AI Video by ZrythCreate yours

Tutorial on rendering video from code with a seek(t) pipeline

245 views·1 saves·Posted 2 Oct 2026 · 9:00 · Explainers

Our note

A nine-minute Chinese tutorial on making video with a language model, covering a seek(t) frame model, spring easing, frame versus time driving and deterministic render rules; the creator says the script, animation, music and edit were AI-made, with a cloned voice from VoxCPM2.

The creator’s post

这几天 Opus 5.5 做的视频刷屏,我也用它做了一条教程:怎么用大语言模型做视频。 写稿、动画、配音、配乐、剪辑,全是 AI 做的。没有一帧来自视频模型,也没有一刀是人工剪的。我只做了两件事:提意见,拍板。 整条产线已经开源,链接在评论区。 画面不是「生成」的,是「算」出来的。Claude 写的是代码:每一帧只由时间 t 决定,浏览器按 t 把画面画出来,无头 Chrome 八个进程并行逐帧拍摄,FFmpeg 拼成视频。 同一份源码,任何时候重新渲染,每一帧都逐像素一致。 配音是时钟:每个镜头多长,由实际配音决定;镜头里的每个动作都挂在旁白的某个短语上,换了配音,画面自动跟着伸缩。 声音也是代码:配乐和音效用 numpy 合成,音效挂在画面动作的那一帧上。 我先把公开的教程和开源 skills 让 Claude 学了一遍,整理成一套可复现、可交付的 pipeline,再和 Claude 一轮轮迭代。我没写代码,也没动剪辑软件,做的事情就是看片、提意见、拍板。 模型听不到声音,也看不了连续播放,所以每次出片前跑一遍自动质检:空白帧、静止段、跨进程重放是否逐像素一致、声画同步、响度、背景声有没有盖过配音、物理校验。全部通过才交给我看。 最后的配音试过火山引擎,音色始终不满意,最后用开源的 VoxCPM2 克隆了我自己的声音。 最后开源了渲染引擎、并行渲染器、配音与校对、配乐与混音、卡拉 OK 字幕、出厂质检,项目规矩文件,一个可以直接装进 Claude Code 的 make-video 技能,还有这次踩过的 20 个坑和修法。 终于理解了梁文峰之前在 DeepSeek 内部讲话了,AGI 的主线还是语言模型,智能的上限目前还在语言模型上 Scaling。代码能力突破之后,语言就能变成画面、运动和物理。语言模型,也可以是世界模型。

A prompt to recreate it

Written by Zryth

The creator didn’t share their prompt, so we wrote this one from the video. It will get you close, not an identical copy.

Make a 540-second animated explainer as one self-contained HTML file on a fixed 1920×1080 (16:9) stage.

What it shows: A nine-minute Chinese tutorial on making video with a language model, covering a seek(t) frame model, spring easing, frame versus time driving and deterministic render rules; the creator says the script, animation, music and edit were AI-made, with a cloned voice from VoxCPM2.

Look: clear diagrams and labels; strong editorial typography as a main design element; physically plausible simulation.
Structure: Follow a clear arc: hook question, 3 to 4 steps that each add one idea on screen, then a one-line takeaway. Keep every label legible for at least 2 s.
Build: Use plain HTML, CSS and JavaScript (canvas or SVG where it helps); no external assets unless I provide them.
Timing: drive everything from one master timeline exposed as window.seek(t) (t in seconds, 0 to 540), so it can be captured frame by frame at 30 fps with no dropped frames. Ease every move; nothing should pop in.
Finish: no scrollbars or overflow, everything stays inside the stage, and the last frame holds cleanly.

After the first version, I'll send notes on timing and detail; keep the same file and revise it.

Make one like this

  1. 1 · Copy the prompt

    Use the prompt above as your starting point, then change the subject and the style.

  2. 2 · Direct it like an editor

    Ask for one self-contained HTML file on a fixed 1920×1080 stage, link this video as the reference, and send two or three rounds of specific notes.

  3. 3 · Render it to MP4

    Screen-record a short piece, or capture it frame by frame for exact timing. How →

How it was made

The creator didn’t name their tools. Most videos like this are HTML, canvas or SVG animation written by Claude and screen-recorded or rendered to MP4. How that works →

Look

Skills that can make this

All skills →

Open-source skills for Claude Code that fit this video’s style and tools.

Views as of collection. Video and post © their creator; recreation prompt by Zryth.

Create with Zryth

Want a video like this for your brand?