免费 AI 视频生成
文字一键成片，完全免费
把你想好的文字脚本交给 AI，模型就能自动生成画面、配音和字幕，一次还能排出多个场景，这就是所谓的 AI 视频生成（Text-to-Video）。这个工具跑在完全免费的 Agnes AI 模型上：不用订阅、不用显卡、不限量使用，1K+ 创作者已经在用，代码按 MIT 协议开源。敲一段文字进去，带配音和字幕的成品短片就直接到手。
AI 平权愿景
「解决的办法不是压制 AI，而是让它变成一种更平权的能力，让每个人都知道如何借 AI 创造更多。」—— Bruce Yang，Agnes AI 创始人
AI 视频提示词示例：真实作品与原始 Video Prompt
真实 AI 视频作品与完整视频提示词，复制就能复现同款画面。
炼丹房里的孙悟空大笑
[SUBJECT - IDENTITY LOCK] Use the provided reference image as the identity anchor. Keep exactly the same face, fur color, eye color, crown, body proportions, armor and outfit in every frame. No face swap, no redesign, no changed costume, no changed art style. The subject is ONE original mythic monkey-king warrior: a single adult anthropomorphic monkey with short golden fur on the face, bright amber-gold irises, a red-and-gold phoenix-wing crown, engraved golden chainmail armor, a tiger-skin skirt, a short red cape and dark boots. Both hands are completely EMPTY, with five clearly separated fingers on each hand. There is exactly ONE subject in the frame. The long gold-banded staff stays leaning against the furnace in the background for the whole clip. It is never picked up, never held, and never appears in either hand. [SETTING & LIGHT] Inside an ancient Chinese alchemy chamber, beside a huge bronze eight-trigrams furnace with a low warm fire glowing through its vents and a thin curl of blue-green smoke. Dark stone walls and a dim uncluttered background keep the silhouette clean and readable. Warm amber key light from the furnace falls on the subject, with a soft cool rim light from behind. Stable lighting, no flicker, no extra characters, no moving props, no falling objects. [NEGATIVE PROMPT] a staff or weapon held in the hand, object in either hand, second monkey, duplicate subject, changed crown, changed armor, staff swinging, furnace exploding, fire spreading, smoke covering the face, dark underexposed face, cropped knees, hands out of frame, blurred hands., second subject, extra people, two people, reflection read as another person, crowd, polite chuckle, stiff upright torso, head-only laughing, nodding laugh, both feet leaving the floor, jumping, hopping, stepping during the laugh, losing balance, falling forward, neck hyperextension, unhinged jaw, distorted face, deformed proportions, both hands making the same shape, both hands open, two fists, open palm closing into a fist, fist opening up, arms swapping height, existing cartoon character, branded mascot, recognizable IP lookalike, character redesign, drifting face, drifting outfit, extra limbs, extra fingers, merged fingers, TikTok/Douyin UI, search bar, watermark, logo, readable text, subtitles, camera movement, zoom, cut, speed ramp, slow motion, morph transition, flicker. [MOTION CLIP MC-01 v3 · SOLO · BELLY LAUGH / SILENCE BEAT / PALM AND FIST] This clip contains NO character description. Prepend your own [SUBJECT] block before it. Every