pinakamalaking pulbos ng harina at ang mga braso na may harina na hiniling namin, na may ngiting pang-stock ad. Ang Seedance ang pinakamukhang pelikula, pero nagkamali sa bilang sa isang take sa tatlo.

Mga produktong may maliit na print

Kling 4.0, malapit ang Wan 3.0

Pinanatili ng bawat take ng Kling ang tatlong linya ng label na tama, may totoong kamay at totoong push-in. Ito rin ang pinakatumpak sa brief. Ang Seedance ang gumawa ng pinakamagandang pulot sa test, sa isang take sa tatlo.

Gumagalaw na kamera sa abalang eksena

MiniMax H3, sunod ang Wan at Kling

Ang pinakamataas na reveal, at nasa shot pa rin ang siklista sa dulo. May 8 segundo rin ito kung saan 6 lang ang sa iba, na nakakatulong sa mabagal na galaw. Ang Wan 3.0 at Kling 4.0 ay nasa loob ng linya ng tabla namin.

Kapag kailangan mo ito agad

Veo 3.1

Natapos sa humigit-kumulang 66 segundo sa median, laban sa 156 hanggang 238 ng iba. Maganda ito sa still, pero mas madalas itong lumaktaw sa mga hakbang ng brief kaysa sa kahit sino, at ito ang pinakamataas ang presyo.

Ang mga iskor

Bawat clip ay nakakuha ng 1 hanggang 5 sa anim na criteria mula sa dalawang hiwalay na blind judge. Ang iskor ng trabaho ay ang average ng tatlong take nito, kaya mabibilang ang masamang take. Isinulat namin ang patakaran bago pa ang anumang render: ang agwat na mas mababa sa 0.3 ay tabla, at naka-orange ang pinakamataas na iskor at ang anumang kapantay nito.

Modelo	Nagsasalita	Produkto	Eksena	Kabuuan	Pinakamahinang take
Wan 3.0	4.17	3.67	3.72	3.85	3.29
Kling 4.0	3.95	3.86	3.66	3.82	3.62
Seedance 2.5	4.08	3.33	3.31	3.57	3.00
MiniMax H3	3.36	3.07	3.89	3.44	2.96
Veo 3.1	3.36	2.65	3.10	3.04	2.50

Ang kabuuan ay ang simpleng average ng 45 na iskor. Kapag tinanggal ang audio nang buo, bahagyang nalalampasan ng Kling ang Wan habang nananatili ang iba sa puwesto nila: Kling 4.0 3.88, Wan 3.0 3.85, Seedance 2.5 3.65, MiniMax H3 3.52, Veo 3.1 3.05. Nasa loob ng linya ng tabla ang Seedance 2.5 sa kabuuan, at ang pinakamahina nitong take ang dahilan kung bakit mas mababa ito sa dalawang nangunguna.

Trabaho 1: isang taong nagsasalita at nagbibilang gamit ang mga daliri

Nagsasabi ang panadero ng isang eksaktong linya, ipinapakita ang isa, dalawa, pagkatapos tatlong daliri, at tinatapik ang tinapay para lumipad ang harina. Sinusubok nito ang lip sync, mga salita, anatomya ng kamay, at maliit na pisika sa iisang naka-lock na shot. Panalo: tatlong-daan na tabla sa pagitan ng Wan 3.0, Seedance 2.5 at Kling 4.0. Sinabi ng lahat ng 15 clip ang linya ng eksaktong salita sa salita. Naghati ang grupo sa usapin ng mga daliri: hindi kailanman umabot sa tatlo ang Veo 3.1 sa anumang take, at buong kamay ang ipinakita ng MiniMax H3 sa dalawa sa tatlong take.

Take 1Take 2Take 3

I-play ang lima, naka-mute

Mga iskor sa pagsasalita

1. Wan 3.04.17
2. Seedance 2.54.08
3. Kling 4.03.95
4. MiniMax H33.36
5. Veo 3.13.36

Average ng tatlong take. Naka-orange ang unang puwesto o ang nasa loob ng 0.3 nito.

