world-execute-me-dsh-pv README

The project's README from github.com/MisakaZentai/world-execute-me-dsh-pv. As one markdown file: README.md.

world.execute(me); · 大肥鱼眼中的 world.execute(me)

preview

一支用代码逐帧渲染的 TUI 风格 PV,非官方同人作品:

这个仓库是这支片子的全部出片代码。你自备歌曲,一条命令就能从头渲染出来。

English: the complete source of a code-rendered, unofficial fan PV for Mili's "world.execute(me);". Every frame is a pure function of song time. On the left is a DeepSeek Harness chat window, rendered as per-frame HTML pages built from dsh's own CSS and screenshotted with Playwright. On the right is a TUI model-visualisation engine drawn with PIL. Bring your own copy of the song, then run python build.py all. The code is MIT; the artwork is CC BY-NC-SA 4.0; the song and lyrics are not included (see NOTICE.md).

快速开始

需要:

npm install
npx playwright install chromium
# 把歌曲放到 input/song.mp3(见 input/README.md)
python build.py all            # 检查 → 歌词 → 替身舞者 → 页面截图 → 渲染
python build.py all --4k       # 另出 3840×2160 版

60 帧导出(可选后处理)

生成无损 720p 母版后,可以用 FFmpeg 做运动补偿插帧,导出 1080p / 60 fps 或 2560×1440 / 60 fps 播放版,无需重新生成网页截图和角色缓存。方法、实测大小,以及文件大小与运行时内存的区别见 60 帧导出说明,其中提供 bash / zsh 和 Windows PowerShell 的复现命令。

不在仓库里的,以及替代办法

东西 为什么 替代办法
歌曲音频 Mili 的版权 自备,放到 input/song.mp3;build.py check 用 sha256 核对版本
歌词原文 歌词版权 build.py lyrics 从 LRCLIB 下载到你的本机,再和仓库里的逐词时间(不含文字)合成
原片的舞者画面 涉及第三方 MMD 模型和动作的使用条件;AI 参考与衍生画面分发权限尚未确认 构建时用鲸鱼娘立绘生成替身,所以你渲染出的片子在这些地方和原片不同
Consolas、微软雅黑等 Windows 字体 不可再分发 用本机的;其他系统可以换字体

重建出的片子和成片的差别

仓库结构

目录名保留了制作时的工作名,代码之间靠相对位置互相找到,请不要改名。

路径 内容
build.py 一键出片
tools/ 歌词合成(lyrics.py)、替身舞者(placeholder_h3.py)
data/ 不含文字的逐词时间、歌曲指纹、舞者帧的清单
film/pv_dsh_frontend_20260927/ 左边的 dsh 窗口:各组页面脚本 batch_*.py、截图脚本 seg_shot.mjs、合成与补丁 dsh_her.py、dsh_patch_*.py
film/tui_pv_world_execute_20260926/ 右边的 TUI 引擎:tuikit.py,第一版镜头 full/,第二版连续性和转场 continuity_full_v2/ 等
film/mmd_motion_eval_20260927/pv_full.py 舞者的时间表
film/third_party_references/ 鲸鱼娘立绘和表情(CC BY-NC-SA 4.0)
film/vendor/ dsh 前端(dsh-web-frontend)和 Cordis 组件包(dsh-client-ui-cordis)的 CSS 和 JS(MIT,DeepSeek)
docs/ 制作原理、字体、60 帧导出

