去年冬天我妈住院,我在医院陪床。
隔壁床的阿姨,儿子在深圳,一个月回来一次。有天下午她拿着手机戳了半天,抬头问我:“小伙子,你帮我看看,我儿子发的照片在哪儿?”
我接过来,微信九十多条未读,照片埋在某个群的第三十几条消息里。翻出来递给她,她盯着屏幕看了很久,说了句:“他这是在哪儿啊。”
那一下我挺不是滋味的。
我们做技术的总有种错觉,觉得给爸妈换台新手机、装几个 APP,就叫尽孝了。可微信里那个层层嵌套的相册入口,对他们来说跟一道解不开的数学题没区别。
腾讯数码前阵子有篇文章,标题挺扎心:《一年就回几趟家,讲真,少给父母买智能产品》。我一开始不服气,看完觉得说得对。你买的智能音箱他们不敢用,你装的摄像头他们觉得被监视。因为那些东西的交互逻辑是给你设计的,不是给他们设计的。
所以这两年我一直在找一个东西:不用学、不用点、放桌上就能用。
前几天在立创开源刷到一个项目,我觉得找到了。
先看这是个什么东西
AI智能相框(小智电子相册 PhotoTVbox)作者:ccy-studio | 立创开源·星火计划2025
数据:2.1万浏览、218 收藏、25 条评论
授权:GPL 3.0,原理图、PCB、BOM、3D外壳、固件源码、上位机源码,一个不落
复刻成本:约 200 元
功能说起来一句话就完了:一个 3.5 寸的桌面相框,你在小程序里发照片它自动显示,你对它说话它能回,到点了它提醒吃药。
听着简单。但把它拆开看,工程含量比外观高太多了。
一、为什么是 ESP32-S3:一个正在发生的趋势
先说个行业观察。
最近圈里在聊 OpenClaw、MimiClaw,“养龙虾”刷屏了小半个月。热闹背后其实是一个很明确的信号:AI 智能体正在从云端往端侧走,从 Linux 往 MCU 走。
搁两年前,你想做个会说话的设备,第一反应是上树莓派跑 Python。现在越来越多人直接拿几十块的 ESP32 就干了。前几天还看到一篇《从 OpenClaw 到 MimiClaw:AI 助手开始往 ESP32 走》,讨论的就是这件事。
这个相框,正好是这波趋势里一个挺完整的样本。
主控 ESP32-S3,8MB PSRAM + 16MB Flash。
PSRAM 是关键。LVGL 跑双缓冲、解 JPEG 图片,内存吃得凶,片内那点 SRAM 根本不够看。16MB 的 Flash 则装得下 UI 资源、中文字库,还留得出 OTA 双分区。
选型上没什么炫技的地方,刚好够用,也不浪费。做过产品的人才会这么选。
二、最见功力的地方,在电源
我看开源项目有个习惯:先翻电源部分。
因为电源最能看出作者是玩票的,还是真做过量产。
这个项目的电源,我看完的评价是:作者被坑过。
核心是 BQ24075,做的是动态电源路径管理(DPPM)。翻译成人话:插着 USB 的时候,系统由适配器直接供电,同时给电池充电;拔掉 USB,无缝切到电池。整个过程系统不断电、不重启。
没做过的人可能觉得,不就是充电吗。
实际上没有路径管理的方案,插拔那一瞬间电压会跌,MCU 直接复位,屏幕黑一下。这种 bug 在实验室不一定复现,到了用户手里天天遇到,还查不出来。
但真正让我停下来看了两遍的,是它怎么处理“插电自动开机”。
设备用了一颗专用长按开关机芯片(EC190708)做电子开关。矛盾就来了:用户希望“没电关机后,一插上 USB 就自动开机”,可开关机芯片的逻辑是等你手动按下去。
作者的解法是加一路 NMOS,检测到 USB 插入,就把开关机芯片的按键脚拉低。相当于电路自己帮你按了一下。
思路一点都不复杂。但它属于你没踩过就想不到、踩过一次终身难忘的那类经验。开源项目里愿意把这种细节公开的,真不多。
避坑提醒BQ24075 的 NTC 脚,要么接热敏电阻做温度保护,要么用 10K 电阻拉到设定值。两个都不接,充电直接不工作,而且它不报错。你多半会以为是芯片买到假的。
三、UI 用 LVGL,但值钱的是它的开发方式
界面是 LVGL 做的,这不稀奇。稀奇的是作者怎么开发它。
他搭了一套 SDL2 的 PC 模拟器:UI 代码先在电脑上跑起来,鼠标点,实时看效果,调顺了再用 sync_esp_code.py 同步进 ESP-IDF 工程编译烧录。
做过 LVGL 的都懂那个痛苦循环:改一个控件坐标,编译,烧录,上电,眯着眼看效果,一轮两三分钟。一个界面调完,半天就没了。
用模拟器,这个循环压到几秒,还能直接上电脑的调试器打断点。
这套方法跟相框本身没半点关系,但它能原样搬进你自己的任何一个 LVGL 项目。整个项目里,我觉得这块的复用价值最高,比任何一张电路图都值钱。
四、语音:站在小智的肩膀上
对话能力是基于小智 AI 二次开发的。
硬件是标准三件套:ES7210 做双麦采集(带回声消除、降噪),ES8311 做播放解码,NS4150B 功放推喇叭。这套组合在小智生态里被验证过无数次,稳。
有意思的是软件层。相框有两个身份:平时是相册,你叫它的时候是语音助手。作者的做法是监听小智的 Idle 事件,对话结束进入空闲,自动切回相册页;被唤醒,跳到对话页。
说白了就是个状态机,但落到体验上,就是“它知道什么时候该退下”。
对老人来说这点特别重要。他们不用去搞懂什么叫“模式”,也不用记得该切到哪一页。
五、再往上:云端、小程序、Web 一整套
如果只做到本地,它就是个电子相框。这个项目往上还搭了一整套:
服务端用 Flask + MQTT,MQTT 推指令和消息,Flask 管账户和文件。
Space 账户隔离,多设备多用户,各看各的,不串号。
客户端用 Uni-app,一套代码同时出小程序和 Web。
这里有个挺现实的细节:小程序因为资质问题上不了架,作者转做了 Web 端。
个人开发者迟早都要撞上这么一次。所以技术方案最好从一开始就留条后路,别把宝全押在一个分发渠道上。
从硬件到驱动,到 RTOS、UI、AI、云端、前端,这一个项目把嵌入式全栈完整跑了一遍。
