APIMODELSAPIMODELS

LTX 2.5 自然与风景提示词

收录 80 多条 LTX-2.5 视频提示词,来自 Civitai、fal.ai 的 LTX-2.5 示例页和 X。每条都逐字全文收录、署名作者、链回原帖,原帖有视频的,卡片里直接播放。这些提示词都不是我们写的。Civitai 上的大多是作者在自己电脑上用 LTX-2.5 开源权重跑的(ComfyUI 里的 dev 或蒸馏版),模型是同一个家族,但没有经过我们的 API。

← 回到全部 89 条 LTX 2.5 提示词

最近更新: · LTX 2.5 Fast API 价格与在线体验 · API 文档

一句话摘要

LTX-2.5 是 Lightricks 在 2026 年 8 月 11 日发布的开源权重视频模型,每条视频都会连声音一起生成。apimodels.app 上有两档:LTX 2.5 Fast(ltx-2.5-fast)可以只给提示词,也可以再加一张首帧图或一段音频,输出 720p、1080p 或 1440p,24 到 50 fps,6 到 20 秒,$0.081/秒起;LTX 2.5 Pro(ltx-2.5-pro)是 Lightricks 定位给成片用的一档,只做文生和首帧图生,720p 或 1080p,6、8、10 秒三档,$0.108/秒起。提示词最长 5000 字符;没有 4K、尾帧、参考图和视频输入;失败不扣费。本页收录 80 多条别人公开发布的 LTX-2.5 提示词,来自 Civitai、fal.ai 的 LTX-2.5 页面和 X。Civitai 上很多是在本地用开源权重、配合作者自己的参数跑的,同一段文字放到 API 上,出来的效果可能不一样。

出处:Lightricks LTX(模型方官方站点)

悬停预览 · 点击有声

秋日瀑布与林间木屋

A stylized illustration depicts a serene natural scene with a waterfall cascading down a rocky cliff into a blue river. Two trees with vibrant orange foliage stand prominently in the foreground, while a small wooden cabin is nestled on a hillside behind the river. The landscape is filled with green grass and scattered rocks, under a clear blue sky.

来自 Esteban099009看原帖
悬停预览 · 点击有声

日落雪山风景

A serene mountain landscape at sunset, with snow-capped peaks reflecting golden light, photorealistic.

来自 Esteban099009看原帖
悬停预览 · 点击有声

异星巨石水池

A single unbroken continuous wide establishing shot in smooth digital gouache and watercolor painting style, maintaining a constant camera distance with both flanking monoliths and the right-bank seated woman fully framed throughout. The camera glides in a slow, weightless lateral parallax arc with a subtle upward float between the monoliths, creating rich three-dimensional depth separation between the towering foreground basalt-and-sandstone walls, the midground turquoise pool and hovering golden rings, the terraced alien foliage, and the distant sunlit cerulean sky. Across the glassy turquoise pool, the three concentric standing-wave ripples continuously expand and shimmer outward with golden-turquoise reflections as steady vertical streams of luminous water droplets ascend rhythmically toward the three tilted golden harmonic torus rings, which spin smoothly around their tilted axes and pulse with warm radiant energy. Abundant golden pollen motes and translucent backlit mist swirl in slow three-dimensional eddies through shafts of warm golden-hour sunlight, silver-teal alien ferns, glowing cyan-white flora pods, and hanging vines sway fluidly in a lively garden breeze, recessed gold waveguides and cyan-white oval quartz resonators along the monoliths throb with a gentle rhythmic luminescence, and wispy cirrus clouds drift slowly overhead. On the mossy right bank, the seated woman in her draped white-and-gold tunic remains peacefully poised in left-facing side profile with her hands resting in her lap, breathing gently as the warm breeze softly stirs the draped folds of her tunic and sunlight glints off the tiny gold point behind her ear. The tranquil ambient soundscape layers a pure, resonant crystalline singing-bowl harmonic chord radiating from the spinning golden rings with delicate, rhythmic water-droplet chimes rising from the pool and the soft, airy rustle of leaves in a warm garden breeze.

