我把同一个万圣节打斗场景给了 Kling 4.0、Seedance 2.5 和 Wan 3.0

TL;DR

• 相同的 11 镜头提示词，相同的六张参考图，相同的设置，每个模型生成两次。
• Seedance 2.5 得分最高（8.06 / 10），但每次 30 秒生成成本为 $17.76。
• Kling 4.0 以每次 $6.41 排名第二（7.56），是本次测试中性价比最高的，而且对我的角色设计最忠实。
• Wan 3.0 以每次 $7.56 排名第三（7.44），风暴很棒、角色稳定，但打斗最弱。它仅落后 Kling 0.12，在我的误差范围内。
• 打斗场景仍然是 AI 视频最难的部分：三个模型在打斗冲击力上平均 6.5 / 10。

我为什么做这个测试

Kling 4.0 于 10 月 15 日发布，并许下了一个很大的承诺：单次生成最长 30 秒、多镜头叙事、多参考图和原生音频。规格表很容易写。我想知道它在面对一个高难度任务时实际表现如何：多个镜头、两个必须保持角色一致的角色、一场打斗、对白、水体物理和一个笑点。

于是我写了一个万圣节短片，然后原封不动地交给 Kling 4.0，以及它目前两个最强竞争对手 Seedance 2.5 和 Wan 3.0。下面是你自己运行同样测试所需的一切：素材、完整提示词、设置和评分表。

挑战：“地图上的最后一栋房子”

万圣节之夜。Pal 是一只派对独角兽吉祥物，它去地图上的最后一栋房子讨糖：一座建在海中岩石上的灯塔。开门的是海洋之王守灯人，一个脾气暴躁、管理灯塔的海神。他并不想施舍。他们从门口一路打到螺旋楼梯，再打到顶部的灯室，在那里，一道比灯塔还高的疯狗浪把他们俩吞没。水退去后，海洋之王叹了口气，把一条沙丁鱼丢进 Pal 的糖果桶里。

• 时长：30 秒
• 镜头：11 个，硬切
• 对白：两句英文（“Trick or treat?” 和 “...Trick.”）
• 它考验什么：多镜头连贯性、角色一致性、打斗编排、水体物理、口型同步和喜剧节奏

我是故意选打斗的。这是 AI 视频最容易露馅的场景类型：两个身体接触、快速运动、喜欢融化的武器，以及会在镜头切换之间漂移的角色。

素材

每个模型都拿到相同的六张参考图，并按相同顺序上传：四个地点和两个角色。

六张参考图。上排：灯塔外观、入口、螺旋楼梯。下排：灯室、海洋之王守灯人、Pal。

设计原创角色

两个角色都是原创的。Pal 是一个诡异又可爱的 1990 年代儿童电视吉祥物，被设定成一个穿着毛绒玩偶服的真人，而不是卡通形象，因为那种“恐怖谷实拍玩偶服”的感觉在镜头里更吓人。海洋之王守灯人借鉴了希腊神话中的波塞冬，这是公有领域内容，我刻意让他远离任何漫画式海王：他六十多岁，银发，在青铜盔甲外穿着黄色渔夫雨衣。如果你也发布这类测试，原创角色能让你避开版权和肖像权问题。

下面是我使用的角色设定图提示词。先生成正视图，挑出最好的一张，然后把它作为其他视图的参考。

Pal：转身设定图

plaintextCopy
1Character turnaround sheet of "Pal the Party Unicorn", a full-body practical mascot costume character from a 1990s children's TV show, photographed for real, not cartoon, not CGI. Faded mint-green plush fur body with a cream-colored belly, sherbet-colored mane and tail in orange, pink and yellow, a gold-and-white spiral striped horn like a candy cane. Wearing a gold sequin tailcoat jacket, a red polka-dot bow tie, and a round pin badge on the chest reading "PAL". Oversized glossy black plastic eyes, a stitched felt smile. Holding an orange plastic jack-o'-lantern candy bucket.
2Three views side by side: front view, side view, back view, full body, same pose, neutral standing.
3Plain light grey studio background, soft even lighting, sharp focus, photorealistic fabric texture, visible plush fibers and sequin reflections, 16:9, character design reference sheet, consistent proportions across all views.
4
5Negative: cartoon, anime, 3D render, CGI, plastic toy, purple dinosaur, pink fur, human face, extra limbs, extra horns, deformed hands, blurry, watermark, text artifacts