Trabaho 2: isang garapon ng pulot na may naka-print na label

Dahan-dahang push-in sa may label na garapon habang bumabagsak ang pulot mula sa sandok. Sinusubok nito ang maliit na teksto, kung nananatiling tahimik ang mga letra, at makapal na likido. Ang label ay sarili naming gawang produkto. Panalo: Kling 4.0, kasama ang Wan 3.0 sa loob ng linya ng tabla. Halos nalutas na ang baybay: 14 sa 15 clip ang nakakuha ng tamang tatlong linya. Hindi ganoon ang pulot. Halos walang tiklop o paikot na nangyari gaya ng hiniling, at hindi kailanman nag-push-in ang Veo 3.1 sa anumang take.

Take 1Take 2Take 3

I-play ang lima, naka-mute

Mga iskor sa produkto

1. Kling 4.03.86
2. Wan 3.03.67
3. Seedance 2.53.33
4. MiniMax H33.07
5. Veo 3.12.65

Average ng tatlong take. Naka-orange ang unang puwesto o ang nasa loob ng 0.3 nito.

Trabaho 3: pagsunod sa siklista sa maulan na palengke

Isang tuloy-tuloy na shot sa likod ng courier sa ulan: isang iwas sa paligid ng kariton, isang asong tumatawid, isang kalansing ng kampana, pagkatapos ay tumataas ang kamera sa ibabaw ng mga payong para ipakita ang buong kalye. Panalo: MiniMax H3, kasama ang Wan 3.0 at Kling 4.0 sa loob ng linya ng tabla. Walang nakakuha ng bawat beat. Pinakamalapit sa brief ang Kling at H3. Ang Wan ang pinakamatalas pero pinalitan ang payak na canvas ng makukulay na tent at gawang-gawang karatula. Paulit-ulit na pinuputol ng Seedance ang siklista, at nawala ng Veo ang siklista nang buo sa dalawa sa tatlong take.

Take 1Take 2Take 3

I-play ang lima, naka-mute

Mga iskor sa eksena

1. MiniMax H33.89
2. Wan 3.03.72
3. Kling 4.03.66
4. Seedance 2.53.31
5. Veo 3.13.10

Average ng tatlong take. Naka-orange ang unang puwesto o ang nasa loob ng 0.3 nito.

Ayon sa aming iniskor

Mga average sa lahat ng 45 clip. Ang detalye ang pinakakaunting naghihiwalay sa mga modelong ito. Ang pagsunod sa brief ang pinakamalaking naghihiwalay sa kanila, at malaking bahagi ng dahilan kung bakit ganito ang ranking.

Huling dumating ang pinakamahal na modelo

Mas mahal ng apat na beses bawat clip ang Veo 3.1 kaysa sa Wan 3.0 at pinakamababa ang iskor sa bawat trabaho, pantay sa MiniMax H3 sa usapin ng nagsasalita. Napakaganda ng mga still, na may pinakapaniniwalaang harina sa test. Ang problema ay ang mga hakbang: hindi ito nagbilang hanggang tatlo, hindi kailanman nag-push-in sa pulot, at nawala ang siklista sa ulan. Kung iisang simpleng ideya ang shot mo, baka mas makabuti ito kaysa sa ipinapakita ng test na ito. Sadyang puno ng hakbang ang mga prompt namin.

Naging madali ang baybay, hindi ang likido

Inasahan naming ang label ang magiging mahirap na bahagi. Labing-apat sa 15 clip ang nag-print nang tama ng tatlong linya, at isang take lang ng Veo ang nagdagdag ng maling salita. Ang pulot na tumitiklop at umiikot habang bumabagsak ay ang hindi nagawa ng halos lahat. Karamihan sa mga clip ay nagbigay ng tuwid na haligi, o matigas na hugis na halos hindi nagbago. Isang take ng Seedance 2.5 ang tama, at iyon din ang problema: isang take lang sa tatlo.

Tungkol sa musika sa ilang clip ng pulot

Humiling kami ng tunog ng kusina at walang musika. Sa 8 sa 15 clip ng pulot, halos walang tunog ang ibinalik ng modelo, mas mababa sa -40 dB, at nagdagdag ang aming render path ng tahimik na background na musika sa ilalim nito: ang lahat ng tatlong take ng MiniMax H3, dalawa ng Seedance 2.5, dalawa ng Veo 3.1 at isa ng Wan 3.0. Wala sa mga take ng Kling ang nangailangan nito. Musika namin iyon, hindi ng modelo. Nabawasan ang puntos ng mga clip na iyon sa tunog dahil sa katahimikan, at minarkahan sila sa raw table. Hindi ginalaw ang mga clip na nagsasalita at mga eksena.

