代码驱动动态视频 · 3B1B 风格讲解 · 视觉自纠错 · K3 vs Gemini 消融
Kimi K3 的「视频自动化」= 作为编程 Agent 编写代码，把概念转化为动画图示、转场与信息可视化，并借助原生多模态在「代码 ↔ 渲染画面」间迭代打磨。本页所有画面均为模型生成代码、逐帧渲染合成的真实视频（非静态图、非文生视频）。所有 Kimi 结果均由 Kimi K3 官方 API（api.kimi.com/coding，模型 k3，思考开启）跑出，已核验含推理块，未混入其它模型；对照方为 Gemini 3.1 Pro。将鼠标悬停即可播放。
K3 的招牌能力：把一个技术概念做成多场景动态讲解视频（像它给自己架构做的 3Blue1Brown 风格讲解）。给一句需求，K3 写出自包含 HTML，含场景时间轴与转场，逐帧渲染成视频。以下每个概念都用完全相同的提示词跑 K3 与 Gemini 3.1 Pro 对比。
Create a 15-second, 1920x1080 3Blue1Brown-style ANIMATED EXPLAINER of a Mixture-of-Experts (MoE) layer. 3 timed SCENES with a fade transition between each: Scene 1 (0-5s) "Router": tokens flow from the left into a Router block. Scene 2 (5-10s) "Top-k Routing": the router lights up lines from a token to its top-4 of 16 expert blocks. Scene 3 (10-15s) "Combine": the 4 experts pulse and their weighted outputs merge; show y = Σ gᵢ·Eᵢ(x). Dark background (#0b0b12), deep-blue/yellow 3B1B palette, clean typography, persistent title "Mixture of Experts" + an updating scene caption. Driven by requestAnimationFrame timestamp.
Create a 15-second, 1920x1080 3Blue1Brown-style ANIMATED EXPLAINER of self-attention. 3 timed SCENES with a fade transition between each: Scene 1 (0-5s) "Q, K, V": a row of word tokens, each emitting three small vectors labelled Q, K, V. Scene 2 (5-10s) "Scores + Softmax": an attention grid fills by value (cell = Qᵢ·Kⱼ), then one row normalizes into bars summing to 1. Scene 3 (10-15s) "Weighted sum": the focus token pulls colored V vectors weighted by attention; show "softmax(QKᵀ/√d)V". Dark background, blue/yellow palette, clean typography, persistent title "Self-Attention" + updating caption. Fully requestAnimationFrame-timestamp driven.
Create a 15-second, 1920x1080 3Blue1Brown-style ANIMATED EXPLAINER of a diffusion image model. 3 timed SCENES with a fade transition between each: Scene 1 (0-5s) "Noise": a field of random static noise, title "Diffusion Models". Scene 2 (5-10s) "Denoising": the model removes noise step by step and a crisp shape emerges (show a step counter t=T→0). Scene 3 (10-15s) "Text guidance": a text prompt token stream steers the result toward the final image. Dark background, tasteful gradient accents, per-scene captions, persistent title. Fully requestAnimationFrame-timestamp driven.
社区里 K3 最受关注的动态内容是 3D 动画解说与仿真（如 generative-ai.review 的古埃及工程 3D 解说系列、Windows XP 仿真、System-2-Arena 游戏推演）。这里复现两类：伪 3D 工程解说与轨道仿真，同提示词跑 K3 与 Gemini。
Create an 8-second, 1920x1080 animated 3D-style "explainer" of how ancient Egyptians raised a heavy stone block using a water-lock / rising-raft method — in the spirit of a 3Blue1Brown-meets-Kurzgesagt engineering explainer. A pseudo-3D scene (isometric) on a sandy ground plane, a large stone block on a wooden raft inside a rectangular stone lock. Over 8s animate: (1) water rises, (2) raft + block float upward, (3) it's nudged onto a ledge. Depth cues: shaded block faces, soft shadow, parallax. Title "Water-Lock Lifting" + an updating stage caption. Warm desert palette. requestAnimationFrame driven.
