TypeTale: a Windows studio that chains prompts, storyboards and video models into one flow

A desktop studio aimed at creators of short video from text

TypeTale is a Windows application for AIGC creation, described in the repository as an entirely free tool mainly used for novel promotion videos, AI short drama and AI film. The homepage for downloads is typetale.top, and the README is written in Chinese, which is consistent with a product aimed at a Chinese creator market.

The free claim has a specific shape that is worth reading carefully. The README says the application is currently completely free and promises that the existing features and future basic features will remain free permanently. Everything else in the product, though, runs on a compute platform that is metered, and the changelog makes that visible with prices per second of video generation.

The audience is creators rather than developers. The feature list runs from script preparation through to a finished edit, and the system requirements section is the giveaway: Windows only, 8 GB of memory recommended, and a note that some features need a GPU with AI support while the rest work without one. In other words, the heavy generation runs on the platform, and the desktop application is the workspace, the orchestration layer and the editor around it.

The support channels are also product-shaped rather than project-shaped. The README invites feature requests and problem reports through GitHub issues and a QQ channel, and a separate link points to a Feishu wiki with more documentation. The repository itself contains a single file, README.md, so what is published on GitHub is documentation and a changelog rather than code.

The chain runs from copy editing to a CapCut draft

The longest section of the README describes an end-to-end production chain, and it is the clearest statement of what the software actually is.

The chain starts with copy preprocessing, with built-in tools for viral openings, expanding a script, condensing it and polishing its style, all through a language conversation. From there the software pulls in a list of trending videos that can be cloned as a starting point, and cuts a trending video into a storyboard automatically. Storyboards can be reordered by hand or by the model.

The reverse direction is also supported. A storyboard image can be turned back into an image generation prompt, and a language model can extract the plot and a list of characters with descriptions from the text, and rewrite the copy in a specified style. The prompt reasoning step then combines the reversed prompt, the story, the character descriptions and the storyboard text to infer the character information and the image prompts for each shot.

Generation follows. Images are produced from those prompts with the built-in styles, audio is generated with EdgeTTS, a local Gpt-SoVITS or the platform's own server, and video is generated with models such as Wan, Veo3 and Sora2 according to the storyboard timing and the audio duration. A lip sync step then aligns the mouth movement to the generated audio.

The export is the part that makes it a production tool rather than a toy. The video drafts are exported in bulk to CapCut, with keyframes and storyboard effects added automatically, and a CapCut template can be chosen at export time to set subtitles and add a watermark. The whole flow is what the term novel promotion video means in practice, and it is the reason the product exists.

The plugin model is the extension point, and it has three scopes

The README is explicit that the software opens up its image and video generation capabilities for secondary development. A Python plugin can be written to operate the software and produce images with arbitrary logic, and to build custom model pages. ComfyUI support goes further, since the application claims complete support for ComfyUI capabilities and the ability to develop ComfyUI nodes for arbitrary image and video sources.

The plugin system itself has three scopes, which is a sensible design for a creative tool. A panel-scoped plugin appears on a storyboard button, a project-scoped plugin sits in the sidebar of the right of a work, and an app-scoped plugin sits in the sidebar of the top left of the application. The same code can therefore be available at the level of a single shot, a single project, or the whole application.

A plugin IDE is built into the product, and the changelog says that writing a plugin inside TypeTale requires no programming background because you talk to the AI assistant instead, with a recommendation to use the platform's own language model for the job. The changelog also records that a plugin system called Codex was given more capabilities over time, that multiple Codex sessions can be opened side by side, and that an earlier capability called OpenClaw was removed once Codex could replace it.

The ComfyUI integration is more than a checkbox. The changelog for one build describes dispatching jobs to a suitable machine based on ComfyUI machine load, enabling or disabling individual machines in the settings, recovering automatically from a task disconnect caused by network instability, cancelling a task that is stuck on a remote server, and fixing a machine that could hang permanently. That is the language of a system that has been operated at scale rather than demoed on a laptop.

The application is a front end for many hosted back ends

The provider matrix is long, and it explains where the boundaries of the product are.