我很少单独推荐某个项目。但要找个东西系统性练手,这个的性价比是真高。
想复刻的看这里
成本约 200 元(含屏幕,外壳 3D 打印另算)。资料包含原理图、PCB、BOM、3D 模型、固件源码和上位机源码,GPL 3.0 全开放。
三个提醒:
1. NTC 那个坑,前面说了,别忘。
2. 作者的测试服务器不保证长期开着,建议自己私有化部署一套,Flask + MQTT,一台最便宜的云服务器就够。
3. 外壳直接打作者的模型,尺寸和 PCB 是对上的,自己改容易装不进去。
最后说点别的
写这篇的时候,我又想起医院里那位阿姨。
我们这行有种挺常见的傲慢:默认用户会学。会注册、会点击、会自己找入口。可对我爸妈那代人来说,“不用学”本身就是最高级的功能。
一个放在桌上的相框,不用解锁,不用找入口,照片自己就来了,你想说话它就应你。
这里面没一样是前沿技术。ESP32-S3 不新,LVGL 不新,MQTT 更不新。
难的是有人肯花心思去想这么一件事:一个七十岁的人第一次看到它,会伸手去点哪儿。
声明:这个项目是立创开源作者 ccy-studio 的作品,不是我做的。我只是完整拆了一遍,觉得好,值得让更多人看到。想复刻的直接去原项目,记得给作者点个收藏。
原项目:oshwhub.com/yc_chen/ai-smart-photo-frame
你有没有给爸妈做过什么小玩意儿?评论区聊聊,我挺想看的。
想动手的,回复关键词【相框】,我把这个项目的原理图、PCB、BOM、固件源码、外壳 3D 模型打包整理好发你。
另外附一份我自己整理的 LVGL + SDL2 模拟器搭建步骤。
觉得有用的话,点个「在看」
让更多做嵌入式的朋友看到
My neighbor’s mom, whose son works in Shenzhen, visits him once a month. One afternoon, she fumbled with her phone for a while, then looked up and asked me, “Kid, can you help me find my son’s photos? He sent them, but I can’t figure out where they are.”
I took the phone and scrolled through over 90 unread WeChat messages. The photos were buried in the third or fourth line of a group chat. I retrieved them and handed the phone back to her, but she just stared at the screen for a long time before saying, “I wonder where he is…”
That moment really struck me.
As techies, we often think that giving our parents a new phone or installing a few apps is enough to fulfill our filial duties. But the deeply nested album feature in WeChat is like a mathematical puzzle to them — impossible to solve without guidance.
Tencent Digital published an article not long ago titled “Since you only go home a few times a year, honestly, stop buying smart products for your parents.” At first, I disagreed, but after reading it, I realized they’re right. The smart speakers we buy, they don’t use; the cameras we install, they feel are spying on them. Because the interaction logic is designed for us, not for them.
So for the past two years, I’ve been searching for something that doesn’t require learning, doesn’t need tapping — something that can be used as soon as it’s placed on the table.
A few days ago, I found a project on the LCSC Open Source platform that I think hits the mark.
What is this?
In a nutshell, it’s a 3.5-inch desktop photo frame. You send photos via a mini-program, and the frame displays them automatically. You can talk to it, and it responds. It even reminds you to take medicine at certain times.
It sounds simple, but the engineering depth is far greater than its appearance suggests.