Create an 8-second, 1920x1080 animated "explainer" of the inner solar system (Sun + Mercury, Venus, Earth, Mars) as an elegant orbital simulation. The Sun glowing at center; four inner planets on orbits, each at a DIFFERENT correct-relative speed (Mercury fastest), positions driven by the timestamp. Faint orbit rings, planet labels, Earth with a small moon orbiting it. Title "Inner Solar System" + a live day counter. Dark space background with subtle stars, glow on the Sun. Deterministic requestAnimationFrame motion.
给一句需求，模型输出自包含 HTML 动画并渲染成视频。以下每支同提示词跑 K3 与 Gemini。
Create a 6-second, 1920x1080 "3Blue1Brown-style" motion-graphics explainer of Mixture-of-Experts top-k routing. Show: a grid of 16 "expert" boxes labeled, a stream of tokens flowing in from the left, and for each token highlight the top-4 experts it routes to (draw connecting lines that light up). Include a title "Top-k Routing (k=4, 16 experts)". Use a dark background (#0b0b12), clean typography, smooth eased transitions. Deep blue / yellow palette like 3Blue1Brown. Animation must be fully driven by requestAnimationFrame timestamp so it renders deterministically.
Create a 6-second, 1080x1920 (vertical) animated visualization of bubble sort on 12 bars. Bars start shuffled; over the 6 seconds they sort ascending with the comparisons highlighted (compared pair turns orange, swap animates smoothly). Show a title "Bubble Sort" and a live step counter. Dark background, vibrant bars. Motion fully driven by requestAnimationFrame timestamp (t in ms). At t=0 fully shuffled, by t=6000ms fully sorted.
Create a 5-second, 1080x1920 kinetic-typography opening title for a short drama titled "逆袭：她回来了" (Chinese text must render). The title words fly in one by one with easing, a subtle particle/glow background pulses, and it ends on a held composed frame with a small subtitle "第一集". Cinematic, high-contrast, gold-on-black. Motion fully driven by requestAnimationFrame timestamp (t ms).
面向短剧生产的高频动态物料，规格取自短剧设计惯例（安全区、字号占比、类型化配色、缓动）。同提示词跑 K3 与 Gemini。
1080×1920 竖屏短剧开场片头，3 秒。标题「豪门第一日：她不再退让」用金属渐变金字（#FCE9A0/#D4AF37/#8A6D1F）配深色描边与投影，字高约屏宽 13%（大）。副标题「第一集」金/白。标题在约 0.7 秒内缩放淡入（cubic-bezier(0.05,0.7,0.1,1)），副标题延迟约 0.15 秒错位跟进，约第 1 秒一道白色流光扫过标题，随后从 1.3 秒定格到结尾。近黑背景，全部文字在安全区内。