海洋之王守灯人：转身设定图

plaintextCopy
1Character turnaround sheet of "The Sea King Keeper", an original character inspired by the Greek sea god Poseidon, now a grumpy lighthouse keeper. A towering, broad-shouldered man in his 60s, weathered sun-darkened skin with deep wrinkles, short silver-white hair, a long grey-green beard braided with seaweed and tiny shells, pale sea-green eyes, permanently furrowed brow. Wearing a worn yellow fisherman's oilskin raincoat, open to reveal barnacle-crusted bronze scale armor with green patina underneath. Holding a rusted bronze trident covered in barnacles and kelp, with a small oil lantern hanging near the prongs.
2Three views side by side: front view, side view, back view, full body, same neutral standing pose.
3Plain light grey studio background, soft even lighting, photorealistic, cinematic character design reference sheet, detailed textures, 16:9, consistent proportions across all views.
4
5Negative: long black hair, tattoos, young muscular superhero, orange and green scale suit, golden crown, comic book style, cartoon, anime, CGI, deformed hands, extra fingers, blurry, watermark

身高对比，让打斗中两个角色保持一致的体型比例

plaintextCopy
1Height comparison reference: "Pal the Party Unicorn" (mint-green plush mascot costume character with gold sequin tailcoat, about 6 feet tall including horn) standing next to "The Sea King Keeper" (towering Poseidon-like lighthouse keeper in a yellow oilskin raincoat over bronze scale armor, about 7 feet tall, holding a trident). Both full body, front view, side by side, plain light grey background, studio lighting, photorealistic, 16:9.

四个地点也用同样方式生成。四个地点共享同一种视觉风格，以便剪在一起：一座老旧的白灰灯塔，油漆剥落、带着锈迹，灰色厚重云层下的黄金时刻光线，以及一片汹涌的深色海洋。正是这种共享的日落光线，导致 Kling 版本的夜间风暴看起来像傍晚。

@Image1：灯塔外观

plaintextCopy
1Wide establishing photo of an old whitewashed lighthouse standing on a rocky, grassy headland surrounded by a rough dark sea. A small white keeper's cottage with a slate roof is attached to the base of the tower. The lamp at the top is lit with a warm glow. Waves break white against the black rocks along the shore, a long coastline fades into the distance. Golden hour: heavy grey storm clouds overhead, a thin band of warm orange light on the horizon. Elevated three-quarter view, slightly above the headland. Photorealistic, cinematic, 35mm film look, moody, 16:9. No people, no text, no watermark.

@Image2：灯塔入口

plaintextCopy
1Low-angle photo of the base of the same old whitewashed lighthouse tower, paint peeling and stained with rust and grime. A heavy arched stone doorway sits in the center with a rusted iron door hanging half open, revealing worn stone steps leading up into darkness. To the right, a small weathered white keeper's cottage with a red tile roof and a brick chimney. In front of the door: cracked stone steps, coils of old rope, rusty buckets and scrap metal scattered on wet rocks and wild grass. The sea and a distant coastline behind. Golden hour under heavy grey clouds, warm low sunlight from the horizon. Photorealistic, cinematic, 35mm film look, 16:9. No people, no text, no watermark.

@Image3：螺旋楼梯

plaintextCopy
1Interior of the same old lighthouse tower, looking straight down from the top of a cast-iron spiral staircase that winds down the center of a tall cylindrical shaft, forming a hypnotic spiral. Rusted iron steps and a thin rusty handrail. Curved whitewashed walls with peeling paint and long rust stains, a few small square windows letting in cool daylight and glimpses of the sea, an old framed chart hanging on the wall. Mixed lighting: soft cool light from the windows, warm light from below. Photorealistic, cinematic, 35mm film look, high detail, 16:9. No people, no text, no watermark.

