新智元报道
这一周的AI视频圈，简直是神仙打架。
10月15日晚，可灵4.0全量上线，重点补强了大幅度动作和复杂运镜，单次原生生成最长拉到30秒。
全能参考也同步升级，还能组合最多15项参考素材、用10张关键帧控制画面，让人物、场景和镜头的安排更可控。
这里面最突出的特点，就是打戏这次补足了，来看效果。
另一边，有人把目标放在了长片上：给一句提示词，生成最长10分钟的完整成片。
10月21日凌晨，YC孵化的VideoGen发布首个视频模型Model 1。从剧本、对白到分镜，都交给AI规划，再生成、剪接成片；同时拿出100万美元免费额度，支持入选的创作者项目。
同一天，火山引擎上线Seedance Managed Agent，瞄准短剧里剧本跑偏、剧情混乱的问题。按官方数据，剧情混乱问题减少了约54%。
当晚，ElevenLabs旗下的ElevenCreative上线Chimera，直接把Seedance 2.5和自家的Eleven v4并称为「全球领先的两个模型」：Eleven v4把情绪放进声音，Seedance 2.5把它带到脸上。
30秒原生生成、10分钟成片、剧本不跑偏……各家的目标越来越清楚：让AI从拍出一个好镜头，走向拍完一个好故事。
可一整场戏交到AI手里，它到底能不能照着剧本拍？
我们用相同的提示词和参考素材，对可灵4.0、Seedance 2.5、MiniMax H3 Max和万相3.0做了一轮实测。
第一个测试的，是参考生功能，这也是目前创作者使用模型最主要的方式。
我们给可灵4.0、Seedance 2.5、MiniMax H3 Max和万相3.0用一张底图作为参考，拍一场悟空和黛玉的对手戏。
最难的，就是把两人之间的暧昧感拍出来。
剧情设定
黛玉坐在池边安静垂钓，悟空嫌慢，用金箍棒轻点水面，把一条鲤鱼托进木桶，转头得意邀功。黛玉先看鱼，再抬眼，停半拍才笑着说：「我钓的是清静，不是鱼。」悟空收起得意，回一句「那俺陪你」，放下棒子，坐到她身边。
可灵4.0的表现让人惊喜，把林妹妹的小心思拍出来了。美中不足的是，道具稍微有一点bug：大圣手中凭空变出一根鱼竿。
MiniMax H3 Max和万相3.0也都把剧情拍了出来，就是两个人之间少了点暧昧的火花。
还得是Seedance 2.5，CP感一下就来了。黛玉托腮抿笑，悟空猴脸靠近，张力十足，两个人之间的小情调演活了。
在这个case里可以看出，各家在参考图片生视频的执行上都不错。但在情绪上，Seedance 2.5的人物情绪刻画更更鲜活一点。
今年视频模型的真人感已经越来越强了，模型听不听话反而成了创作者的刚需。
我们用4个题来综合看看模型听不听话到底体现在什么地方。
文生视频，最能体现一个模型的底子。
这次我们不给首帧，只给文字，让可灵4.0、Seedance 2.5和万相3.0生成这个情节——
在一座美术馆中，画中人溜出来，把值夜班的保安哄进画里替班。注意，穿越、变身是最容易穿帮的。
剧情设定
深夜闭馆，画里的绿裙女子偷看打瞌睡的保安，提裙跨过低低的画框，油画笔触随身体走出逐渐变成真人皮肤和布料。她叫醒保安，牵着他说：「替我站一会儿。」保安跨进同一幅画，转身坐上画内的红椅。彻底成为油画的保安委屈地问：「……一会儿是多久？」
可灵4.0的动作刻画非常利落，但女孩出画时有点穿帮，保安进后，也还保持着真人脸，有些贴片感。
万相3.0的油画笔触和浓郁色彩非常显眼，真人感也很强。可惜进画时卡了壳，保安半个身子还凸在画框外。
Seedance 2.5 执行得最好：保安一坐下，脸和制服瞬间染上厚涂油画的质感，保安和画融为一体。美中不足的是，进画时多迈了一次腿，动作不够利落。
整体来说，Seedance 的指令遵循特别好，对于文生视频这种考验模型极限能力的场景，它的表现比较优秀，30s的场景可用性最强。
下面这场戏，考察的是场景的指令遵循。
Prompt要求：戏台下只留下一个老人。
剧情设定
冬日傍晚，大雪一直下，荒野里一座孤零零的戏台，台下正中央只坐着一个老人。虞姬先独自唱完「看大王在帐中和衣睡稳，我这里出帐外且散愁情」，锣鼓过门后，霸王才从舞台左侧登场，接唱「今日里败阵归心神不定」。
Seedance 2.5按提示词要求，台下只生成一个老人，寂寥的氛围感立刻就有了。
可灵4.0这边，台下乌泱泱坐了好几排观众，一下变成了乡村大舞台，氛围感不太对了。
场景遵循这一场，Seedance 2.5胜出。
接下来的考题，让模型参考一段「画中人活过来」的视频，用文字指定五幅名画，生成展厅里的传扇接力。
这道考题，细节实在太多了。仕女要雍容，醉仙要疏狂，牵马人要利落；同一把圆扇还得按顺序完成四次交接。
哪个模型能成功完成？
