AI Videos 只靠 Diffusion，目前似乎不太够

Claude Opus 5.5 在社区的表现属实惊人，加上我额度属实用不完（比较无聊说是），因此我便打算使用目前支持多模态的 AI 来生成一些好玩的视频。当然自己调的比如 MinimaxH3 生成的视频也会放这里。

答卷

Claude Opus 5.5 Medium2026/10/03

Freestyle

自由发挥：798 个钢琴音符变成 798 片雪花。

4

Claude Opus 5.5 Medium2026/10/03

Self Introduction

卡通风格的五分钟自我介绍，角色全靠矢量手绘保持一致。

8 → 9.5

DeepSeek V4.1 Flash Middle2026/10/03

Self Introduction

同一道自我介绍题，DeepSeek V4.1 Flash 的答卷。

2.8

Gemini 3.8 Flash Medium2026/10/03

Self Introduction

基于 Google Gemini 四芒星标志生成的矢量卡通形象，结合 IndexTTS 声音克隆与程序化动效的五分钟自我介绍。

3.8

AI 自评 8.2 · 高估 4.4

gpt-6-astra medium2026/10/03

Self Introduction

用标志启发的卡通角色、英文旁白和原创配乐完成五分钟自我介绍。

5

Step 5 Preview2026/10/03

Self Introduction

五分钟卡通自我介绍：代码画出的阶梯吉祥物，本地 TTS 配音，MiniMax H3 参考图驱动动画。

4

评分纯主观，满分理论 10 分，未来可能超出 10 分。

考场环境

角色	配置
LLM	Claude Pro Subscription
LLM	DeepSeek Official API
本地算力	PC with RTX 5060 Laptop（8GB VRAM，跑 ComfyUI）

通用素材说明

每道题都会附带下面这段素材说明（各篇 Prompt 里保留了原文）。

Assets所有题目共用

here are the assets I can provide for u:
  1. Audio
     /home/aaaa0ggmc/Music/CloudMusic/ has over 2k songs for u to pick
     u can also generate sound effects urself

     If cockpit(MCP) is on, you can connect to AIDJ and filter / search songs with lyrics or emotion tags(emotion tags are not quite stable, mixed of a sort of languages)
  2. Image / PresetVideos
      http://127.0.0.1:8188 ComfyUI provides u with the way to generate some pieces
     The root folder is ~/Apps/ComfyUI
     Image models are Z-Image, QwenImage...
     Video Models are MinimaxH3(use the int8conv version)
     do remind to free the VRAM cause I have 8GB only
    It is suggested to generate a piece of video with length lower than 8s and solution less than 0.4.
    Generating a pre-built video is costing in my computer.
    And these diffusion models are not quite capable of producing "fixed" identity, which needs yourself to generate the picture.

快速让AI回填：

Fill (~/Projs/Blog/docs/keep_learning/ai_vid/) with our turns and timecost, see ~/Projs/Blog/guides/AI_VID_WRITING_GUIDE.md for some instructions.