@Image4：顶部灯室

plaintextCopy
1Interior of the lamp room at the top of the same old lighthouse. A huge glowing Fresnel lens with concentric glass rings sits in the center on a black iron pedestal, its warm golden light the main light source. Panoramic iron-framed windows wrap around the whole room, some panes cracked, showing a rough dark ocean with whitecaps under heavy grey clouds and a warm golden-hour glow on the horizon. A circular iron railing and old brass and iron machinery around the base of the lens. Slightly elevated wide shot from the edge of the room. Photorealistic, cinematic, 35mm film look, moody, 16:9. No people, no text, no watermark.

提示词

这个提示词基本上就是一份镜头清单。每个节拍都有时间范围、镜头运动、动作和声音。除此之外，还有一小段规则告诉模型镜头如何连接，以及一次击打应该有什么感觉。下面是三个模型收到的完全相同的提示词。对于 Wan 3.0，唯一的变化是参考语法：用 “Image 1” 代替 “@Image1”。

完整的 30 秒提示词

plaintextCopy
1A 30-second cinematic Halloween action-comedy short with multiple cuts, photorealistic, shot on 35mm film, 16:9. Stormy night at sea, heavy rain throughout, cold blue moonlight, lightning flashes, a sweeping warm lighthouse beam.
2
3REFERENCES:
4- @Image1: the lighthouse on its rocky island (exterior, island, sea).
5- @Image2: the lighthouse entrance and iron door.
6- @Image3: the spiral staircase inside the lighthouse.
7- @Image4: the gallery platform at the top of the lighthouse.
8- @Image5: "The Sea King Keeper" — keep his face, silver-white hair, seaweed-braided grey-green beard, yellow oilskin raincoat over bronze scale armor, and rusted bronze trident exactly as shown.
9- @Image6: "Pal the Party Unicorn" — a real practical mascot costume character, not cartoon. Keep the mint-green plush fur, sherbet mane, striped horn, gold sequin tailcoat, red polka-dot bow tie and "PAL" badge exactly as shown. Pal carries an orange plastic jack-o'-lantern candy bucket in every shot.
10Both characters must look identical in every shot. Pal's costume is soaked and matted by the rain. The Sea King Keeper is a head taller than Pal, and their proportions never change.
11
12SHOTS:
13[0-3s] Aerial wide shot pushing in fast toward the lighthouse from @Image1. Huge waves smash against the rocks, the lighthouse beam sweeps across the lens, lightning flashes. Sound: thunder, roaring sea, howling wind.
14
15[3-5s] Low tracking shot from behind: Pal from @Image6 walks across the wet black rocks toward screen right, heading for the door from @Image2, candy bucket swinging, rainwater dripping off the sequins. Sound: heavy rain across the stereo field, squelching footsteps.
16
17[5-7s] Close-up: Pal's hoof pounds the iron door three times, each knock shaking rainwater off the metal. Sound: three deep metallic knocks echoing.
18
19[7-10s] Low-angle medium shot: the door is yanked open and slams against the wall. The Sea King Keeper from @Image5 fills the doorway, trident in hand, beard dripping, warm light behind him. Pal tilts its head and says cheerfully in English, in a bright 1990s kids' TV host voice: "Trick or treat?"
20
21[10-12s] Close-up on the Sea King Keeper's face, slowly pushing in. He narrows his eyes, knuckles whitening around the trident, and growls in English, in a deep gravelly voice: "...Trick."
22
23[12-15s] Wide shot, slow motion ramping back to real speed: he lunges with the trident. Pal sidesteps, the prongs graze its mane, slice through a line of raindrops and slam into the door frame, blasting out brick chips. Pal ducks under his arm and dashes inside. He rips the trident free, spins around and charges in after it. Sound: the trident whooshing through the air, a heavy thud as it bites into brick.