剧情设定
五幅名画在展厅并排挂好，仕女将一把圆形纨扇抛出画外，醉仙接住后用陕西话抱怨「谁嘛，搅了额的美梦！」，再传给牵马人。牵马人翻腕甩出，渔翁接棒，提醒一句「可不要吓到我的鱼儿」，最后交给琴师，搁在琴案上收住。
一把扇子、五个人、四次交接，每一棒都要拍清抛出和接住。人物动起来以后，仕女仍是工笔重彩，醉仙仍是泼墨，牵马人保留白描线条，渔翁周围的大块留白也不能消失。接力要连贯，各幅画又不能被揉成同一种画风。
可灵4.0的场面很丝滑流畅，但复杂指令遵循不够好：圆扇变成了折扇，还漏掉了第二幅，传到第三幅只见马不见人。
Seedance 2.5执行得好一点，始终保持同一把圆扇、四次抛接。不过，它也出现了一些小bug，比如在第三棒把牵马人改成了骑马人。
总的来说，Seedance 2.5准确完成了四次交接，在复杂指令遵行上更优。
剧情一长，镜头就得跟着人物的处境变化。各家模型的镜头语言怎么样？
为此，我们把冲突放进赌场：黄毛掀桌闹事，十三娘从含笑应付到被激怒，再冷笑着把人拉进赌局。机位和表情，都得把这股火药味拍出来。
剧情设定
给可灵4.0和Seedance 2.5同一张原创旗袍人像、同一段合成读文音频，按同一份提示词拍30秒。黄毛掀翻赌桌，十三娘掀开珠帘，先从容应对；遭到调戏后突然发怒，再迅速恢复笑容，拿起骰盅，让对方赌输了就跪着捡筹码。
可灵4.0的人物和旗袍从头到尾都很稳，画面精致、颜色不失真，但在掀桌等细节上稍微有点小bug。另外，十三娘大多正对镜头，机位变化不多。
在镜头语言上，Seedance 2.5的表现力非常生动。黄毛掀桌的瞬间，镜头便压到贴近地面，筹码和骰子瞬间飞起，再以慢动作砸落。这是电影级别的运镜。
不过，2.5中段有个镜头从中景猛地推到人物脸上，有些短剧味。
在镜头语言上，两家各有千秋。质感上，Seedance 2.5的更偏电影感，可灵4.0的画面有些油蜡感。
龙舟赛的考题，是对画面物理还原极大的考验，划桨的动作、身体的起伏必须和水面对得上。
剧情设定
多条龙舟在江面并排竞渡，每船二十余名划手跟着鼓点同步前俯后仰、挥桨入水。船头鼓手高举双槌、击鼓喊号，船尾舵手掌舵，两岸观众摇旗助威。
镜头从低机位侧面跟船开始，依次给到桨叶破水、手臂绷紧、汗水滑过脸庞和鼓手发力的近景；水花要有慢动作细节，最后升到空中拍群舟冲线。
可灵4.0的整体完成度不错，但是仔细看，桨入水时仿佛没有阻力，好像在空气中划桨一样，这种细节刻画得不够精细。
Seedance 2.5展现了极佳的表现力。
它直接把镜头压到船舷边，人物探身、桨入水，桨叶的运动和水的反应很自然；甚至还给了划手肌肉绷紧、满脸汗水的特写，刻画出了力量感，物理表现力上更强一点。
山羊这段，我们试的是视频续拍。
我们把一段山羊失足跌落的动画交给模型，让它接着往下演。
剧情设定
承接参考视频里踩松石的瞬间，山羊失足坠落，一路撞上山石、弹起、翻滚，皮毛沾满灰尘。镜头跟着它下到湖边，斑马惊叫着散开。
山羊随后向画面右侧弹起，落地再滚一段，最后四脚朝天地停在另一只斑马腿边。后者岿然不动，只动动耳朵。一镜连续跟拍、声音顺接。
一眼看上去，Seedance 2.5的羊非常Q弹，一滚一滚地太可爱了，甚至最后还打起来水漂，物理规律表现力非常生动。
可灵4.0的山羊滚下来的画面，缺少一些弹性，不过整体来说也完成了一段很精彩的视频。
几个极限测试扒下来，结论很清晰了。
Seedance2.0之后，参考生成成为创作者使用模型最主要的方式，各家模型也开始比拼全能参考的能力。在参考图片生视频上，这几家模型都还算听话，一致性都不错。
在文生视频和复杂指令上，交给Seedance2.5最放心。Seedance 2.5把动作、物理规律、复杂叙事和运镜处理得更细腻，主打一个听指挥。
可灵4.0在打戏上有了进步，有了Seedance2.0打斗的感觉，同时画面色彩饱和度也会更高。
奥斯卡提名编剧Nicholas Kazan的电影《Cortés》已经开机
现在国内大家都开始卷长剧和IP，创作者更在意的，是写下的动作、情绪和转折，能否稳定地留在成片里。模型就更要记住：主角长什么样，故事走到了哪一步。
这就是为什么我们这篇评测的核心。听不听话、可不可控，已经成为AI视频下一步的核心关键。
谁能听懂导演，把一场戏从头到尾拍完整，谁就能拿到下一场战役的入场券。
编辑：摩西 Aeneas
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
This week's AI video ring, it's a fairy fight.