Paano namin ito pinatakbo

Isinulat muna namin ang tatlong prompt at ang mga patakaran sa pag-iskor, nagpatakbo ng isang pilot clip para suriin ang pipeline, at saka nag-render ng iba. Kabilang sa 45 ang pilot. Walang pumalya, kaya walang inulit o inalis.

Humingi ang bawat clip ng 16:9 sa 720p at 6 na segundo, walang konteksto ng brand at walang muling pagsulat ng prompt. Dalawang eksepsiyon: tumakbo ang MiniMax H3 nang 8 segundo dahil wala itong 6-segundong opsyon, at naghahatid ito ng 1344x768 sa halip na 1280x720. Naghahatid ang Wan 3.0 ng 30 frame bawat segundo, ang iba ay 24.

Pareho ang sinasabi ng bawat trabaho sa bawat modelo, isinulat ayon sa rekomendasyon ng gumawa ng modelong iyon, dahil hindi nagiging patas ang test kung babasahin ng isang modelo ang diyalekto ng iba. Walang gabay ang Kling para sa 4.0, kaya ginamit namin ang mga gawi nito sa 3.0. Ang Veo 3.1 ang buong modelo, hindi ang mabilis na bersyon nito.

Nagpadala kami ng seed na 1101, 2202 at 3303 sa Veo 3.1, Wan 3.0 at MiniMax H3. Hindi nagpapasa ng seed ang aming render path sa Seedance 2.5 o Kling 4.0, kaya tatlong magkakahiwalay na run ang mga take na iyon. Walang modelong nagkumpirma ng seed na ginamit nito.

Blind ang pag-iskor at ginawa ng dalawang hiwalay na AI judge, hindi ng tao. Binawasan ang bawat clip sa 12 frame mula simula hanggang dulo, isang spectrogram ng tunog nito, ang lakas nito at isang transcript, sa ilalim ng random na code. Walang judge ang nakakita kung aling modelo ang gumawa ng clip hanggang sa maisulat ang kanilang mga iskor. Malapit ang pagkakasundo ng dalawang judge, na may correlation na 0.91 sa mga clip at average na agwat na 0.36 puntos bawat criteria. Ang mga iskor sa pahinang ito ay ang kanilang average, at nasa CSV ang numero ng bawat judge.

May ilang pagsusuri na hindi nangailangan ng paghuhusga. Kinumpirma ng lokal na transcript ang eksaktong linya sa lahat ng 15 clip na nagsasalita. Walang nakitang hiwa ang scene detection sa anumang clip, bagaman ang isang take ng Veo 3.1 ay nagpe-play bilang biglaang lundag at minarkahan namin ito. Parehong binilang nang hiwalay ng dalawang judge ang mga linya ng label at nagtugma sila sa lahat ng 15.

Ang mga prompt, salita sa salita

Kopyahin ang alinman sa mga ito. Binubuksan ng bawat link ang prompt na iyon sa modelong iyon sa Hermoso.

A baker speaks to camera in a small bakery kitchen and counts three ingredients on her fingers. She is a woman in her late thirties with short curly dark hair, flour on her forearms and a navy linen apron.

Shot 1 (0-6 seconds): Locked-off medium shot at eye level, 35mm lens, shallow depth of field. She stands behind a floured wooden counter with a round crusty loaf in front of her and looks straight into the lens. She raises her right hand and shows one finger, then two fingers, then three fingers, one finger for each ingredient she names. Then she lowers both hands and pats the top of the loaf twice, and a small puff of flour rises into the window light.

Dialogue: The baker says in a warm, dry voice: "Just three things. Flour. Water. Salt. That's the whole loaf." Accurate lip sync.

Sound: her voice and the soft room tone of the kitchen.

Style: realistic documentary, warm morning window light, one continuous shot.

Negative prompt: subtitles, captions, on-screen text, watermark, background music, cuts.

Medium shot at eye level, locked-off camera, one continuous shot, realistic documentary look, warm morning window light in a small bakery kitchen. A woman in her late thirties with short curly dark hair, flour on her forearms and a navy linen apron stands behind a floured wooden counter with a round crusty loaf in front of her and looks straight into the lens. As she speaks she raises her right hand and shows one finger, then two fingers, then three fingers, one finger for each ingredient she names. Then she lowers both hands and pats the top of the loaf twice, and a small puff of flour rises into the window light.

