Bernini-R 14B / OmniVideo2-A14B · 固定 caption 对比

Bernini-R 14B · T2V 12/12 · OmniVideo2-A14B · T2V 12/12 · Bernini-R 14B · RV2V 12/12 · OmniVideo2-A14B · V2V 12/12

T2V 使用相同的最终目标描述、seed、81 帧、25 fps、40 步;Omni 的 Qwen 自由扩写已关闭,只提取既定文本的特征。Bernini 保留原生任务前缀。相同 caption 正文不代表不同模型的特征或采样器相同。

编辑分支同样使用该目标描述与相同源视频。Bernini RV2V 另取源视频第 0/27/53/80 帧作原生参考;Omni 保留原生视觉条件与双份文本上下文。Real target 仅供观察,完整目标片段已统一到 3.24 秒以同步播放。

比较条件与原生差异

共同条件:相同最终 caption 正文、逐例 seed、81 帧、25 fps、40 步;编辑使用同一源视频。T2V 均为 832×480。

Bernini 保留固定任务前缀和原生 RV2V 参考帧;编辑输出为 832×464。Omni 保留原生源视觉条件和双文本上下文;编辑输出为 832×480。两边引导方式不同,UniPC 有效 flow shift 分别为 3 和 5。本页控制 caption 一致,其余原生差异如实保留。

所有结果原样展示。视频文件有效不代表目标动作成功。

12 例固定输入与 caption 清单
1.3B 旧 caption 流程对比本批运行记录

提杠铃 · Barbell lift

Editing instruction:Edit the action so that the woman in a home gym lifts a barbell from the floor to her waist.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman in a home gym lifts a barbell from the floor to her waist. She wears black activewear. The home gym has a tiled floor and exercise equipment. A static full-body shot keeps the woman and the barbell visible.

fb7725c351ba · seed 2026081901 · caption SHA-256: b3337cc9b01e086c7969770d68fc6c6572e454e5fe1be64f770a6b7e1e30500d

狗转头 · Dog head turn

Editing instruction:Edit the action so that the dog turns its head further to the right while sitting in the grass.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The dog turns its head further to the right while sitting in the grass. The dog is a reddish-brown and white Pembroke Welsh Corgi in a sunny green field. The camera is stationary in a medium shot.

3fadbe0c2961 · seed 2026081902 · caption SHA-256: aaeb58cbd788e3a815962e6289b1853ca09f2ef13c69d98a1af484e67f9da178

长椅上写字 · Write in notebook

Editing instruction:Edit the action so that the young woman sitting on a park bench thoughtfully writes in her notebook with a pen.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The young woman sitting on a park bench thoughtfully writes in her notebook with a pen. She has shoulder-length light brown hair and wears a pink shirt and blue jeans. The wooden bench is in a sunny park. The camera is stationary.

7077cfb160cf · seed 2026081904 · caption SHA-256: 0acdffedc18b04d2c2bef7d0a4e4e51544b75d4dd1f904e080d9e0490be24fb1

转身倚车 · Turn and lean on car

Editing instruction:Edit the action so that the woman in a black leather jacket and red pants turns and leans back against the hood of the black car.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman in a black leather jacket and red pants turns and leans back against the hood of the black car. The setting is an underground parking garage with steady red lighting. The camera is stationary.

7e488525099e · seed 2026081906 · caption SHA-256: ed6115b291eb39710054250e5a6fd3e51f92ed3675fe1af1c5b45617396c3082

闭眼尝汤 · Taste from ladle

Editing instruction:Edit the action so that the woman in an apron closes her eyes as she tastes from a metal ladle.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman in an apron closes her eyes as she tastes from a metal ladle. She wears a patterned apron in a modern kitchen. The camera is stationary.

d5221a1b69d5 · seed 2026081907 · caption SHA-256: 043306a7c3eff9ad7a50db5bbf0d3b4a2d6dec27fb8c5b5f6a316382992bc71b

放下手机后翻身睡觉 · Put phone away and sleep

Editing instruction:Edit the action so that the man in bed puts his smartphone on the nightstand and turns over to go to sleep.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The man in bed puts his smartphone on the nightstand and turns over to go to sleep. He has dyed-blond hair, wears a pink tank top and a nose ring, and is in a dark bedroom with a nightstand beside the bed. The camera is stationary.

40712e1341dc · seed 2026081911 · caption SHA-256: 24857d19e619899acd653ea104c6a5b68b895cdd4b5cc3637df661f694521907

Jellyfish contraction

Editing instruction:Edit the action so that the jellyfish pulsates, its bell contracting to reveal a ring-like shape in the blue water.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The jellyfish pulsates, its bell contracting to reveal a ring-like shape in the blue water. A small, translucent jellyfish with white spots is surrounded by deep blue water. The underwater camera is stationary.

10af6c651dbc · seed 2026081910 · caption SHA-256: 790782c54c3fea6d88d28f9129501ef060ebc863cd6ead101f7a12563d43e383

Playful selfie

Editing instruction:Edit the action so that the woman in a red bikini top and straw hat takes a playful selfie with her phone, sticking her tongue out and winking.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman in a red bikini top and straw hat takes a playful selfie with her phone, sticking her tongue out and winking. She is framed in a static medium shot against a uniform green screen background.

7fa338474c75 · seed 2026081900 · caption SHA-256: f7d22ce2d98ed6adff42d53192e7204016d9035165349de21317c756823ac96e

Scarf spin

Editing instruction:Edit the action so that the woman wearing a green skirt and black shirt gracefully spins in a circle while holding a scarf.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman wearing a green skirt and black shirt gracefully spins in a circle while holding a scarf. Her scarf is red and white. A flower-lined fence and a rustic wooden house form the background. The camera is stationary.

81533c9e56ec · seed 2026081915 · caption SHA-256: f6e77c2c4c9a431ed92369c65484776cb036d47a4d72c6a19e11d2a4d5b09579

Rub eyes then cover face

Editing instruction:Edit the action so that the man lying on the couch rubs his tired eyes and covers his face with both hands.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The man lying on the couch rubs his tired eyes and covers his face with both hands. He has dark skin, dreadlocks, and a gold septum ring. The couch is dark and the room is dimly lit. The camera is stationary.

d977ce147d80 · seed 2026081905 · caption SHA-256: a217f2a0d32d07e58d5c5781d12d401ed7a08f25d9286ea9f0bee5c25677e960

Place bottle on treadmill

Editing instruction:Edit the action so that the woman turns her head to the left and places her water bottle onto the treadmill.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman turns her head to the left and places her water bottle onto the treadmill. She has long reddish-blonde hair and wears a mint green athletic outfit. The modern gym has blue and white striped pillars. The camera is stationary.

840b214afead · seed 2026081913 · caption SHA-256: fe3b668a66485813f97192ae443c497bc34efcc6edbd10bbdf8af1724af85247

Bow to clap

Editing instruction:Edit the action so that the woman in a taekwondo uniform stands in profile and claps her hands together in front of her.

Source
Real target
Bernini-R 14B · T2V
实际文本与运行记录
OmniVideo2-A14B · T2V
实际文本与运行记录
Bernini-R 14B · RV2V
实际文本与运行记录
OmniVideo2-A14B · V2V
实际文本与运行记录

两模型共用的最终 caption 正文(无扩写):
The woman in a taekwondo uniform stands in profile and claps her hands together in front of her. Her uniform is white with a black collar. The background is black with side lighting, in a static medium shot.

d908552fb691 · seed 2026081903 · caption SHA-256: 025f04cb7e601915763666a56d60e01f88ed30d6efa85a172c0cf38eff834896