For images the application integrates StableDiffusion WebUI, ComfyUI, Midjourney, Nano Banana and Qwen. For video it lists ComfyUI with Wan, Veo3 and Sora2, plus Midjourney, with support for text to video, image to video, and first frame or first and last frame driving. For audio there is EdgeTTS, GPTSoVITS and IndexTTS2. The language model list includes Aliyun, SiliconFlow, DeepSeek, OpenAI, Gemini and Claude, plus arbitrary custom platforms and a relay platform, and the vision model list covers Aliyun, OpenAI, Gemini and again custom or relayed platforms.

Two of these are worth separating out. The relay platform means the application can sit in front of a provider it does not natively know, which is how a product keeps working when a model vendor changes its API. The custom platform option is the same idea for a private endpoint.

The changelog also shows how the video model set moves. A build added the Jimeng CLI into the engine, which is limited to advanced members and requires the software to detect the account after login. Another build added Seedance 2.0, including the ability to work with a real person directly without uploading a reference image. The MiniMax H3 model was adapted twice in a month, once for the platform and once for better ComfyUI integration with more reference images and multiple reference audio and video.

There is a local path too. The README mentions complete images on two compute clouds and a complete local AI bundle, so the same workflows can run either against hosted infrastructure or against a local installation of the models, depending on what the user has.

The paid layer inside the free promise

Because the README is transparent about pricing in the changelog, the cost model is visible, and it is not uniform.

The compute platform sells generation time. One changelog entry states a price of 0.08 yuan per second for MiniMax H3 at 480P, and another puts a membership price for Seedance 2.0 at 720P as low as 0.34 yuan per second, with credits shared between the web and desktop clients. The Jimeng CLI is restricted to advanced members, and a plugin update mentions a Grok model charged per use.

The second paid item is more unusual: encrypted prompt templates. A changelog entry describes a prompt template encryption feature aimed at enterprises and individual creators, which must be bought from the compute platform's feature center. Templates can be uploaded encrypted to the cloud and distributed to a whitelist of users, who log in with a permitted phone number, receive the template at startup, and can use it without being able to read it. The restriction is that encrypted templates only work with the platform's own language model, so the protection is enforced by the choice of model rather than by cryptography alone.

The third is a creator revenue programme. A first co-creation feature lets a creator package a fixed AI video production flow as an application-level plugin, which the team then encrypts and sells jointly, with the revenue split mentioned in the entry.

None of this contradicts the free promise as stated, since the promise covers the application and its basic features. But a project that generates a few hundred seconds of video a day is paying a metered amount, and that belongs in any estimate of what a production run will cost.

What a README-only repository can tell you, and what it cannot

The repository contains one file. There is no source, no build system and no test suite to read, the repository listing records the licence as unknown, and the primary language field is not filled in. The last push was on 2026-08-21, and there are no GitHub releases, so the versioning you see is the application version in the changelog rather than a tag.

That changes how the changelog should be read. Entries such as 9.0.19, 9.0.11, 9.0.0 and a long run of 8.0.x releases document a desktop application that is updated frequently, and the entries are written from a product perspective, describing new views, new models and fixed bugs rather than API changes. For a user deciding whether to adopt the tool, that is exactly the right kind of information. For an engineer deciding whether to build on it, there is nothing to build on.

The other difference worth noting is that the most interesting parts of the product are not in the README at all. The 3D director's desk view, with 360 degree panorama generation, world model generation and character modelling, appears in the changelog and is described in passing. The canvas view, the standalone OmniBoard product, the knowledge base with similar image search, the trending video library that can be cloned, the prompt tag library, the character template library and the voice model library are all listed as features without much explanation of how they behave.

So the honest verdict is that this is a product announcement with a good feature list, not an open source project in the sense this site usually reviews. It is worth trying, since the download is a desktop client from the project's own site and the basic features are promised to stay free. It is not worth reading as source, because there is none here to read.

Editorial conclusion

Adopt TypeTale if you produce novel-driven short video or AI short drama and want one desktop application that holds the whole chain from script to a CapCut draft, with a plugin system so you can add your own generation back ends. Do not adopt it as an open source dependency, because the repository contains only a README, the licence is not stated in the repository listing and there is no code to read, and do not assume the free promise covers the compute, since the video models and the encrypted prompt templates are metered separately. Before you commit to a project, price the per-second compute on the platform, check whether the model you need is available in your region, and confirm that the export format you want is the CapCut draft rather than a finished video file.

Frequently asked questions

Official sources

1. Issues
2. README
3. TypeTale/TypeTale on GitHub