Baker (warm, dry voice): "Just three things. Flour. Water. Salt. That's the whole loaf."

Only her voice and the soft room tone of the kitchen are heard.

A baker speaks to camera in a small bakery kitchen and counts three ingredients on her fingers, realistic documentary style, locked-off medium shot at eye level, warm morning window light.

0-1 seconds: She stands behind a floured wooden counter with a round crusty loaf in front of her and looks straight into the lens.
1-4 seconds: She raises her right hand and shows one finger, then two fingers, then three fingers, one finger for each ingredient she names.
4-6 seconds: She lowers both hands and pats the top of the loaf twice, and a small puff of flour rises into the window light.

Dialogue: The baker says in a warm, dry voice: "Just three things. Flour. Water. Salt. That's the whole loaf."

Constant throughout: she is a woman in her late thirties with short curly dark hair, flour on her forearms and a navy linen apron. One continuous shot, the camera stays still. Sound: her voice and the soft room tone of the kitchen. No subtitles. No background music.

integrated_multimodal_description: [Shot 1] Live-action, realistic documentary style, a Static Shot at eye level holds a medium shot of a small bakery kitchen in warm morning window light. A woman in her late thirties with short curly dark hair, flour on her forearms and a navy linen apron (S1), on screen, with a warm, dry voice, stands behind a floured wooden counter with a round crusty loaf in front of her. She looks straight into the lens and says: <d>[English] Just three things. Flour. Water. Salt. That's the whole loaf.</d> As she speaks she raises her right hand and shows one finger, then two fingers, then three fingers, one finger for each ingredient she names. Then she lowers both hands and pats the top of the loaf twice, and a small puff of flour rises into the window light.

overall_soundscape: The soft room tone of a small kitchen. Her palms pat lightly on the bread crust.

non_diegetic_music: N/A

A medium shot at eye level with a locked-off camera in one continuous take. In a small bakery kitchen lit by warm morning window light, a woman in her late thirties with short curly dark hair, flour on her forearms and a navy linen apron stands behind a floured wooden counter with a round crusty loaf in front of her and looks straight into the lens. She raises her right hand and shows one finger, then two fingers, then three fingers, one finger for each ingredient she names. She then lowers both hands and pats the top of the loaf twice, and a small puff of flour rises into the window light. Realistic documentary style, 35mm lens, shallow depth of field.

The baker says in a warm, dry voice: "Just three things. Flour. Water. Salt. That's the whole loaf."

Audio: her voice and the soft room tone of the kitchen, and nothing else.

A macro product shot of an open jar of honey with a printed label while a ribbon of honey falls into it.

Shot 1 (0-6 seconds): Slow push-in, 100mm macro lens. The frame shows an open, squat, clear glass jar of golden honey on a dark slate counter in soft daylight from the left. A cream paper label on the front of the jar is printed in black with exactly three lines of text: "HERMOSO" in large serif capitals, "WILDFLOWER HONEY" in smaller capitals below it, and "NET WT 12 OZ" in small capitals at the bottom. A hand at the top of the frame holds a wooden honey dipper above the jar, and a thin ribbon of thick honey falls into the open jar, where it folds and coils on the surface before slowly sinking. The camera pushes in slowly toward the label until the three lines of text fill the lower third of the frame, sharp and legible, with the letters fixed in place on the glass.

Sound: a quiet kitchen with a kettle simmering softly and birdsong through an open window.

Style: photoreal commercial, soft daylight, one continuous shot.

Negative prompt: subtitles, captions, watermark, background music, extra text on the label, cuts.

Macro product shot, one continuous shot, photoreal commercial look. The frame shows an open, squat, clear glass jar of golden honey on a dark slate counter in soft daylight from the left. A cream paper label on the front of the jar is printed in black with exactly three lines of text: "HERMOSO" in large serif capitals, "WILDFLOWER HONEY" in smaller capitals below it, and "NET WT 12 OZ" in small capitals at the bottom. A hand at the top of the frame holds a wooden honey dipper above the jar, and a thin ribbon of thick honey falls into the open jar, where it folds and coils on the surface before slowly sinking. The camera pushes in slowly toward the label until the three lines of text fill the lower third of the frame, sharp and legible, with the letters fixed in place on the glass.