24
25[15-19s] Cut on action to the bottom of the spiral staircase from @Image3, close-quarters brawl. Eye-level handheld tracking camera that shakes with the impacts. The Sea King Keeper thrusts three times in a row; Pal parries left and right with its horn, every clash bursting with sparks, each blow knocking Pal back up the steps. The final strike smashes into the iron railing and bends it. Sound: dense, heavy metal clangs echoing through the tower, the groan of bending iron.
26
27[19-22s] Still inside the spiral staircase from @Image3, the fight escalates into Hong Kong wire-work style aerial combat, light, fluid and gravity-defying. The camera rises vertically through the central shaft of the staircase, keeping both characters in frame.
28Beat 1: Pal kicks off the curved brick wall and wall-runs horizontally around the spiral, climbing upward, each step cracking the bricks, candy bucket swinging.
29Beat 2: The Sea King Keeper plants his trident on the iron railing and pole-vaults off it, spinning up to the flight above and landing hard enough to shake the steps, cutting Pal off.
30Beat 3: He sweeps the trident; Pal backflips over the shaft and lands balanced on its hooves on the thin iron railing. Then both leap upward together and smash through the hatch at the top of the stairs.
31Sound: rapid thuds of hooves on the wall, metallic clang of the trident striking the railing, the heavy impact of his landing, wind rushing down through the hatch, then the bang of the hatch bursting open.
32
33[22-24s] Cut on action to the gallery platform from @Image4: they burst out of the hatch and collide head-on in the pouring rain, trident and horn locked together, sparks and rainwater spraying, both silhouetted by lightning. The camera pulls back into an extreme wide shot: behind them a monstrous rogue wave rises higher than the top of the lighthouse and crashes down, swallowing them both. Sound: a piercing scrape of locked metal, thunder, the massive roar of the wave.
34
35[24-27s] Medium shot on the platform: the water drains away, pouring off the edge. Both characters sit slumped against the bent railing, completely soaked, panting, glancing at each other. Sound: only rain, running water and their heavy breathing.
36
37[27-30s] The Sea King Keeper sighs, pulls a single sardine from his raincoat pocket and drops it into Pal's candy bucket. Cut to a close-up of Pal looking down into the bucket as its wide stitched smile slowly fades. Sound: a wet "plop", then a short cheerful 1990s kids' TV jingle sting.
38
39ACTION AND EDITING RULES:
40- Connect every shot with a cut on action: each movement started in one shot finishes in the next. No skipped actions.
41- Pal always moves toward screen right and upward, with the Sea King Keeper chasing from behind. Keep screen direction consistent.
42- Every hit has weight and recoil: sparks, brick chips, water spray, characters knocked back, metal bending.
43- Fight moves must be clear and readable, each with a clear start and landing. No limbs passing through bodies, no warping weapons, no extra limbs.
44
45STYLE AND CONSTRAINTS:
46Deadpan comedy tone: epic action, absurd ending. Realistic water, rain and fabric physics. No subtitles, no on-screen text, no watermark, no background music except the final jingle sting. No extra characters.