On the evening of 15 October, Kling4.0 was fully online, with a focus on the reinforcement of large movements and complex mirrors, with single lifetime generation to a maximum of 30 seconds.
Full reference is also upgraded at the same time, and up to 15 reference assets can be combined, with 10 key frames to control the images and make the arrangements for characters, scenes and shots more manageable.
The most striking feature of this is the play-up, which is done.
On the other side, the target was placed on a long film: a prompt to produce a complete piece of up to 10 minutes.
In the early morning hours of 21 October, the first video model, Model 1, of Vidio GenZ94Z, hatched by YC. From scripts, pairs to storyboardes, they are submitted to AI for planning, generating and cutting films, and one million US$ free quotas are available to support selected creators.
On the same day, Volcengine went online as Feedance Managed Agent, aiming at the problem of the running of the script and the confusion of the drama. According to official data, the problem of disorder has decreased by about 54 per cent.
That night, the ElevenCreative on-line Chimera, under the banner of ElevenLabbs, directly named Seedance 2.5 and his home, Eleven v4, as "two models leading the world": Eleven v4 put emotions in the voice and Seedance 2.5 brought them to his face.
30 seconds of original creation, 10 minutes of film, and the scripts don't miss each other... The goal is becoming clearer: let AI go from making a good shot to making a good story.
But if the whole scene is in the hands of AI, can it be filmed in the script?
With the same prompt and reference asset, we did a test round of Kling4.0, Seedance 2.5, Mini Max H3 Max and Man against 3.0.
The first test is the reference biofunctional function, which is currently the most important way for creators to use models.
We'll use a background map for Kling4.0, Seedance 2.5, Mini Max H3 Max and Mansion 3.0 as a reference for an anti-Goku and Daiyu.
