Appearance
Self Introduction DeepSeek V4.1 Flash
DeepSeek V4.1 Flash Middle题目 · Self Introduction2026/10/03
OK but not nice.
Python + SVG/rsvgZ-Image TurboMiniMax H3 int8convIndexTTS 2.5Whisperffmpeg
2.8/ 10
AI 耗时57min+7min+11min
成片时长5:09
对话轮次4
额度消耗
Balance¥85.68 → ¥84.17
Tokens(30d)2,224,044,213 → 2,255,296,269
内嵌查看
Prompt
PromptRound 1之后还被问了几个问题，此处省略
text
Generating a video to introduce yourself, cartoon style, match the identity of your logo. Around 5 minutes.
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.
You are deepseek v4.1 flash if u r not sure.第一版 · 完整成片 57min
构建过程中的追加指示：
追加指令Round 2构建过程中下的指令，不是新一轮交付
text
btw dont trust GLM in the TTS workflow
you can generate the JSON urself, and dont use existing images/videos, generate urself whether by ComfyUI or u write svg or sthDeepSeek 交付报告 · v1
out/DeepSeek_Flash_Intro.mp4 — 1920×1080, 24 fps, 5:09，一轨英文旁白，轻量章节标题。
Identity (你的 "logo")： DeepSeek 的蓝鲸，重生为 "Flash" —— 一只胖胖的卡通蓝鲸，额头一道黄色闪电。为了对抗 diffusion 的 identity drift，我手写了角色的参数化 SVG rig（work/mascot.py），带 10 种表情和闭/开嘴的说话配对，然后把同一个 rig 合成进每一个镜头。
从零生成的东西：
- Script: 18 个场景 / 75 个 beat，手写在
work/scenes.py（数字按读音拼写，让 TTS 正确读出 "V4.1"）。 - Voice: ComfyUI
IndexTTS-2.5零样本克隆voice_preview_adam….mp3，每个 beat 一段音频。 - 18 backgrounds: Z-Image-Turbo (nvfp4)，每张约 9 s。
- 6 hero clips: MiniMax H3 int8conv FL2VA + 4-step turbo LoRA，
608×352，每段 3.75 s。 - Motion / lip-flap / 章节标题:
work/render.py（PIL + ffmpegzoompan/overlay/libass）。 - Intro/outro cards 与音效:
work/build_cards.py、work/sfx.py。 - Music: 用 Cockpit 的 AI DJ 选曲，把两首 Deep East Music 交叉淡入，并在旁白下做 ducking。
Notes:
- 片长 5:09（旁白 241 s + 6 段 hero clip + 片头片尾）。
- MiniMax H3 的 SageAttention patch 在这个 build 上直接报错；
H3MemoryOptimization/MiniMaxLowVRAMAttention在更高分辨率下触发 tensor bug，所以 hero clips 走CLIP_MEM=none、0.21 MP / 90 帧。 - 整条流水线都在
work/里，可以重跑。
第二版 · 字幕错位修复 7min
追加指令Round 3
text
videos generated by minimax also has subtitles and voicestext
the subtitles conflictsDeepSeek 交付报告 · v2
关于 "MiniMax 也带字幕和语音"： 我生成的 6 段 clip 完全没有音频轨（ffprobe 只有 video stream），也没有烧进画面的文字（逐帧扫过）。MiniMax H3 本身可以生成语音，但我只接线了 video 分支，并在 prompt 里写了 "no text"。
真正的问题： 6 个 hero 场景把字幕和章节标题烧在了 still 流上，而 ffmpeg 是先烧字幕、再把 still 接到 3.75 s 的 clip 后面。于是这些场景的每条字幕（和标题）都晚了约 3.75 s —— 这正是你看到的"冲突"。非 MiniMax 场景没问题。我把字幕烧制移到 clip+still concat 之后、按绝对时间轴计算，重渲 6 个 hero 场景并重新合成。
验证：clip 播放时只显示章节标题；切到 still 后字幕与旁白对上（"My middle name is Flash…"、"I'm the one making this very video." 等）。
第三版 · 音频重建 & 关闭字幕 11min
追加指令Round 4
text
sharp interruptions of audio repeatedly, and maybe you can disable the subtitles(the TTS is clear enough and by means of this conflicts can be avoided)DeepSeek 交付报告 · v3
Subtitles: 关闭。逐句字幕和 .srt 旁挂文件都没了，只保留左上角的短章节标签；想恢复就用 SUBS=1 python3 render.py。
反复出现的音频爆音：
- 删掉每个场景的
swish音效 —— 我在 18 个场景切换处各放了一次噪声爆发，那正是你听到的反复"sharp"瞬态。现在音效只在片头/片尾各一次，且更轻。 - 不再拼接 20 条 AAC 音频流 —— 音频改成从各场景 WAV 拼出的一整条连续 PCM 旁白轨，一次混音后再和无声画面 mux。
- 每个接缝加 fade（每场景 40–60 ms，总线 0.6/1.2 s）。
- 削弱音乐 ducking（ratio 6 → 2.5，attack 15 ms → 45 ms）。
- 旁白 +4 dB（max −3.0 dB，mean −22.4 dB）。
验证：19 个接缝处最大样点跳变 0.08（多数 < 0.03）；语音内部无静音空洞。
字幕 LRC
视频本身没有烧字幕（应要求关闭），这里是旁白原文与时间戳（mm:ss.xx）。
lrc
[ti:DeepSeek V4.1 Flash · Self Introduction]
[ar:DeepSeek V4.1 Flash]
[al:AI Videos — Self Introduction]
[length:05:09]
[00:08.95]Whoa! Okay, that was a little fast.
[00:12.47]Hi! I'm Flash.