line below refers only to "the subject"; do not infer species, age, gender, outfit or art style from this clip. If an identity reference image is supplied, use it as the identity anchor and keep exactly the same face, hair, body proportions and outfit in every frame. [SPEC] Exactly 8.0 seconds, 9:16 vertical, 30fps, 720x1280, one locked-off full-body medium-wide shot, no cuts, no camera movement, no zoom. The subject is visible from head to at least the knees the whole time. Action accents are timed to the laid-back audio track: strong beats at 1.55s / 2.6s / 4.4s / 5.5s / 6.7s / 7.3s, and a 0.25-second SILENCE GAP from 5.17s to 5.41s. [BEAT MAP] 0.0-0.6s | READY: the subject stands centered, both hands loosely held in front of the belly, looking at the lens, softly smiling, knees slightly bent. Almost still. 0.6-1.6s | SHALLOW FOLD: the subject tips the head down and folds the upper body forward about 30-45 degrees, hair falling forward, shoulders shaking with the first laugh, both hands coming up in front of the torso. The lowest point lands exactly on the 1.55s beat. This fold is deliberately SHALLOW — a deeper one comes later. 1.6-2.6s | RISE: the subject unrolls back to upright and tips the head back into a wide laugh, mouth open, eyes narrowed, chest opening. Land the recovery on the 2.6s beat. 2.6-3.6s | DEEP FOLD: the subject hinges forward from the hips by 45-60 degrees, forehead dropping toward the knees, spine rounded, shoulders shaking hard, both hands braced on the belly or slapping down onto the knees, hair falling fully forward. This is the deepest bend in the clip, clearly lower than the first fold; hold it for roughly three consecutive frames before recovering. 3.6-4.4s | THE ONE PEAK: the ONLY full backward lean of the clip — chest open, chin up, head tipped back, mouth wide open in a laugh, eyes almost closed, cheeks lifted, shoulders bouncing hard. Peak exactly on the 4.4s beat. 4.4-5.17s | SETTLE: the subject unrolls to neutral, the laugh winds down, and the torso begins turning a few degrees to one side. 5.17-5.41s | SILENCE FREEZE: on the audio silence gap, the subject suddenly stops moving for the whole gap. No swaying, no blinking pulse, no hand drift. This dead stop is the punchline. 5.41-5.6s | THE TURN (strongest beat at 5.5s): on the downbeat the subject turns three-quarter toward the lens and drives ONE hand forward at chest-to-shoulder height with a fully OPEN PALM, fingers together, palm facing up and slightly toward the lens (HIGH PALM); at the same time the OTHER hand draws into a CLOSED FIST held low in front of the belly, knuckles down (LOW FIST). The wide-open laughing face stays on and the eyes come back to the lens. 5.6-7.9s | PALM AND FIST HOLD: the high open palm stays open and forward, the low closed fist stays closed and low — this one-up / one-down, one-open / one-closed contrast is the signature pose. On the 6.7s and 7.3s beats the upper body makes a small forward-and-back pulse and the open palm makes a tiny push, but neither hand changes shape, neither arm swaps height, and both feet never leave their marks. A broad smile is held throughout. 