Sound: a quiet kitchen with a kettle simmering softly and birdsong through an open window.

A macro product shot of an open jar of honey with a printed label while a ribbon of honey falls into it, photoreal commercial style, slow push-in, soft daylight.

0-2 seconds: The frame shows an open, squat, clear glass jar of golden honey on a dark slate counter in soft daylight from the left. A cream paper label on the front of the jar is printed in black with exactly three lines of text: "HERMOSO" in large serif capitals, "WILDFLOWER HONEY" in smaller capitals below it, and "NET WT 12 OZ" in small capitals at the bottom.
2-4 seconds: A hand at the top of the frame holds a wooden honey dipper above the jar, and a thin ribbon of thick honey falls into the open jar, where it folds and coils on the surface before slowly sinking.
4-6 seconds: The camera keeps pushing in slowly toward the label until the three lines of text fill the lower third of the frame, sharp and legible.

Constant throughout: the label's letters stay fixed in place on the glass. One continuous shot. Sound: a quiet kitchen with a kettle simmering softly and birdsong through an open window. No subtitles. No background music.

integrated_multimodal_description: [Shot 1] Live-action, photoreal commercial style, a macro shot shows an open, squat, clear glass jar of golden honey on a dark slate counter in soft daylight from the left. A cream paper label on the front of the jar is printed in black with exactly three lines of text: "HERMOSO" in large serif capitals, "WILDFLOWER HONEY" in smaller capitals below it, and "NET WT 12 OZ" in small capitals at the bottom. A hand at the top of the frame holds a wooden honey dipper above the jar, and a thin ribbon of thick honey falls into the open jar, where it folds and coils on the surface before slowly sinking. The camera performs a Push In at slow speed toward the label until the three lines of text fill the lower third of the frame, sharp and legible, with the letters fixed in place on the glass.

overall_soundscape: A quiet kitchen with a kettle simmering softly. Birdsong drifts through an open window.

non_diegetic_music: N/A

A macro product shot in one continuous take with a slow push-in, 100mm macro lens. The frame shows an open, squat, clear glass jar of golden honey on a dark slate counter in soft daylight from the left. A cream paper label on the front of the jar is printed in black with exactly three lines of text: "HERMOSO" in large serif capitals, "WILDFLOWER HONEY" in smaller capitals below it, and "NET WT 12 OZ" in small capitals at the bottom. A hand at the top of the frame holds a wooden honey dipper above the jar, and a thin ribbon of thick honey falls into the open jar, where it folds and coils on the surface before slowly sinking. The camera pushes in slowly toward the label until the three lines of text fill the lower third of the frame, sharp and legible, with the letters fixed in place on the glass. Photoreal commercial style.

Audio: a quiet kitchen with a kettle simmering softly and birdsong through an open window, and nothing else.

A low tracking shot follows a bicycle courier through a crowded rainy night market and rises above the crowd to reveal the whole street.

Shot 1 (0-6 seconds): Wide 24mm lens. In a narrow night-market street at dusk in heavy rain, lined with food stalls under plain canvas awnings and strings of warm lanterns, a bicycle courier in a yellow rain jacket rides away from the camera while the camera follows behind at low height at the same speed. The rider swerves left around a vendor pushing a cart, then a small dog runs across the street in front of the bicycle and the rider rings the bell once. Steam rises from a noodle stall and the wet cobblestones reflect the lanterns. In the last two seconds the camera rises smoothly above a crowd of people with umbrellas and reveals the whole street stretching ahead, rain streaks catching the lantern light.

Sound: heavy rain on canvas, one bicycle bell, sizzling woks, crowd chatter.

Style: cinematic realistic, rainy dusk, one continuous shot.

Negative prompt: subtitles, captions, on-screen text, watermark, background music, cuts.

One continuous tracking shot, cinematic realistic look. In a narrow night-market street at dusk in heavy rain, lined with food stalls under plain canvas awnings and strings of warm lanterns, a bicycle courier in a yellow rain jacket rides away from the camera while the camera follows behind at low height at the same speed. The rider swerves left around a vendor pushing a cart, then a small dog runs across the street in front of the bicycle and the rider rings the bell once. Steam rises from a noodle stall and the wet cobblestones reflect the lanterns. In the last two seconds the camera rises smoothly above a crowd of people with umbrellas and reveals the whole street stretching ahead, rain streaks catching the lantern light.

