docs(docs): DevLog 2025.10.20 (#679)
This commit is contained in:
Lovehsigure_520
Binary file not shown.
Binary file not shown.
|
After Width: | Height: | Size: 433 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 176 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 119 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 227 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 162 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 301 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 353 KiB |
@@ -0,0 +1,97 @@
|
||||
---
|
||||
title: DevLog @ 2025.10.20
|
||||
category: DevLog
|
||||
date: 2025-10-20
|
||||
excerpt: |
|
||||
Sharing the latest progress on the AIRI project, from Tauri to Electron migration, new Live2D models, and various open-source project updates.
|
||||
preview-cover:
|
||||
# TODO
|
||||
---
|
||||
|
||||
Long time no see, everyone!
|
||||
|
||||
AI trading bots are super hot recently—we've got similar research to share, starting with development...
|
||||
|
||||
## From Tauri to Electron Migration
|
||||
|
||||
Tauri went viral again a couple days ago. We were early adopters back in March, loved its plugin design and encapsulated tons of crates. Finally released v0.7.2 in June, but to deliver the voice chat everyone wanted, we struggled for 3 months... 3 months... wrestling with Tauri's WebKit + the painfully awkward Web Audio API and DevTools... all the way to September...
|
||||
|
||||
...Finally couldn't hold out anymore—switched fully to Electron over National Day!
|
||||
|
||||
<img src="./assets/electron.png" alt="electron.png" />
|
||||
|
||||
Now, building on the original Electron foundation, we've added Linux support, introduced what we call the Control Island, and it can even overlay on the interface during full-screen mode on macOS.
|
||||
|
||||
Compatibility is excellent—I absolutely love it. Yesterday, we finally got the caption overlay working, so now we can have subtitles like Neuro-sama to see what the AI is outputting!
|
||||
|
||||
<img src="./assets/control-island.png" alt="control-island.png" />
|
||||
|
||||
<div style="text-align: center; font-size: 0.875rem; color: #666; margin-top: 0.5rem;">
|
||||
Control Island
|
||||
</div>
|
||||
|
||||
## New Live2D Model
|
||||
|
||||
Sharp-eyed folks might've noticed—our model got updated! Yes, updated! I absolutely love this new model (sadly, not ready to shove it directly into the open-source repo just yet).
|
||||
|
||||
This model was honorably improved through collaboration with an artist who's worked with Neuro-sama's official team and a super talented modeling expert—new animated expressions are incredibly rich too
|
||||
|
||||
(whispering) Maybe if we get more sponsors, I'll be willing... (x
|
||||
|
||||
<video src="./assets/airi.mp4" alt="airi.mp4" controls></video>
|
||||
|
||||
## Three.js MMD Support
|
||||
|
||||
The models you have on hand or can find might not all be Live2D/VRM—actually, the most abundant and best ones are still MMD models.
|
||||
|
||||
We're also using Three.js for 3D rendering, but the reality is that Three.js no longer has a working MMD implementation. Thanks to kwaa's work, we now have a repo for this!
|
||||
|
||||
If you're interested too, come join us in maintaining it! [moeru-ai/three-mmd](https://github.com/moeru-ai/three-mmd)
|
||||
|
||||
## Velin: Write Prompts with Vue
|
||||
|
||||
>"You can write prompts with [Vue](https://velin-dev.netlify.app/#/)"!
|
||||
|
||||
Remember back in May when we shared our own prompt library? Thanks to RainbowBird hard work and donation, Velin is now officially part of Moeru AI! Almost all of AIRI's prompts are powered by Velin—but no worries about cross-platform capabilities, Velin works great in Node.js environments too!
|
||||
|
||||
<img src="./assets/velin.png" alt="velin.png" />
|
||||
|
||||
## Eventa: Event-Driven IPC/RPC
|
||||
|
||||
>"Events are all you need"
|
||||
|
||||
We once shared a project [netlify](https://velin-dev.netlify.app/#/) that lets you do pure local inference in the browser using Vercel AI SDK-like methods.
|
||||
|
||||
All these local inferences can only run in Web Workers / worker_threads, which communicate via events. Electron IPC works the same way, but we felt it wasn't elegant enough. Thanks to RainbowBird, we now have eventa—the library for driving event-based IPC/RPC implementation. [Eventa](https://github.com/moeru-ai/eventa) is now officially part of Moeru AI!