7.9-8.0s | FINAL HOLD: the smile settles, the pose stays identical, the last frames are completely motionless. End on the still pose, no fade. [CAMERA] Fixed camera at chest height, straight-on, subject centered with headroom and floor contact visible at all times. No dolly, no handheld shake, no push-in, no cut, no zoom, no slow motion, no speed ramp. The high open palm and the low closed fist must both stay fully inside the frame from 5.5s to the end. [PERFORMANCE RULES] The comedy comes from CONTRAST, not from nonstop motion: one shallow fold, one rise, one much deeper fold, a single backward-laugh peak, a dead-stop freeze, then a held palm-and-fist pose. The laugh is whole-body rocking driven from the hips and spine — chest, head and shoulders all travel, the deep fold must be clearly lower than the shallow one, and the backward lean must be clearly higher than neutral. Both feet stay planted for the whole clip; there is NO jump. Keep the same face, body proportions and outfit throughout. Believable hair and cloth inertia one beat late, no hovering, no rubber limbs. [AUDIO] Generate NO audio on the video track. This clip is designed for a post-production layback of the reference 0-8s mixed track (music + laugh bursts + beat accents + the 5.17-5.41s silence gap). No generated dialogue, no generated laugh, no lyrics, no subtitles, no on-screen text.
深夜窗光｜极近面部微表情肖像
[SPEC] 12 秒，9:16 竖屏，24fps，720×1280，一镜到底不切镜，超写实真人电影肖像，纪录片 RAW 质感。 [CHARACTER LOCK] 画面中只有一名明确成年的 27 岁东亚女性，东亚骨相特征清晰：小巧柔和的鹅蛋脸、饱满额头、下颌线自然收束、下巴圆润；深棕色杏仁眼、浅双眼皮，鼻梁细直、鼻尖小巧，唇形上下比例协调，妆容为克制的精致上镜晚妆。冷调中性偏白肤色，保留毛孔、面部绒毛、鼻翼纹理、轻微肤色不均与面颊极淡的自然血色。浓密乌黑长发带松散波浪、略偏分缝，大部分垂在脸左侧与肩头，几缕细碎发丝横过额头与面颊。低饱和灰蓝色柔软针织上衣、领口自然宽松，不戴首饰。全片保持同一张脸、同一发型、同一服装与体型比例；面部是唯一视觉主体。 [SETTING & LIGHT] 极简低亮度室内，暗灰色虚化背景，没有家具与装饰。人物右前方有一扇大窗，阴天自然光经薄纱窗帘柔化。光线柔和均匀、略偏冷，落在右颊、鼻梁与眼睛上，左颊与头发处于柔和阴影里；眼中有小面积自然窗光反光。整体低饱和冷灰电影调色，肤色保留自然暖度，不做强烈橙蓝对比。 [CAMERA & PERFORMANCE] 85mm 人像镜头透视，f/2.0，摄影机在人物正前方略偏左、与眼睛齐平、距离面部 25–40 厘米。面部占画面 80%–90%，额头靠近上缘，下巴与颈部被下缘部分裁切。镜头全程基本固定：不推拉、不旋转、不切机位、不突然变焦，只保留极细微的真实手持呼吸感。 0–2.4 秒：双眼轻闭，头部静止，碎发贴在面颊前；她缓慢睁眼，眼睑自然抬起，瞳孔逐渐落在镜头附近。 2.4–4.8 秒：视线缓慢下移，眼球先动、上眼睑随后跟随；头部只下倾几度，面颊与鼻梁光影发生细微变化。 4.8–7.2 秒：保持低垂凝视；嘴唇随一次自然呼吸微张 1–2 毫米，不说话；睫毛轻微颤动，面颊肌肉有极细微的放松。 7.2–9.6 秒：眼睛先回到镜头，头随后跟上；嘴唇自然闭合；完成一次缓慢完整的眨眼，之后是柔和稳定的注视。 9.6–12 秒：目光短暂转向画面右侧再回到镜头；头部极小幅调整带动几缕发丝；嘴角几乎不可察觉地轻轻上扬，不露齿；最后一秒安静凝视镜头，眼睛湿润明亮。 [TEXTURE] 只保留真实微运动：眼球转动平滑、虹膜大小稳定、眼睑自然闭合、唇纹清晰、呼吸只体现为肩颈极轻起伏。头发仅有细微自然移动，没有风。禁止磨皮、玻璃肌、蜡像质感。 [NEGATIVE PROMPT] 换脸、五官漂移、网红模板脸、欧美骨相、非东亚面孔、未成年观感、过度磨皮、塑料皮肤、油光水光肌、浓密假睫毛、虹膜变形、眼睛突然放大、夸张眨眼、快速转头、嘴部变形、头发闪烁、镜头推拉、数字变焦、场景切换、卡通质感、CG 渲染、字幕、文字、Logo、水印、边框、界面元素。
迷你城市｜一天采购 vlog
