Help center帮助中心

Getting started

1. Install

• Windows — run the setup.exe and follow the steps. Pick the CUDA build for NVIDIA cards, Vulkan for AMD or Intel.
• macOS — open the .dmg and drag changji into Applications. The app is signed and notarized by Apple; the first launch checks this online.
• Linux — unpack the .tar.gz and run changji-desktop.sh.

Not sure which build to take? The download page explains the options.

2. Choose models

Model weights aren't bundled with the app. On first launch changji asks which models to use for writing, images, video and voice, checks what fits your graphics memory, and downloads them — from ModelScope in mainland China and Hugging Face elsewhere. You can change the choice later in Settings.

If a remote GPU machine will do the rendering, you can skip the large image and video downloads here — the remote machine needs them instead (see Remote GPU).

Your first film

1. Start a new conversation and describe the film — a sentence, a synopsis, or paste a chapter of a novel.
2. changji writes a story outline and asks you to confirm it. Tell it what to change in plain words.
3. It writes the chapter, adapts it into a script and breaks it into a storyboard. Open the panels on the right to see the story, the cast, the script and every shot.
4. It draws reference images for the characters and locations, then first frames, then video, voiceover and subtitles for each shot.
5. Watch the chapter when it's done. To judge the direction early, ask for just the first few minutes.

Not happy with a shot? Say so, or edit it directly — only that shot is redone.

Render on a remote GPU

No GPU on this computer, or not a fast enough one? Run changji on a machine that has one — a workstation of your own, or a GPU server rented by the hour — and link the two. Your projects stay on this computer; the rendering goes to the remote card and the results come back.

On the remote machine

1. Install changji. On a Linux GPU server, take the CUDA engine package from the download page (changji-linux-x64-cuda.tar.gz), unpack it and install ffmpeg; changji --doctor names anything else that's missing. On a Windows or Mac computer, install the desktop app as usual.
2. Start it so other machines can reach it: changji --host 0.0.0.0 --port 8080. It prints its addresses and its access token. The desktop app listens on the network by itself; its address and token are under Settings → Network → This machine's address and token.
3. Open the printed address in a browser once and download the image, video and voice models there. The remote machine does the rendering, so it's the one that needs them.
4. Make sure your computer can reach that port. On a rented server, open it in the provider's firewall or security group — ideally for your own IP address only.

On your computer

1. Open Settings → Network and, under Add machine, enter the remote address (for example http://203.0.113.7:8080) and its token.
2. The machine joins the list and shows what it can do. Choose which steps it takes — video only, for example — and work starts going to it.

Only what each job needs is sent to the remote machine — a shot's reference images and first frame, for example — and nothing passes through us. If you rent the server by the hour, shut it down when the chapter is done.

快速上手

1. 安装

• Windows：运行 setup.exe 按提示安装。NVIDIA 显卡选 CUDA 版，AMD 或 Intel 显卡选 Vulkan 版。
• macOS：打开 .dmg，把场记拖进「应用程序」。程序已经过 Apple 签名和公证，首次打开时会联网校验。
• Linux：解压 .tar.gz，运行 changji-desktop.sh。

不确定下载哪一个？下载页上有说明。

2. 选择模型

模型权重不随程序附带。首次启动时，场记会让你为写作、出图、出片、配音分别选择模型，按你的显存判断能否装下，然后自动下载——中国大陆从魔搭（ModelScope）下载，其他地区从 Hugging Face 下载。之后可以在设置中更改。

如果打算让远程的 GPU 电脑出片，本机可以跳过出图、出片这些大文件的下载——要下载的是远程那台（见远程 GPU）。

第一部片子

1. 新建一条对话，描述你想拍的片子：一句话、一段梗概，或者直接贴一章小说。
2. 场记写出故事大纲并请你确认。想改哪里，直接用大白话告诉它。
3. 它写出这一章，改编成剧本，再拆成分镜。打开右侧的面板，可以查看故事、角色、剧本和每一个镜头。
4. 它为角色和场景画参考图，然后逐镜出首帧、出片、配音、生成字幕。
5. 完成后观看这一章。想先判断方向，可以让它只做前几分钟。

对某一镜不满意？直接说，或者手动修改——只会重做那一镜。

在远程 GPU 上出片

这台电脑没有显卡，或者显卡不够快？在一台有显卡的机器上也跑一份场记——你自己的工作站，或者按小时租的 GPU 服务器——两台互联。片子留在这台电脑上，出片交给远程那张卡，结果再传回来。

在远程那台上

1. 安装场记。Linux GPU 服务器从下载页取 CUDA 版命令行引擎（changji-linux-x64-cuda.tar.gz），解压后装上 ffmpeg；还缺什么，changji --doctor 会说。Windows 或 Mac 电脑照常安装桌面端即可。
2. 让别的机器连得上它：changji --host 0.0.0.0 --port 8080。启动后它会打出自己的地址和访问口令。桌面端本来就在网络上监听，地址和口令在设置 → 互联 → 本机地址与口令里。
3. 在浏览器里打开它打出的地址，把出图、出片、配音模型下载到那台上。出片由远程那台来做，模型要装在它那儿。
4. 确认你的电脑连得到那个端口。租来的服务器要在服务商的防火墙或安全组里放开它，最好只对你自己的 IP 放开。