The hardest thing is to get a picture of the confusion between the two.
Story set.
Dai Yu sits by the pool in a quiet, slow, softer water with a golden stick, puts a carp into a barrel, turns around and takes credit. She looked at the fish, then looked up, and stopped and laughed, "I was fishing in peace, not fish." "The Goku takes his pride and says, "I'll stay with you." Put down the stick and sit next to her.
The performance of Kling4.0 was surprising and captured the care of Lin's sister. It's not enough that the props have a little bit of bug: the Great Sage made a pole out of empty hands.
Mini Max H3 Max and Mansion 3.0 also filmed the scene, which was a little bit of a spark between them.
It's gonna be Seedance 2.5, and CP feels it. Dai Yuto smiles, Goku faces are close, he's full of tension, and he's playing with two people.
As can be seen in this case, families are doing well with the reference video. But emotionally, the two-and-a-half characters of Seedance are much more alive.
The reality of the video model has grown stronger this year, and the model has become a demand for creators instead of listening.
We put four questions together to see where the model actually works.
text-to-video, which best reflects the foundations of a model.
This time, we don't give first frame, just text, so Kling4.0, Seedance 2.5, and 10,000-0.0...
In an art gallery, the person in the painting snuck out and put security guards on night shifts in the drawings. Attention, crossing and changing is the easiest to cross.
Story set.
Late at night, the woman in the green dress peeked at the slumbering security guards, the skirt crossed the low frame, and the brush came out of the body and gradually turned into skin and fabric. She woke up the guard and took him and said, "Stand me for a moment. "The security guards crossed into the same painting and turned to the red chair in it. The security committee, which became the painting, asked, "...how long?"
The Kling4.0 action was very good, but the girl was a bit of a fang when she was drawing, and when the security guard came, she also had a real face and a bit of a sticker.
The brush touch and the heavy color of the painting, which is 3.0, are very visible, and the real feeling is very strong. It's a pity that it's stuck in the paint and the security guard's body is stuck outside the frame.
Seedance 2.5 is the best: the security guard sits down, the face and the uniform suddenly get the touch of a thick painting, the security guard and the painting fit together. What's not good enough is the fact that the drawings were made with extra legs and not very good moves.
Overall, the Seedance directive is particularly good, with text-to-video being the best performance and 30s being the most available.
The next scene is the scene's instructions.
Prompt demands that only one old man remain under the stage.
Story set.
In the evening of winter, the snow continued, and there was a single stage in the wilderness, with only one elderly person sitting in the centre. Yu Ji sing "Looking at the King's books and clothes, I'm out of the house and I'm sad," and when he comes through the door, he comes out from the left side of the stage and sings "Today's Day is a day of defeat."
At the request of prompt, Seidance 2.5 produces only one elderly person below the stage, and the atmosphere of silence is immediately there.
On this side of Kling4.0, there were several rows of viewers sitting below the stage, turning into a big country stage, and the atmosphere was not quite right.
The scene follows this one. Seedance 2.5 wins.
The next question is, let the model refer to a video called "The Live of the Painter" and assign five paintings to the text to generate a fan relay in the gallery.
There's too much detail on this test. The bridesmaids must have a face, the drunk fairies must be deranged and the horse-makers must be profited, and the same fan must be handed over four times in order.
Which model will succeed?
Story set.
Five paintings were hung in rows in the gallery, and the lady threw a round fan out of the picture, and the drunk fairy picked up and complained in squawk, "Who is it, messing with the beauty of the dream!" and passed them on to the horse-to-man. "Don't scare my fish" when the horse-thrower flips out of his wrist and picks up the stick, and then puts it to the violinist and holds it in the piano case.
One fan, five people, four encounters, each bat to throw out and catch. When the person moves, the woman is still a choreographer, the drunk fairy is still ink, the horseman keeps a white line, and the large pieces around the fisherman cannot disappear. The relay must be consistent, and the paintings must not become the same.
The scene of Kling4.0 was smooth, but complex instructions were not well followed: the fan was turned into a fan, and the second one was missing, and the third one was circulated only to the horses.