|
||||
|
||||
## Project Development Status
|
||||
|
||||
Now both Moeru AI and Project AIRI have grown into massive organizations, with over 50 original repositories covering machine learning, data processing, frontend, backend, and more—using TypeScript/Python/Rust/Go and other languages.
|
||||
|
||||
Total followers across all have exceeded 800 people. This was unimaginable when we first started a year ago—truly, thank you all so much for your support!
|
||||
|
||||
<img src="./assets/moeru.png" alt="moeru.png" />
|
||||
<div style="text-align: center; font-size: 0.875rem; color: #666; margin-top: 0.5rem;">
|
||||
Moeru AI
|
||||
</div>
|
||||
|
||||
<img src="./assets/project-airi.png" alt="project-airi.png" />
|
||||
<div style="text-align: center; font-size: 0.875rem; color: #666; margin-top: 0.5rem;">
|
||||
Project AIRI
|
||||
</div>
|
||||
|
||||
## Pure Rust TTS Implementation
|
||||
|
||||
Little teaser: Recently teamed up with kwaa to port the well-known TTS model chatterbox to a pure Rust implementation—no more worrying about tricky Python environment setup!
|
||||
|
||||
~5s inference per run on 4080S, absolutely love it.
|
||||
|
||||
Basically 1:1 recreated the Python model architecture in Rust, hoping to build it into a super streamlined local TTS inference engine leveraging other SOTA TTS models.
|
||||
|
||||
<img src="./assets/rust-tts.png" alt="rust-tts.png" />
|
||||
|
||||
## Final Words
|
||||
|
||||
That's all for today's "one more thing"—hope you enjoyed this long thread, one after another!
|
||||
|
||||
We'll continue updating tomorrow, bringing you tons more goodies, introducing our explorations in VLA/VLM gaming, how we're approaching it, and what results we're seeing.
|
||||
Binary file not shown.
Binary file not shown.
|
After Width: | Height: | Size: 433 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 176 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 119 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 227 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 162 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 301 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 353 KiB |
@@ -0,0 +1,95 @@
|
||||
---
|
||||
title: DevLog @ 2025.10.20
|
||||
category: DevLog
|
||||
date: 2025-10-20
|
||||
excerpt: |
|
||||
和大家分享 AIRI 项目的最新进展，从 Tauri 到 Electron 的迁移，新的 Live2D 模型，以及各种开源项目的更新。
|
||||
preview-cover:
|
||||
# TODO
|
||||
---
|
||||
|
||||
今天 AI 操盘炒币真的很火，我们也有类似的研究想和大家分享，先从开发开始聊起吧...
|
||||
|
||||
## 从 Tauri 到 Electron 的迁移
|
||||
|
||||
Tauri 前两天又火了一波，我们最早是在 3 月尝鲜的，期间因为喜欢 plugin 设计封装了许多 crate，虽然 6 月总算是发了 v0.7.2，但后面为了实现大家想要的语音对话功能，我们又顶着 Tauri 目前用的 WebKit + 巨难用 Web Audio API 和 DevTools 重构了 3 个月... 3 个月... 到了 9 月...