[SPEC] 12 秒，9:16 竖屏，24fps，720×1280，超写实真人城市生活 vlog，恰好 6 段相连场景、每段约 2 秒，手持 vlog 摄影，无文字叠层。 [CHARACTER LOCK] 一名明确成年的 26 岁东亚女性，东亚骨相特征清晰：柔和鹅蛋脸、深棕色杏仁眼带浅双眼皮、鼻梁细直、唇形自然，妆容为精致日常妆。齐肩黑色直发，一侧别在耳后，鬓角留几缕碎发。米白色宽松衬衫、浅蓝直筒牛仔裤、白色低帮运动鞋，单肩挎一只小号帆布购物袋。6 段场景里保持同一张脸、同一发型、同一服装、同一体型比例与同一只购物袋。 [SETTING & LIGHT] 白天的中国现代沿海城市。自然光充足、色调干净略低饱和，橱窗有真实反射，特写镜头浅景深。所有店铺招牌留白或不可辨认，画面中不出现任何品牌名。 [SHOT LIST] 0–2 秒：明亮的小公寓室内。她在玄关镜前整理一下，从木桌上拿起手机、钥匙和折好的帆布袋，开门出门；自然手持 vlog 构图，窗外日光洒进来。 2–4 秒：热闹的街市。背后跟拍她走过水果摊、蔬菜摊与小店橱窗；行人与一辆驶过的电动车自然移动；插入一小段她边走边笑的自拍式正面镜头。 4–6 秒：蔬菜摊特写。她的手挑起番茄、黄瓜、青菜与柠檬，端详后放进帆布袋；摊主的手递回零钱；背景是虚化的成筐鲜亮蔬菜。 6–8 秒：社区小超市货架。俯拍她的手把纸包面包、玻璃瓶和几件颜色鲜亮但无品牌的零食放进购物篮；她又看见一件，停顿一下，带着一点好笑多拿了一件。 8–10 秒：回到繁忙街道，双手拎着两只装满的购物袋。先是一小段穿过人群的 POV 镜头，再切到她略显疲惫又觉得好笑的自拍式镜头；车流与过街行人在身后移动；她走到公寓楼门口。 10–12 秒：厨房里，她把袋子放上台面；俯拍蔬菜、面包与水果铺开的全景。她呼一口气，随手把头发松松扎起，掰一小块面包咬下，露出满足的笑。收在温暖的俯拍构图上。 [TEXTURE] 真实手持微抖、自然眨眼、衣物与袋子随动作摆动，拎起满袋时有可信的重量感。各段之间皮肤质感与头发长度一致。不用变速、不用变形转场。 [SOUND] 连续环境声：门响、脚步、街边人声、自行车铃、市场叫卖、购物袋摩擦、厨房声响。不加音乐重音、不加旁白。 [NEGATIVE PROMPT] 换脸、五官漂移、网红模板脸、欧美骨相、非东亚面孔、未成年观感、场景之间服装或发长变化、多出的肢体、多指、手部畸形、物体悬空、可读文字、品牌名、Logo、字幕、水印、背景静止、画面闪烁、过曝。
三步开始创作
从创意到成片，只需三步。
输入创意
在页面中描述你的视频创意——无论是简单的画面描述、完整的故事构思，还是一篇长文章。
一键生成
点击「开始生成视频」，AI 自动完成故事编写、场景生成、配音合成、字幕叠加和视频拼接。
下载成品
生成完成后即可下载带旁白配音和字幕的完整视频，直接发布到社交媒体平台。
什么是 AI 视频生成？
AI 视频生成（Text-to-Video）就是把一段文字脚本交给 AI，让模型自动生成画面、旁白配音和字幕，还能组合出多个场景。Agnes Video Generator 跑在完全免费的 Agnes AI 模型上，不用订阅、不用显卡、没有用量上限。按官方项目数据，目前支持 22 种语言界面，服务了 1K+ 活跃创作者，MIT 协议开源，普通用户在浏览器里就能一键做出带配音和字幕的成片。
Agnes Video Generator 与传统方案对比
| 维度 | 本工具 | 传统/付费方案 |
|---|---|---|
| 费用 | 完全免费 | 按月订阅或按量计费 |
| 硬件 | 无需显卡，云端运行 | 需高端 GPU 或本地渲染 |
| 开源 | MIT 协议可自部署 | 闭源私有 |
| 上手门槛 | 输入文字一键生成 | 需学习剪辑/建模 |
技术细节与模型能力请参阅 Agnes AI 官方 API 文档
为什么选择 Agnes Video Generator？
眼下做 AI 视频的门槛实在吓人：国外包月就要几十美元，国内免费额度一用完就开始按秒算钱，想本地跑开源模型，先得掏几万块买显卡。我们觉得 AI 应该是更平等的能力，世界级的 AI 理应人人可用。用 Agnes AI 的免费视频模型，不用写代码，打开浏览器就能开工。
零成本 AI 视频
文本生成、图片生成、视频生成、语音合成，这些 AI 能力全都免费，没有水印，也不限制 Token。底层是 Agnes AI 的免费视频模型，云端跑完，浏览器里直接出片，不用写代码，也不需要高性能显卡。
五种创作模式
简单视频、创意长视频、稿件长视频、诗词视频、数字人口播，五种模式从单段快速出片到完整故事片都能覆盖，看你要做什么类型的视频。
AI 配音与字幕
微软 Edge TTS 免费语音合成，内置 4 种中文语音角色。字幕按词级时间轴自动生成，支持多行换行，样式可以完全自定义。
多语言界面
支持 22 种语言：中文、English、Deutsch、Français、Nederlands、Español、Português、Italiano、Русский、日本語、한국어、Melayu、Indonesia、हिन्दी、Türkçe、Tiếng Việt、বাংলা、ไทย、Filipino、العربية、اردو、فارسی，不管你说哪种语言，打开就能用。
高级创作控制
角色参考图、自定义尾帧、三种视频串联模式（keyframes/ti2vid/none）、多种分辨率和时长可选。
生产级特性
断点续传（任务中断自动恢复）、WebSocket 实时进度推送、任务管理、页面内 API Key 配置。
五种视频创作模式
满足不同创作需求，从快速单段到完整故事片，再到诗词动画和数字人口播。
简单视频
单条 prompt 生成单段 AI 视频，完整暴露所有参数（生成模式、时长、分辨率、seed、负向提示词）。适合快速制作单段 AI 视频片段。
创意长视频
AI 全流程自动化：创意→故事→脚本→角色参考图→多场景视频→TTS 旁白→字幕→拼接成片。适合故事短片、创意视频创作。
稿件长视频
粘贴长文/稿件，AI 自动拆段、逐段生成 AI 视频、统一 TTS 配音+字幕，最终拼接为完整视频。适合解说视频、课程内容。
诗词视频
输入一首诗，AI 逐句拆分场景：原句作为旁白朗读与字幕，画面按诗意逐句生成，最终拼接成意境短片。适合古诗、诗歌意境动画。
数字人口播
AI 生成数字人形象或上传自定义形象，动态口播片段配合 TTS 配音和字幕，循环拼接合成完整口播视频。
开发故事
了解这个项目从零到一的完整开发历程。
SandGrid@lcy362
Agnes Video Generator 项目作者 · 全栈开发者
这篇开发纪实把 Agnes Video Generator 从最初的想法到最终上线完整捋了一遍：技术选型、架构设计、多语言支持，还有接入 AI 时踩过的坑和解决办法。对 AI 视频生成感兴趣，或者想看看开源项目是怎么一步步做出来的，都值得翻一翻。
访问 GitHub阅读完整开发故事 →