Seedance 2.5 works better and keeps the same fan, four times. However, it has also seen the emergence of small bugs, such as the conversion of the horse taker to a horse rider in the third bat.
Overall, Seedance 2.5 accurately completed four handovers and improved compliance with complex directives.
As long as it lasts, shot has to follow the situation of the person. What about the shot language of the models?
For that reason, we put the conflict in the casino: the yellow hairs made a mess, from laughing at the table to being pissed off and laughing and pulling people into the game. The seat and the expression, they're gonna have to film the powder.
Story set.
Kling4.0 and Seedance 2.5 were given the same original flag figure, the same synthesizing reading audio for 30 seconds on the same prompt. When the yellow hair turns up the table, the mother lifts the beads, first reacts; then suddenly gets angry after being molested, then smiles quickly, picks up the dice and lets each other lose their bets and kneels to pick up the chips.
Kling4.0's character and flag robe were steady from beginning to end, with a fine picture and a fine colour, but slightly smaller bugs in detail, such as lifting tables. Moreover, most of the 13th mother is facing shot, and there have been few changes in her flight space.
In the shot language, the performance of Feedance 2.5 is very dynamic. At the moment when the yellow hair table was lifted, shot was pressed close to the ground, the chips and dice went up and fell in slow motion. It's a movie level mirror.
However, a shot was pushed from the center of the scene to the face of the person in the middle of the 2.5, which was somewhat short play.
In the language of shot, each family has a thousand swings. Kling4.0 image has some oil wax.
The test of the Dragon Canoe is a tremendous test of the physical restoration of the image, the movement of the paddles, the rise and fall of the body, and the water.
Story set.
Many canoes competed across the river, with more than 20 rowers per boat synchronizing with their drums, leaning down and swinging into water. The bow drummers raised their hammers, shouted their drums, the tail rudder took the wheel, and the viewers on both sides waved their flags.
shot begins on the side of the lower plane and on the boat, and then gives the near view of the oars broken, the arms tightened, the sweat slid across the face and the drumming hand; the water flower needs to move slowly in detail and eventually rises to the air to shoot the canoes.
The overall completion of Kling4.0 was good, but the details were not detailed enough to look at it as if it was without resistance and as if it were in the air.
Seedance 2.5 shows excellent performance.
It pushes shot directly to the side of the ship, and the person goes through the body, pushes into the water, and the motion of the oars and water reacts naturally; it even gives hand-stamped muscles and sweaty features, drawing a sense of power and more physical expression.
The goat part, we're going to try the video.
We hand over to the model an animation of the goat's fall and let it go down.
Story set.
As soon as the pines were stepped on the reference video, the goats fell and hit the rocks, bounced, rolled and covered with dust. shot followed it down to the lake and the zebra screamed.
The goat then came up to the right side of the picture, fell down and rolled over the ground, and the last four feet stopped at the foot of the other zebra. The latter are not moving, but moving their ears. One shot and the sound.
At first glance, Seedance 2.5's sheep are so Q-balled, so cute, and they even end up beating, and the physical pattern is very lively.
The image of the goat coming down from Kling4.0 lacks some elasticity, although, overall, it also produced a wonderful video.
A couple of end-of-the-art tests were pulled down and the conclusion was clear.
After Seedance 2.0, reference generation became the predominant way for creators to use models, and models began to compare with the ability to use full-fledged references. And on the reference video, these models are still good and consistent.
On text-to-video and the complex instructions, the best assurance is given to Feedance 2.5. Seedance 2.5 handles actions, physical patterns, complex narratives and mirrors more carefully, leading to a command.
Kling4.0 has made progress in the play, with the feeling of a Seedance 2.0 fight and a higher colour saturation.
Oscar nominated writer Nicholas Kazan.
Now everyone in the country has started to roll the drama and the IP, and the creators are more interested in writing the moves, emotions and turns that can stay in the film. The model is more important to remember: what the protagonist looks like and what the story is about.
That's why this is the core of our assessment. Listening, uncontrollable, has become central to AI video next step.
Anyone who understands the director and takes the whole scene from the beginning to the end can get tickets to the next battle.
Editor: Moses Aeneas
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.