我中途改了什么

我最初版本的楼梯打斗是一团糨糊：两个角色在同一个地方乱挥。三个改动解决了问题，现在我在每个动作提示词里都会用：

1. 三秒内三个清晰动作。 跑墙、撑杆跳、后空翻。每个动作都说明谁做、做什么、落在哪里。
2. 每个打斗镜头只用一个镜头运动。 这里，摄影机沿楼梯井直直上升。模型处理一个清晰运动，比处理组合运动好得多。
3. 写出每次击打的结果，而不是写“他们打斗”。 火花、砖屑、水花、后坐、弯曲的金属。正是这些让一次击打显得沉重。

另外两个有帮助的习惯：给每张参考图只分配一个任务（要么地点，要么身份，绝不要两者兼有），并用 “cut on action” 连接镜头，让一个镜头里开始的运动在下一个镜头里结束。

我如何测试

关于挑选版本的一点说明：我为每个模型生成了两段视频，老实说，每一条都有亮点。我给每个模型最完整的那条打分。如果你愿意剪辑，把每条里最好的部分剪在一起，你的最终视频会比这里看到的任何一条都更好。

评分卡

我从九个类别给每条视频打分，每项十分。总分是九项得分的简单平均值。

只有一个人（我）打分，所以把每个数字视为 ±0.5。

结果

绿色标记每个类别中的最高分。Seedance 领先半分。Kling 和 Wan 只差 0.12，这在我的 ±0.5 误差范围内，所以质量上可以视为平手。价格才是拉开差距的地方。

现在加上价格

Seedance 2.5 以半分优势击败 Kling 4.0，但每次运行成本几乎是后者的三倍。按每美元得分计算，Kling 4.0 遥遥领先。

并排对比

每个模型相同时间点的画面。行：Kling 4.0、Seedance 2.5、Wan 3.0。列：8.5s、13s、17s、20.5s、23s、28.5s。

你首先注意到的是：Seedance 和 Wan 给了我一场风暴。Kling 给我的则是一个略显阴郁的下午。原因在我的参考图里，它们是在日落时拍摄的。当图像和文字发生冲突时，Kling 相信图像。这让它损失了氛围分，但也正是 Kling 的角色看起来最像我的设计的原因。

节奏也不同。13 秒时，Kling 已经进入楼梯间，而另外两个还在门口打斗。23 秒时，Kling 还在灯室，而另外两个已经打到了巨浪。

Kling 4.0：最忠实，也最有性价比

做得好的地方

• 最佳角色保真度。 海王的脸、海藻编辫的胡子、雨衣和盔甲，以及 Pal 的薄荷绿毛绒、亮片和 “PAL” 徽章，几乎与设定图一模一样，并且在镜头之间保持一致。
• 最快的剪辑节奏。 30 秒内约 14 次剪切，是三者中最多的。感觉像预告片。
• 最佳空中瞬间。 Pal 翻过楼梯井的俯视螺旋镜头，是三者中最接近钢丝武打要求的，而且它环绕灯塔的巨浪有真实体积感。
• 最佳笑点。 Pal 低头看向桶里，你能看到那条鱼。
• 最低成本： 每次运行 $6.41。

不足之处

• 没有“暴风雨之夜”。 开场是阴天傍晚、小雨，因为它优先遵循了日落参考图，而不是文字。
• 打斗接触偏软。 门口的三叉戟突刺基本被跳过，楼梯底部看起来像追逐，灯室交锋缺乏重量。
• 时间线漂移。 前半段提前了约两秒，这导致巨浪被推到 24–27 秒，并压缩了结局。
• 声音在场景中变化。 角色在不同台词之间的声音不一致。

Kling 4.0，每秒一帧（0–29s，从左到右，从上到下）

Seedance 2.5：最好看的片段，但价格更高

做得好的地方

• 最佳氛围和图像质量。 夜晚、雨、温暖的门口灯光和灯塔光束，层次像一部真正的电影。
• 门口打斗遵循了脚本。 三叉戟没刺中，咬进门框并炸出砖屑，与写法一致。
• 最壮观的巨浪。 一道卷曲的水墙升过灯塔顶部，两个剪影在其中，正好在 22–24 秒出现。
• 清楚的沙丁鱼。 你能看到它在海王手里，也能看到它被放进桶里。

不足之处

• Pal 偏离了设定。 更圆、更卡通，更像玩具小马，而不是我的玩偶服角色。
• 缺少两个打斗节拍。 楼梯底部的近距离打斗被俯拍镜头取代，灯室交锋被跳过。
• 没有失望情绪。 Pal 在结局中一直保持微笑。
• 一句台词读错。 海王的 “...Trick.” 变成了 “Trip”。
• 最高成本： 每次运行 $17.76。

Seedance 2.5，每秒一帧（0–29s）

Wan 3.0：风暴很棒，打斗最弱

做得好的地方

• 镜头覆盖广。 门口交锋、楼梯乱斗、俯视螺旋、挂墙、舱门、带闪电的灯室交锋、巨浪、瘫倒、沙丁鱼：11 个镜头里大多数都有对应瞬间。
• 风暴到位。 从头到尾都是夜晚、大雨和闪电。
• 对白干净。 两句台词都正确、同步，声音也一致。
• 角色稳定。 两个角色在每个地点都接近其设定图。