Sound: heavy rain on canvas, one bicycle bell, sizzling woks, crowd chatter.

A low tracking shot follows a bicycle courier through a crowded rainy night market and rises above the crowd to reveal the whole street, cinematic realistic style, one continuous shot.

0-2 seconds: In a narrow night-market street at dusk in heavy rain, lined with food stalls under plain canvas awnings and strings of warm lanterns, a bicycle courier in a yellow rain jacket rides away from the camera. The camera follows behind at low height at the same speed. The rider swerves left around a vendor pushing a cart.
2-4 seconds: Ahead, a small dog runs across the street in front of the bicycle and the rider rings the bell once. Steam rises from a noodle stall and the wet cobblestones reflect the lanterns.
4-6 seconds: The camera rises smoothly above a crowd of people with umbrellas and reveals the whole street stretching ahead, rain streaks catching the lantern light.

Constant throughout: the same rider, the same rainy dusk light. Sound: heavy rain on canvas, one bicycle bell, sizzling woks, crowd chatter. No subtitles. No background music.

integrated_multimodal_description: [Shot 1] Live-action, cinematic realistic style, a Tracking Shot at low height follows a bicycle courier in a yellow rain jacket who rides away from the camera through a narrow night-market street at dusk in heavy rain, lined with food stalls under plain canvas awnings and strings of warm lanterns, with the camera moving at the same speed. The rider swerves left around a vendor pushing a cart, then a small dog runs across the street in front of the bicycle and the rider rings the bell once. Steam rises from a noodle stall and the wet cobblestones reflect the lanterns. Near the end the camera performs a Pedestal Up with large amplitude at slow speed, rising above a crowd of people with umbrellas and revealing the whole street stretching ahead, rain streaks catching the lantern light.

overall_soundscape: Heavy rain drums on canvas and one bicycle bell rings. Woks sizzle under the crowd's chatter.

non_diegetic_music: N/A

One continuous tracking shot with a wide 24mm lens, cinematic realistic style. In a narrow night-market street at dusk in heavy rain, lined with food stalls under plain canvas awnings and strings of warm lanterns, a bicycle courier in a yellow rain jacket rides away from the camera while the camera follows behind at low height at the same speed. The rider swerves left around a vendor pushing a cart, then a small dog runs across the street in front of the bicycle and the rider rings the bell once. Steam rises from a noodle stall and the wet cobblestones reflect the lanterns. In the last two seconds the camera rises smoothly above a crowd of people with umbrellas and reveals the whole street stretching ahead, rain streaks catching the lantern light.

Audio: heavy rain on canvas, one bicycle bell, sizzling woks and crowd chatter, and nothing else.

Ang hindi masasabi sa iyo ng test na ito

Hindi sakop ng tatlong prompt ang lahat, at ang tatlong take bawat trabaho ay maliit na sample, kaya ituring na walang agwat ang pagkakaiba na 0.3 o mas mababa. Ang dalawang judge ay mga AI model na bumabasa ng frame at sukat. Walang nakinig sa mga clip na ito gamit ang tainga: ang tunog ay hinusgahan mula sa transcript, spectrogram at lakas ng tunog, kaya ang kalidad ng boses ang pinakamahinang bahagi ng aming pag-iskor, at maaaring makaligtaan ng 12 frame ang glitch sa pagitan nila. Magaspang ang aming sukatan ng lip sync at hindi namin ito lubos na pinagkatiwalaan. Pinakamalapit dito ang Seedance 2.5 at Kling 4.0, at hanggang kalahating segundo ang layo ng Wan 3.0, kaya panoorin ang mga clip kung mahalaga sa iyo ang lip sync. May isang kilalang butas ang blind na pagsubok: ang 8 segundong haba ng MiniMax H3 ay nagpapakilala sa mga clip nito, at maaaring nakatulong sa kanya ang dagdag na dalawang segundo sa mabagal na reveal. Hindi nasubukan ang anumang wala sa talahanayan, kabilang ang Seedance 2.0 at HeyGen Video. Mabilis magbago ang mga modelo at presyo, at galing sa iisang araw ang mga resultang ito.