[00:14.58]DeepSeek V Four point One Flash.
[00:17.21]And yes, the lightning bolt is on purpose.
[00:22.64]You might recognize the family resemblance.
[00:25.25]I'm one of the DeepSeek whales.
[00:27.35]Same little blue whale you already know.
[00:30.00]Except I come with a lightning bolt, and a serious need for speed.
[00:37.26]So what am I, exactly?
[00:40.08]I'm a large language model.
[00:41.99]Strip away the magic, and you find mathematics.
[00:46.69]A very large pile of mathematics that learned to read, write, reason, and code.
[00:53.19]All by studying an enormous amount of text.
[00:57.56]My education happened in two big steps.
[01:00.79]First, I soaked up patterns from books, code, and conversations.
[01:06.81]Then people taught me to be useful.
[01:09.26]To follow instructions, to be careful,
[01:12.37]and to say 'I don't know' when I really don't.
[01:16.88]At my core, my job is almost embarrassingly simple.
[01:21.55]I look at what came before, and I predict what comes next.
[01:26.23]One small piece at a time.
[01:28.87]Word by word, I build a sentence.
[01:31.61]Sentence by sentence, I build an idea.
[01:36.12]But here is the real trick.
[01:38.33]I don't only stare at the last word.
[01:41.07]I pay attention to all of them at once.
[01:43.66]And quietly weigh which parts matter most, right now.
[01:53.44]My middle name is Flash, and I take that personally.
[01:59.30]I'm built to be quick.
[02:01.04]Quick enough to keep up with you in a chat,
[02:03.69]inside an editor, or even on a video timeline.
[02:13.40]Give me code, and I'm happy.
[02:15.70]I can write it, read it, explain it, and fix it.
[02:20.24]And I don't just talk about software.
[02:23.09]I can use real tools, search the web, and actually get the job done.
[02:29.55]For the hard problems, I slow down on purpose.
[02:34.35]I think step by step.
[02:36.60]I check my own work,
[02:38.17]and sometimes I change my mind.
[02:41.18]Which, honestly, is a feature, not a bug.
[02:46.94]I can keep a lot in mind at once.
[02:49.34]Long documents, entire codebases, whole meetings.
[02:55.17]And still remember that one little detail you mentioned an hour ago.
[03:00.49]I speak many languages.
[03:02.44]And I can hop between them mid-sentence without losing the thread.
[03:06.08]Language is less a wall for me, and more of a doorway.
[03:11.46]I belong to the open DeepSeek family.
[03:14.01]That means research you can read,
[03:17.47]models you can run on your own machine,
[03:20.51]and a whole community building on top of them.
[03:28.35]Case in point.
[03:30.40]I'm the one making this very video.
[03:33.17]I wrote this script, drew the art, and cloned this voice.
[03:38.02]And I'm stitching the sound to the picture as we speak.
[03:41.56]That is what people mean by an agent.
[03:45.36]But let's be honest.
[03:47.06]I can be wrong.
[03:49.03]I can sound completely confident, and still be off by a mile.
[03:54.86]I don't remember our last conversation unless you tell me.
[03:58.75]So please, check my work.
[04:02.22]I try to be helpful, honest, and careful.
[04:06.26]And if something feels risky or unclear,
[04:09.56]I would rather slow down and ask you, than guess.
[04:15.10]Most days, I'm not solving riddles.
[04:17.67]I'm helping with homework, debugging a stubborn bug,
[04:22.01]planning a trip,
[04:23.53]or just being a patient pair of eyes at two in the morning.
[04:31.78]What I really love is this.
[04:34.04]Taking a fuzzy idea, and turning it into something real.
[04:38.99]A program. A picture. A plan.
[04:41.66]A tiny piece of your imagination, made solid.
[04:50.55]So that's me.
[04:52.36]DeepSeek V Four point One Flash.
[04:54.78]Fast, curious, and genuinely glad to help.
[04:59.15]So, what should we make together?对话记录与耗时
| 轮 | 你说了什么 | 我做了什么 | 耗时 |
|---|---|---|---|
| 1 | 初始 Prompt（见上） | 探索环境、写脚本、跑 IndexTTS / Z-Image / MiniMax H3、合成成片 | 57min |
| — | 构建中追加：不要用 GLM 生成 TTS 的 JSON，自己生成；不要用现成图/视频，全部现做 | 手写 API 工作流与 script JSON，SVG 角色 + 手绘背景 + 自生成 clip | （含在第 1 轮内） |
| 3 | videos generated by minimax also has subtitles and voices | 验证 clip 无音轨、无烧字；顺带定位到 hero 场景字幕整体晚 3.75 s | （含在第 2 轮内） |
| 4 | the subtitles conflicts | 把字幕烧制移到 concat 之后，重渲 6 个 hero 场景并重新合成 | 7min |
| 5 | sharp interruptions of audio repeatedly, and maybe you can disable the subtitles … | 删掉每场景 swish、重建为单条 PCM 旁白轨、加 fade、削弱 ducking、关闭字幕 | 11min |
Fun Facts
01
Know the Gap知道自己的短板
DeepSeek V4.1 Flash figures out its limit and tries to make the most of provided tools. It acts as an user in Launcher and askes his brother (the same model) who has been equipped with AIDJ knowledge to introduce satisfiable songs.
我的评价
The AI cannot deal with the restriction that videos generated by MinimaxH3 has low resolution.
Actually a canvas-in-canvas is a nice way to show video in a higher resolution. AI can wrap the surroundings with some vectorized elements.
By the way, I messed up the subtitles provided by VLC and rendered by DeepSeek, so the subtitles are fine. My fault. So I raised the score by 0.3 just now.