|
||||
|
||||
...最终还是憋不住，国庆彻底换了 Electron
|
||||
|
||||
<img src="./assets/electron.png" alt="electron.png" />
|
||||
|
||||
现在 Electron 在原有的基础上，有了 Linux 的支持，添加了我们称之为的 Control Island（控制岛），甚至能在 macOS 的窗口全屏的时候也可以 overlay 在界面上
|
||||
|
||||
兼容性很好，孩子真的很喜欢，昨天也总算是有了 caption overlay，可以像 Neuro-sama 一样有个字幕可以看到 AI 输出了什么了
|
||||
|
||||
<img src="./assets/control-island.png" alt="control-island.png" />
|
||||
|
||||
<div style="text-align: center; font-size: 0.875rem; color: #666; margin-top: 0.5rem;">
|
||||
Control Island
|
||||
</div>
|
||||
|
||||
## 新的 Live2D 模型
|
||||
|
||||
眼尖的伙伴可能发现了，我们的模型更新了！是的更新了！我超喜欢现在的这个模型（sadly 目前还是不想在开源仓库里直接塞这个模型）
|
||||
|
||||
这个模型很荣幸是找到了 Neuro-sama 官方也合作过的一位画师和非常厉害的建模老师合作改进的，新的动画表情也非常丰富
|
||||
|
||||
（小声）也许 sponsor 多了就愿意了（x
|
||||
|
||||
<video src="./assets/airi.mp4" alt="airi.mp4" controls></video>
|
||||
|
||||
## Three.js MMD 支持
|
||||
|
||||
大家手头上有的或者能找到的模型不一定是 Live2D/VRM 模型，可能最多的最好的模型还是 MMD 的
|
||||
|
||||
我们也是基于 Three.js 实现的 3D 渲染，然而现状是 Three.js 这边已经没有正常在工作的 MMD 实现，感谢 kwaa 的工作，我们现在有仓库做这个了
|
||||
|
||||
如果你也感兴趣，[请来一起维护！](https://github.com/moeru-ai/three-mmd)
|
||||
|
||||
## Velin：用 Vue 写提示词
|
||||
|
||||
>「你可以用 [Vue](https://velin-dev.netlify.app/#/) 写提示词！」
|
||||
|
||||
还记得 5 月份的时候我们也分享过我们自己实现的 prompt 库吗？感谢 RainbowBird 的努力和捐献，现在 Velin 也是 Moeru AI 的一员了，AIRI 几乎全部的提示词都是 Velin 驱动的，不过不用担心跨平台能力，Velin 在 Node.js 环境也可以用！
|
||||
|
||||
<img src="./assets/velin.png" alt="velin.png" />
|
||||
|
||||
## Eventa：事件驱动的 IPC/RPC
|
||||
|
||||
>「事件就是你所需要的一切」
|
||||
|
||||
我们曾分享过 [Netlify](https://velin-dev.netlify.app/#/) ，一个允许在浏览器里用 Vercel AI SDK 类似的方式进行纯本地推理的项目
|
||||
|
||||
这些本地推理都只能在 Web Worker / worker_thread 中实现，而他们又都是事件通信，Electron IPC 也如此工作，但我们觉得这都不够优雅，还是感谢 RainbowBird ，用来驱动和实现基于事件的 IPC/RPC 的库，[Eventa](https://github.com/moeru-ai/eventa)，现在也是 Moeru AI 的一员了
|
||||
|
||||
## 项目发展现状
|
||||
|
||||
现在 Moeru AI 和 Project AIRI 都已经是非常庞大的组织，拥有超过 50 个原创的仓库，涵盖机器学习、数据处理、前端、后端各种领域，TypeScript/Python/Rust/Go 等多种语言
|
||||
|
||||
总共加起来的 follower 已经超过 800 人，这放在 1 年前我们刚创立的时候是不敢想象的，真的非常感谢大家喜欢
|
||||
|
||||
<img src="./assets/moeru.png" alt="moeru.png" />
|
||||
<div style="text-align: center; font-size: 0.875rem; color: #666; margin-top: 0.5rem;">
|
||||
Moeru AI
|
||||
</div>
|
||||
|
||||
<img src="./assets/project-airi.png" alt="project-airi.png" />
|
||||
<div style="text-align: center; font-size: 0.875rem; color: #666; margin-top: 0.5rem;">
|
||||
Project AIRI
|
||||
</div>
|
||||
|
||||
## 纯 Rust TTS 实现
|
||||
|
||||
小预告：最近和 kwaa 把很知名的 TTS 模型 chatterbox 迁移到了纯 Rust 实现的版本，这样就不需要担心 Python 环境不好配置的问题了
|
||||
|
||||
4080S 差不多 5s 推一次，很喜欢
|
||||
|
||||
几乎是在Rust里 1:1 对着 Python 的模型架构实现了一遍，希望能够依托其他 SOTA 的 TTS 模型变成一个很精简的本地 TTS 推理引擎
|
||||
|
||||
<img src="./assets/rust-tts.png" alt="rust-tts.png" />
|
||||
|
||||
## 最后
|
||||
|
||||
今天的 one more thing 就到这里，希望大家喜欢这么长的一个 thread 一个接一个
|
||||
|
||||
我们明天会继续更新，给大家带来更多超级多的干货，给大家介绍我们在 VLA / VLM 玩游戏上，做了什么探索，如何进行，以及效果如何
|
||||
Reference in New Issue
Block a user