来自 J032看原帖
悬停预览 · 点击有声

VHS火山山谷风景

pixelated, low-resolution, low-angle, looking down VHS video of an volcanic summer scene in fluorescent colors with a blue glowing sky, red rocky mountains in the background, a lush verdant forest of pine and fir and birch trees and astrolabes in the background infront of the mountains, a blooming plain in front next to an atelier inside a wooden shed next to the trees in the background, a thatch of pine and fir and birch trees with a lot of mini R2-D2s all over the third tree on the left side in the middleground, a brutalist glass yurt with a rainbow-colored-smoking chimney with an out-of-focus violent lake in the middleground, a the bottom half of a swath of pine and fir and birch trees on the right side in the middleground, and in the foreground a louvered cantelivered fence and a plate with pastichio on the left side and pistachio on the right side and a few out-of-focus pine and fir and birch trees and an in-focus see-through glass path leading to the top half of a 40-meter tall R2-D2 statue with an long wooden pegleg waiving a wooden placard with glass tubes filled with neon lights that read "sound on" in the middleground. A spaceship with the top half of the USS Enterprise and the bottom half of a TIE fighter quickly spins 1080 degrees down from offscreen in the background on the top right corner of the frame and lands on the using jet engines to hover above the lake and a landing platform explodes forward from it and crashes into the camera and for the rest of the shot there is a crack of glass in the shot and Darth Vader made of glass carrying a glass keg of amber beer, Chewbacca, made of glass pulling a rickshaw, glass, R2-D2, glass, C-3P0, made of glass, glowing green dreadlocks, smoking from a bong, riding on a skateboard ontop of rollerblades and holding a chest freezer hooked up to a generator above his head, Superman, transparent, carrying 5 translucent pizza boxes, and Godzilla, clear, carrying two glass tanks of odorless propane walk ride on a stegosaurus from inside the Millenium Falcon down to a giant glazed donut bounces them all over the screen like a giant trampoline and into a spinning cylcone in the bottom right corner which comically spins them through a grass window in the yurt and they all land comfortably on beanbags and throughout the video a narrator in a high, squeaky, whiney, nasally, Yiddish accent sings: "A woman must be like a fine chicken stock? She must be rich, full-bodied, and able to complement without overwhelming?".

来自 msavitt623看原帖
悬停预览 · 点击有声

立体派秋日木屋风景

cubism-style, tilt-shift lens shot of an idyllic fall scene in pastel colors with a sky, snowy mountains in the background, a snowy forest of pine and fir and birch trees in the background infront of the mountains, a snowy plain in front of the trees in the background, a thatch of pine and fir and birch trees on the first one-fifth of the left side in the middleground, a rustic log cabin with a smoking chimney with an out-of-focus frozen lake in the second and third fifths of the middleground, a swath of pine and fir and birch trees on the fourth and fifth fifths on the frame on the right side in the middleground, and in the foreground a rustic crumbling wooden fence and more snowy fields and a few out-of-focus pine and fir and birch trees and an in-focus dirt path leading to the rustic log cabin in the middleground. The Millenium Falcon spaceship drifts down from offscreen on the top of the frame and lands on the frozen lake and a landing platform descends from it and Darth Vader carrying a keg of beer, Chewbacca pulling a rickshaw in R2-D2 and C-3P0, dreadlocks, smoking from a bong, a chest freezer hooked up to a generator in it, Superman carrying 5 pizza boxes, and Godzilla carrying two tanks of propane walk from inside the Millenium Falcon down the landing platform across the ice and into the rustic log cabin and throughout the video a narrator in a high, squeaky, whiney, nasally, Yiddish accent subtly implants disconcerting thoughts into your mind.

来自 msavitt623看原帖
悬停预览 · 点击有声

佛蒙特雪湖日落

An image with the look of 1960s 35mm film. A snowy landscape in Vermont, with the sun setting beyond a lakeshore.

来自 Franklin_Sports看原帖

怎么写好 LTX 2.5 提示词

  1. 1一个镜头写成一段话,按发生顺序写. 本页 fal.ai 的那些示例,每条都是从头到尾描述一个机位:景别和主体、发生了什么、镜头怎么动、光线,最后才是镜头焦段。大意是「缓慢推近擂台上的拳击手,顶光穿过烟雾打在肩上,50mm、f/2」。一串关键词堆在一起,模型拿不到先后顺序。
  2. 2把声音写进去,反正每条都会出声. LTX-2.5 的声音和画面一起生成,我们的 API 也没有静音选项,所以要不要声音不由你,怎么响可以由你。台词放进引号、写明是谁说的,再写环境音和最关键的那一下声音。本页有几条 Civitai 提示词把两个人的对话一句一句写好,每个人开口前先交代嗓音。
  3. 3动作长度要对得上 API 的时长档. Fast 的文生和图生是 6 到 20 秒、只能取偶数秒,1440p 或 48 / 50 fps 时只有 6、8、10 秒;Pro 只有 6、8、10 秒。本地工作流常用帧数定长度,除以帧率就知道原帖要的是几秒,再挑最接近的一档。API 不会看提示词自己定长度,要自己传 duration。
  4. 4传了首帧,就写动作,别再描述画面. 图生视频里,主体、场景和构图已经由那张图定死了,提示词就专心写什么在动:谁转身、镜头怎么走、听到什么。成片会保持图片的比例;带首帧又传 aspect_ratio,会直接返回 400。
  5. 5音频生视频:音轨说了算,提示词负责选角. LTX 2.5 Fast 收到 audio_url 后,会把这段音频原样当成片的声音,再按它生成画面。提示词写清楚画面里是谁、怎么动、怎么取景就够了,比如我们文档里的例子:明亮的演播室里,一位年轻主持人对着镜头讲话,手势自然,中景。长度默认跟着音频走,也可以自己设 duration。Pro 不收音频。
  6. 6API 收不了的东西,先删掉. 本地 LTX-2.5 工作流里可以挂负面提示词、LoRA 触发词、额外的引导图或尾帧。我们的请求只收提示词、一张首帧、一段音频和生成参数,别的都不收:尾帧、参考图、视频输入会返回 400。也没有负面提示词这个字段,原帖写在负面提示词里的东西,不挪进正文就等于没了。
  7. 7画面文字和很乱的物理,别指望一次成. Lightricks 自己的提示词指南也说了,画面里的文字和混乱的物理运动目前还不稳定。如果一个镜头靠一块看得清的招牌,或者一堆东西互相撞,做好多跑几次的准备,文字也可以剪辑时再加。