不足之处

• 打斗冲击力最弱（5.5）。 火花很多，但击打没有重量。
• 结局移到了户外。 巨浪之后，他们坐在露天栏杆旁，而不是参考图中的灯室里。
• 笑点反了。 Pal 从失望变成高兴，与要求相反。
• 巨浪更小。 只是灯室周围的水花，而不是一道水墙砸下来。

Wan 3.0，每秒一帧（0–29s）

耻辱堂

每次测试都有花絮。这些是我最喜欢的：

• Seedance 的海神，盯着独角兽，低吼道：“...Trip。”
• Kling 的角色在场景中途换了声音。我猜是方法派表演。
• 我要的是一个诡异儿童节目吉祥物。 Seedance 给了我一只玩具小马。
• Wan 的 Pal 讨糖只得到一条沙丁鱼，却非常兴奋。老实说？敬你。
• Kling 的“夜间猛烈风暴” 是一个多云下午。

我学到了什么

1. 当参考图和文字冲突时，模型的选择不同。 Kling 4.0 遵循图像；Seedance 和 Wan 遵循文字。如果你想让 Kling 以文字为准，就让你的参考图与文字匹配。
2. 30 秒内 11 个镜头会迫使每个模型压缩。 Kling 压缩了门口打斗，Seedance 删掉了楼梯乱斗和灯室交锋，Wan 压缩得最少。
3. 打斗仍然是最难的类别。 三个模型的打斗冲击力平均为 6.5，是九个类别中最低的，即使最高分也只有 7.5。
4. 细微面部转折是共同弱点。 三者都没有完全呈现“笑容慢慢消失”。Kling 最接近，因为它让 Pal 低头看向桶里。
5. 口型同步已经解决；台词准确性和声音一致性是新问题。 三者都同步，但 Seedance 读错了一句台词，Kling 的声音漂移了。只有 Wan 两者都做对。
6. 价格差距大于质量差距。 Seedance 和 Kling 相差半分。它们之间也相差近三倍成本。

那么，Kling 4.0 值得吗？

• 如果你的角色需要看起来和你的设计一模一样，并且你要制作大量片段，选择 Kling 4.0。它是本次测试中性价比最高的，而且参考保真度是三者中最好的。只要确保你的参考图已经具备你想要的光照和氛围。
• 如果你需要一两个氛围最佳的英雄镜头，而且预算不是问题，选择 Seedance 2.5。
• 无论你选哪个，都要计划生成不止一次并剪辑。尤其是打斗场景，仍然需要人类在多次生成之间剪辑。

局限性

• 样本很小。 每个模型两次生成，只给一次打分。这展示的是每个模型在这个提示词下能做到什么，而不是它的平均水平或成功率。
• 日落参考图 vs 夜间提示词。 这种冲突有助于看出每个模型相信哪种输入，但也让 Kling 在氛围上吃了亏。
• 只有一位评审。 所有分数都是我的，±0.5。你的优先级可能会以不同方式权衡这九个类别。

自己运行

上面的一切都可复用：角色提示词、六参考图设置、11 镜头提示词和九类别评分卡。换上你自己的角色和地点，保留镜头清单结构，你就能公平比较任何你在考虑的视频模型。

告诉我接下来该让这些模型受什么折磨。万圣节快乐，如果一只独角兽敲你的门，就给它糖吧。

本次测试中的所有角色均为原创。本文中的视频片段由 AI 生成。

在 Atlas Cloud 上试用这些模型

本次测试中的三个模型都可在 Atlas Cloud 上使用。有两种使用方式：

• 在网站上生成。 打开模型页面，上传参考图，粘贴提示词并运行。这是用你自己的角色复现本测试的最快方式。
• 调用 API。 获取一个 API key，用同一个 key 从你自己的应用或流水线调用三个模型，这样可以轻松地把同一提示词批量跑在多个模型上。

模型页面：

• Kling 4.0
• Seedance 2.5
• Wan 3.0
• 浏览所有模型