1080×1920 竖屏 6 秒三段式：(a) 0-1.5s 章节标题「三年后」居中，配一条金色横线 scaleX 0→1 展开（easeOutExpo）；(b) 约 2.0-2.4s 一道金色色块擦除转场 + 1 帧白闪；(c) 2.6-6.0s 下三分之一字幕条从左滑入（power3.out），主名「顾沉舟」68px 粗体 + 身份行「顾氏集团总裁」40px 80% 不透明，置于半透明黑条上，白字+投影，保持到结尾。近黑背景，全部在安全区内。
1080×1920 竖屏人物关系图，5 秒。5 个圆形头像节点（直径≥160px，中文名居中且不裁切、不重叠）：林晚 / 顾沉舟 / 苏曼 / 林父 / 陈默。带标签连线：林晚—夫妻—顾沉舟，苏曼—情敌→林晚，林父—父女—林晚，陈默—下属—顾沉舟，苏曼—合作—林父。节点错位缩放弹入（每个约 0.1s），随后连线描绘出现，第 3 秒起定格。短剧主题色（金/红/深蓝），全部在安全区内。
把渲染画面回传给模型让它自评并修正，再渲染，测「看着自己的成品越改越好」。共 4 组判别性任务：人物关系图、金色片头、剧情时间线、播放数据看板。典型收敛：人物关系图 K3 8→10→10；剧情时间线 Gemini 8→10（首轮节点偏多，看图后修正）。下方为两者各自 3 轮迭代后的最佳成片视频。
1080×1920 竖屏人物关系图，5 秒。5 个圆形头像节点（直径≥160px、中文名不裁切、彼此不重叠）：林晚/顾沉舟/苏曼/林父/陈默，含 5 组带标签关系线；节点错位弹入后连线描绘，第 3 秒起定格。要求近黑背景+主题色、全部在安全区内。〔循环：写HTML→渲染截图→回传自评→输出修正版，共 3 轮〕
1080×1920 竖屏短剧开场片头，3 秒，标题「逆袭：她回来了」金属渐变金字，副标题「第一集」，标题缩放淡入、副标题错位跟进，随后定格；近黑背景，全部在安全区内。〔3 轮 vision-in-the-loop 迭代〕
1080×1920 竖屏短剧剧情时间线，6 秒。一条纵向主轴，6 个按时间排列的事件节点（圆点+年份标签+中文事件短句）：三年前·林晚被逐出家门 / 两年前·隐姓埋名创业 / 一年前·公司崭露头角 / 半年前·与顾沉舟重逢 / 三月前·揭穿苏曼阴谋 / 如今·强势回归豪门。主轴自上而下生长，节点错位弹入，字幕随后淡入，t=4.5s 起定格。标题「逆袭时间线」，全部在安全区、不拥挤、不裁切。〔3 轮 vision-in-the-loop 迭代〕
1080×1920 竖屏短剧播放数据看板，6 秒。标题「《逆袭：她回来了》· 播放数据看板」；4 张 KPI 卡片带数字滚动动画（累计播放 1.2亿 / 完播率 68% / 点赞 340万 / 追剧人数 89万）；下方「近7日播放趋势」7 根柱状图带周一…周日标签。数字滚动、柱高增长均由 requestAnimationFrame 驱动，t=4.5s 起定格；卡片对齐不重叠、柱子共基线、全部在安全区。〔3 轮 vision-in-the-loop 迭代〕
| METRIC | KIMI K3 | GEMINI 3.1 PRO |
|---|---|---|
| 3B1B 讲解视频完整度 | 3/3 一次成功、代码完整（默认配置） | 限制 thinkingBudget 后 3/3 一次完整；默认扩展思考会耗尽 token 致代码截断 |
| 困难任务首轮命中 | 8/10 | 10/10 |
| 多轮视觉自纠错 | 8→10→10（逐轮改进） | 10→10→10（首轮即满分） |
| 单支视频生成延迟 | 约 60–110 秒 | 约 74–174 秒 |
| 画面视觉理解 | 准确识别自渲染帧 | 准确识别自渲染帧 |
| 端点稳定性（观察） | 复杂任务易触发 300s 网关超时，需约束「勿过度思考」 | 稳定 |
官方标价，按每百万 token（USD）。K3 来自 kimi.com/blog/kimi-k3；Gemini 3.1 Pro 来自 ai.google.dev/gemini-api/docs/pricing（≤200k tokens 档；>200k 为 $4/$18）。K3 的最大优势是缓存命中输入极便宜（$0.30/M，官方称编码场景缓存命中率 >90%）。
| 计费项 | KIMI K3 | GEMINI 3.1 PRO |
|---|---|---|
| 输入 · 缓存未命中 | $3.00 / M | $2.00 / M |
| 输入 · 缓存命中 | $0.30 / M | （隐式缓存，命中约 $0.20 / M） |
| 输出（含思考 token） | $15.00 / M | $12.00 / M |
| 场景 | KIMI K3 | GEMINI 3.1 PRO |
|---|---|---|
| 一支 3B1B 讲解视频（≈600 输入 / 2.5–2.9k 输出） | ≈ $0.045 | ≈ $0.031 |
| 一轮 Vision-in-the-Loop（含回传截图 ≈1.1k tok + 规格） | ≈ $0.31（输出常达 18–25k tok） | ≈ $0.039（输出 ≈3k tok） |