提示词公式

景别和主体 + 按先后写的一个清楚动作 + 镜头运动 + 光线和时间 + 焦段或胶片质感 + 引号里的台词和环境音 + 这一档允许的时长(用 duration 传,不写进提示词)

例(本库收录的原文,原样引用,作者 hichuicomfy,发在 Civitai):Style: high-resolution food photography - close-up - The chocolate is poured over the whipped cream and berries. The chocolate flows slowly, creating rich, glossy streaks across the white cream and the blue blueberry, the deep brown contrasting with the vibrant fruit. A soft, rich sound of chocolate dripping and settling accompanies the pour.

常见问题

LTX-2.5 是什么?

LTX-2.5 是 Lightricks 的视频模型,2026 年 8 月 11 日以开源权重发布,每条视频都会把声音和画面一起生成。Lightricks 的 API 分两档:Fast 主打速度和成本,Pro 按它的说法主打画面保真和运动稳定。Lightricks 还说 LTX-2.5 新增了原生多镜头、新的视频解码器和更好的提示词遵循,这些是厂商的说法,不是我们的测试结论。

这些提示词是谁写的?

都是别人写的:每一条都是创作者在 Civitai 或 X 上公开发的,或者是 fal.ai 放在它 LTX-2.5 页面上的,我们一个字都没写、也没改。每张卡片都逐字全文引用提示词、写明作者、链回原帖。提示词和视频归原作者所有;我们自己写的说明和整理的数据按 CC BY 4.0 授权。要更正署名或撤下某一条,发邮件给我们,或者在 GitHub 仓库提 issue。

Civitai 上的提示词,我用 API 跑出来不一样,为什么?

因为本页 Civitai 的提示词大多是作者在自己电脑上用 LTX-2.5 开源权重跑的,不是在我们的 API 上,同一段文字出来的效果自然会有差别。本地 ComfyUI 可能用的是 dev 或蒸馏版权重,还可能挂了 LoRA、调了采样参数、写了负面提示词或者加了引导图,这些都不在提示词文字里。文字能照搬,这些设置搬不过来。卡片上的视频是作者那套设置的结果,在 Fast 或 Pro 上要多试几次才能接近。

在哪能跑这些提示词?多少钱?

在 apimodels.app 上,模型名 ltx-2.5-fast 和 ltx-2.5-pro,接口 POST /v1/video/generations,按成片秒数计费。Fast:720p $0.081/秒、1080p $0.117/秒、1440p $0.171/秒,720p 6 秒 $0.486,1080p 10 秒 $1.17,1080p 20 秒 $2.34。Pro:720p $0.108/秒、1080p $0.153/秒,1080p 10 秒 $1.53。两档都比 Lightricks 官方 API 价低 10%;Fast 的音频生视频按成片长度计到 0.01 秒;失败不扣费。常见的用法是先在 Fast 上试提示词,选中的镜头再拿到 Pro 上出,两档字段一样,只改一个模型名。

本页的提示词放到 API 上被拒或者失败,为什么?

超出 API 限制的请求,会在建任务之前直接返回 400,不扣费。限制有这些:提示词最长 5000 字符;Fast 是 6 到 20 秒的偶数秒,Pro 只有 6、8、10 秒;没有 4K、尾帧、参考图和视频输入;Pro 不收音频、没有 1440p;传了首帧就不能再传 aspect_ratio。任务开始以后,也可能因为内容审核或上游繁忙而失败,同样不扣费。

LTX-2.5 会自己配声音吗?

会:每条视频的声音都和画面一起生成,我们的 API 也没有静音选项。在意声音的话,把台词放进引号,写明环境音。用 LTX 2.5 Fast 还可以直接传你自己的音轨(audio_url),画面会按这段音频来生成。

这些提示词和视频能商用吗?

提示词和视频归原作者所有,商用前先看原帖怎么说。你自己生成的视频,按你用来生成的那个服务的条款执行。如果你在本地跑开源权重,它用的是 Lightricks 的 LTX-2.x Community License,有自己的条件,商用前先读一遍。

在 apimodels.app 上跑这些提示词

LTX 2.5 Fast 已经上线,和另外 180+ 个图像、视频、语音、大模型共用一个 API、一个 key,按量计费、注册送 $0.10 试用额度。

这份提示词库在 GitHub 上是开源的

整份合集以 Markdown 和 JSON 放在 GitHub 上:每条提示词都带作者署名和原帖链接(Civitai、fal.ai 或 X)。提示词逐字全文引用,归原作者所有。视频也归作者所有,只链原帖、不转存。我们自己写的说明和整理的数据按 CC BY 4.0 授权。要更正署名或撤下某一条,提 issue 即可。

在 GitHub 上查看

其它提示词库