在你的电脑上

1. 打开设置 → 互联，在添加机器里填远程那台的地址（例如 http://203.0.113.7:8080）和口令。
2. 那台机器出现在列表里，并显示它能做什么。选好让它做哪些步骤（比如只出片），任务就会派过去。

传给远程那台的只有每项任务需要的内容（例如一镜的参考图和首帧），不经过我们。按小时租的服务器，这一章做完就可以关掉。

More machines

Every computer running changji can take work from the others. With LAN discovery on, machines on the same network show up in Settings → Network automatically; to add one by hand — on the same network or a remote one — enter its address and access token under Add machine. A computer with several GPUs runs one worker per card on its own, so add each computer only once.

You can choose, per machine, which steps it's allowed to do — for example, only video.

Web interface

The engine also serves a web interface. Start it with changji --host 0.0.0.0 --port 8080 and open the address it prints — it includes the access token — in any browser on your network, including a phone.

Your files

• Each film is a folder. Copy it to back it up or to continue on another computer.
• Finished chapters and the full film are in the film's output folder, subtitles in subtitles.
• Every prompt sent to a model and its reply are logged in the llm_log folder of the data directory, so you can see exactly what was asked.
• The app sends anonymous usage statistics — never prompts, stories, file names or keys. Settings → General → What gets sent shows the exact text, and you can turn it off there. See the privacy policy.

Troubleshooting

"This CPU isn't supported"

The x86 builds need AVX2 (Intel Haswell from 2013 or newer). In a virtual machine, set the CPU model to host-passthrough.

A render fails with out of memory

Pick a smaller canvas for the film, close other programs that use the GPU, or hand video to a remote GPU machine with more memory.

Windows shows a SmartScreen warning

If SmartScreen says the app isn't recognized, choose More info → Run anyway. Only download changji from this site or our official releases page.

A remote machine shows as unreachable

Open its address in a browser on this computer. If that doesn't load, the port isn't reachable — on a rented server, open it in the firewall or security group. If it loads, check that the token you entered is the one the remote machine printed.

A remote machine is connected but gets no work

The machine list shows what it can do. If a step is missing, its models aren't downloaded there yet — open its address in a browser and download them. Also check which steps you allowed it in Settings → Network.

Contact us

Didn't find the answer? Get in touch.

多台机器

每台运行场记的电脑都可以接收其他电脑派来的任务。打开局域网发现后，同一局域网内的机器会自动出现在设置 → 互联里；手动添加时（同一局域网的也好，远程的也好），在添加机器中填写它的地址和访问口令。多显卡的电脑会自动为每张卡启动一个工作进程，所以每台电脑只需添加一次。

可以为每台机器单独设定允许做哪些步骤，比如只负责出片。

网页界面

引擎自带网页界面。用 changji --host 0.0.0.0 --port 8080 启动后，在局域网内任意设备（包括手机）的浏览器中打开它打印出的地址（地址中已包含访问口令）即可。

文件在哪儿

• 每部片子是一个文件夹。复制它即可备份，或者在另一台电脑上继续。
• 成片的各章和整部电影在片子的 output 文件夹中，字幕在 subtitles 中。
• 发给模型的每条提示词和回复都记录在数据目录的 llm_log 文件夹中，可以准确查看问了什么。
• 程序会发送匿名使用统计，从不包含提示词、故事、文件名和密钥。设置 → 通用 → 会发送的内容里能看到原文，也可以在那里关闭。详见隐私政策。

常见问题排查

提示「CPU 不受支持」

x86 版本需要处理器支持 AVX2（Intel Haswell / 2013 年以后）。在虚拟机中运行时，请把 CPU 型号设为 host-passthrough。

出片时显存不足

为片子选择更小的画幅，关闭其他占用显卡的程序，或者把出片交给显存更大的远程 GPU 电脑。

Windows 弹出 SmartScreen 警告

如果 SmartScreen 提示无法识别此应用，点击更多信息 → 仍要运行。请只从本网站或官方发布页下载场记。

远程机器显示连不上

在这台电脑的浏览器里打开它的地址。打不开，就是端口不通——租来的服务器要在防火墙或安全组里放开它。打得开，就核对填的口令是不是远程那台打出来的那个。

远程机器连上了，却没有任务派过去

机器列表里会显示它能做什么。缺了某一步，说明那一步的模型还没下载到它上面——在浏览器里打开它的地址下载即可。另外，在设置 → 互联里看一下允许它做哪些步骤。

联系我们

没有找到答案？请联系我们。