Why it exists
Image and video models can make a beautiful few seconds. A film is something else: a story that holds together, characters who look like themselves from one shot to the next, cuts that follow a rhythm, sound that lines up with picture. Getting all of that out of a handful of models is mostly bookkeeping — the kind a script supervisor does on set.
changji is that bookkeeping, automated. It isn't a model; it's the production line that connects models and makes sure every step reads from the same plan.
It also runs the models. Most tools that promise "a drama from one sentence" are a front end around somebody else's paid API: you bring the key, the generation happens on their servers, and the meter runs per second. changji carries the language, image, video and speech models inside one program and runs them on your own GPU. No API key, no account, no subscription — and nothing you make passes through anyone else's machine.
Principles
The storyboard is the source of truth
Every stage reads and writes the same structured storyboard. Stages never pass loose text to each other, so nothing gets lost in translation.
Consistency by construction
A character's appearance is assembled by the program from the asset library, identically for every shot. The model has no field in which to describe a face, so it can't reinvent one.
Sound first
Voiceover is recorded before shots are locked, and shots are timed to the real length of each line.
Unattended means reliable
Every clip passes quality gates, and failures are retried or degraded gracefully. A pipeline that stops and waits overnight has wasted the night.
No API key
The models live inside the app and run on hardware you own: nothing to sign up for, nothing to top up, and your projects stay with you. Need more speed? Run changji on a remote GPU machine — another of your own, or a server you rent — and link it with this one. It's still your hardware, and still no API key.
Honest about AI
Every video changji makes carries a visible "AI generated" mark and a note in its metadata. What's made by AI should say so.
Open source
The changji engine is open source under the Apache-2.0 license — the permissive kind. You can use it commercially, change it and ship it without asking anyone, which is not a given in this category: several comparable projects are "open" under non-commercial or no-hosting terms. The models it runs come with their own licenses — some don't allow commercial use — so check them before you publish. Source code and releases are on GitHub.
为什么做它
出图、出片模型能生成漂亮的几秒钟画面。可电影是另一回事：故事要前后连贯，角色在每一镜里都得是同一个人，剪辑要有节奏，声音要和画面对齐。要让几个模型做出这些，大部分工作其实是「记账」——正是片场里场记做的那种事。
场记把这份记账自动化了。它不是模型，而是把模型串起来、确保每一步都依据同一份计划的生产线。
模型也是它自己跑的。市面上承诺「一句话生成短剧」的工具，大多是套在别人付费 API 外面的一层界面：Key 你自己带，生成在对方的服务器上完成，按秒计费。场记把大模型、出图、出片、配音模型都装在一个程序里，在你自己的显卡上运行。不用 API Key，无需账号，无需订阅——你做的东西也不会经过任何人的机器。
几条原则
分镜表是中枢
所有阶段读写同一份结构化的分镜表，阶段之间不传递零散的文字，信息不会在传递中走样。
一致性由结构保证
角色外观由程序从资产库机械拼接，每一镜逐字相同。模型根本没有可以描述长相的字段，也就无从「重新设计」一张脸。
配音先行
先配音，再锁定镜头，每一镜的时长按台词的真实长度确定。
无人值守就要可靠
每段视频都要过质量闸门，不合格就重试或降级处理。一条停下来等人的流水线，等于白白浪费了一整夜。
不用 API Key
模型就在程序里、在你自己的硬件上运行：不用去哪儿注册，不用充值，片子留在你手里。想要更快？在一台远程的 GPU 电脑上也跑一份场记——你自己的另一台，或者租来的服务器——和这台互联。那依然是你的硬件，依然不用 API Key。
如实标注 AI
场记生成的每段视频都带有可见的「AI 生成」标识，并在文件元数据中注明。AI 做的东西就应该说明是 AI 做的。
开源
场记引擎以 Apache-2.0 许可证开源——宽松的那一种：商用、修改、再发布都不用问任何人。这在同类里并不理所当然：好几个同类项目的「开源」附带不可商用或不许托管的条款。它运行的各个模型有各自的许可证，其中一部分不允许商用，发布作品之前请先确认。源代码和发布包在 GitHub 上。