许可

第三方清单见 NOTICE.md,一手授权来源和核实范围见 docs/ASSET_SOURCES.md。

分享视频

AI 使用情况

代码大部分由 AI 编写(Claude Code、Codex)。总体方案、审美取舍和逐批审片由作者完成。

用途 工具 用量(大致)
写代码:画面引擎、dsh 界面页面、分段实现、出片脚本(主力) Claude Code(Claude Opus) Claude Pro 订阅额度内,没开超额计费,没走 API 按量付费;token 总量没有记录
写代码:逐词歌词时间对齐、整片导演审核 Codex(GPT-6 Astra) Pro 订阅额度内,没走 API;token 总量没有记录
舞者视频(定稿) MiniMax-H3,本地 ComfyUI 约 31 次生成,约 1.7 GPU 小时,本机 RTX 3090,没有费用
舞者视频(试过,没进定稿) Grok Imagine(网页) 提交 31 次(2 张图 + 29 条视频),订阅额度制;单条约占周额度 1%(作者观察,不是计费公式)
舞者视频(试过,没进定稿) Seedance 2.0 Fast(Updream 网页) 2 条试片
音频分析(本地,只做对齐,不生成声音) Demucs(HDEMUCS_HIGH_MUSDB_PLUS)分离人声;MMS_FA 做 CTC 对齐;Whisper medium.en + stable-ts 独立对齐 —

署名

这是非官方同人作品,与 DeepSeek、Mili 没有从属或合作关系,也未经他们认可。

world-execute-me-dsh-pv (MisakaZentai/world-execute-me-dsh-pv)

art
Searching Database…

What is this?The complete source code of an unofficial fan PV (music video) for Mili's song 'world.execute(me);', where every frame is generated by code as a function of song time. The left half shows a DeepSeek Harness (dsh) chat window, the right half a TUI visualization of the model's world.
Does the repo include the finished video, or just the code that renders it?Just the code — the repo is the full build pipeline, not the film. Given your own copy of the song, it outputs out/film.mp4 at 1280×720, a lossless master at out/filmmaster.mp4, and out/film4k.mp4 if you add --4k.
How do I render the full film myself?Install the Python 3.12+ requirements (pip install -r requirements.txt), plus Node 20+ and ffmpeg on PATH, then run python build.py all. Each step can also be run individually, e.g. lyrics or check.
What do I need to provide that isn't in the repo?The song audio: you place your own copy at input/song.mp3 (build.py check verifies it by sha256). Lyrics are downloaded from LRCLIB at build time, Windows fonts like Consolas can be swapped on other systems (docs/FONTS.md), and the original dancer is replaced by a code-drawn whale-girl placeholder.
Will my render look exactly like the published video?Close but not pixel-identical: the dancer is a placeholder, checkpoint numbers differ at 12.5-29 s, three frames near 123.5 s differ in send-button color, plus minor unverified brightness deltas. Verification covered the original Windows build (4968 screenshots); fresh installs on other systems are unproven.
How long does rendering take on CPU, and does a GPU help?A CPU render of the full film takes roughly 10-30 minutes depending on core count. With a CUDA-enabled torch build, adding --gpu --python makes it much faster.
Is this an official DeepSeek Harness (dsh) project?No — it's an unofficial fan tribute, not affiliated with or endorsed by DeepSeek or Mili. The chat window on the left is rebuilt frame by frame from dsh's own MIT-licensed CSS and JS (vendored in film/vendor/) and captured with Playwright screenshots.
If I render my own copy, can I post it online?Yes for personal, non-commercial sharing under Mili's fan-work guidance, but you must clearly label that the video contains AI-generated content, keep the full attribution chain and license links, and confirm separately for commercial uses like sponsorships or paid screenings.
How was the film actually made?Most of the code was AI-written (Claude Code with Opus as the main tool, Codex for lyric timing and review) under the author's direction and shot-by-shot review. The original dancer frames came from MiniMax-H3 via local ComfyUI on an RTX 3090; audio alignment ran locally with Demucs, MMSFA and Whisper.
Can I get a 60 fps version without re-rendering everything?Yes — once the lossless 720p master is built, FFmpeg motion-compensated interpolation exports 1080p/60 or 1440p/60 playback copies without regenerating the web screenshots or character cache. Ready-made bash/zsh and PowerShell commands are in the 60fps doc.
Size
5.8 MB
Stars
591
Updated
CommitOct 06
Age
18d
Commits · 90 Days