**1. Why ESP32-S3: A Trending Shift in the Industry**
Lately, the tech community has been buzzing about OpenClaw and MimiClaw, with discussions on “raising lobster” dominating social media for weeks. Behind the hype is a clear trend: AI agents are moving from cloud servers to edge devices, and from Linux to microcontrollers (MCUs).
Two years ago, if you wanted to make a talking device, you’d reach for a Raspberry Pi and write Python. Today, more and more people are doing the same with a $30 ESP32. A recent article, “From OpenClaw to MimiClaw: AI Assistants Are Moving to ESP32,” explicitly discusses this shift.
This photo frame is a perfect example of this trend.
The core is an ESP32-S3 with 8MB PSRAM and 16MB Flash.
The PSRAM is critical — LVGL (the UI library) requires significant memory for double buffering and JPEG decoding, far exceeding the MCU’s internal SRAM. The 16MB Flash is sufficient to store UI resources, Chinese character fonts, and dual partitions for over-the-air (OTA) updates.
The choice reflects practical engineering rather than showing off — it’s just enough for the job, no more, no less.
**2. The Most Impressive Part: Power Management**
In any open-source project, I first check the power section.
Power design reveals whether the creator is a hobbyist or a professional who has worked on mass-produced products.
This project uses a BQ24075 for Dynamic Power Path Management (DPPM). What that means in plain English is: when USB power is connected, the system runs off the adapter and charges the battery simultaneously; when USB is unplugged, the system seamlessly switches to battery power without rebooting or losing data.
Without this management, the voltage would drop during the transition, causing the MCU to reset and the screen to go blank. These are the kinds of bugs that don’t show up in the lab but drive users crazy.
However, the most intriguing detail is how the device handles “plug-in auto-activation.”
The device uses an EC190708 chip to manage shutdown and startup via a physical button. The challenge is this: users expect the device to turn on automatically when plugged in after it has been shut down due to low battery — but the chip’s logic only activates on manual button press.
The solution? The creator added a path that detects the USB connection and triggers an NMOS transistor to pull the “button” pin low, effectively simulating a manual press.
This is a classic case of “You don’t know until you’ve been bitten, and once you know, you never forget.” In open-source circles, sharing these specific solutions is rare and highly valuable.
**3. UI with LVGL: The Value is in the Development Method**
While the use of LVGL for the interface is standard, the true innovation lies in the development workflow.
The creator set up an SDL2 simulator on the PC. This allows for the simultaneous development of the UI on a computer (with mouse interaction) and the firmware on the MCU. Once the UI is perfected, a simple script (sync_esp_code.py) syncs the changes into the ESP-IDF project for compilation and deployment.
Anyone working with LVGL knows the pain of the edit-compile-deploy-test cycle, which can take several minutes per iteration. Using a simulator cuts that down to seconds, enabling real-time debugging and parameter adjustment.
This is the most reusable part of the project — it’s a methodology applicable to any LVGL project, not just this specific frame.
**4. Speech Capabilities: Building on Existing Foundations**
The conversational ability is built on top of the Ai-Thinker Speech Kit (小智).
The hardware consists of a standard trio: the ES7210 for dual-microphone recording (with echo cancellation and noise reduction), the ES8311 for audio playback, and the NS4150B amplifier. This combination is well-tested and stable.
The software side is interesting. The frame operates in two modes: a photo viewer by default and a voice assistant when spoken to. The author listens for the “Idle” event from the speech engine; when the conversation ends, the system returns to the photo view, and when it detects a wake-up word, it switches to the dialogue view.
This simple state machine ensures the device behaves intuitively — it knows when to step forward and when to step back, without confusing the user with manual mode switching.
**5. The Full Stack: Cloud, Mini-Program, and Web**
The project doesn’t stop at the device itself. It includes a complete ecosystem of services.
The backend is powered by Flask and MQTT, handling authentication, file storage, and push notifications.
The user interface is built using Uni-app, which generates both a WeChat mini-program and a web version.
A significant point for real-world application: the mini-program couldn’t be uploaded due to certification issues. So the creator pivoted to a web-based interface. This adaptability is crucial for independent developers who often face distribution roadblocks.
**6. Conclusion: A Comprehensive Learning Project**
This project covers the entire span of embedded development: hardware design, driver implementation, operating systems (ESP-IDF), user interfaces, artificial intelligence, cloud services, and front-end programming.
For someone looking to build a portfolio or learn the ropes, this is an excellent choice. The cost is reasonable (~$200 including the screen and excluding the 3D printed case), and the documentation is thorough, including schematics, PCB files, a bill of materials, and all source codes.
**A Note on NTC**
One critical error in the design is noted by the author themselves: the NTC (Negative Temperature Coefficient) sensor, used for battery temperature monitoring, is incorrectly placed. It is connected in series with the charging current path rather than in parallel. This causes the sensor to heat up during charging, leading to an inaccurate over-temperature reading and preventing the battery from fully charging. This is a low-level but very costly error to overlook.
**Final Thoughts**
Returning to the hospital visit, I think a device like this could have been a lifesaver for that auntie. It requires no sign-in, no account creation, and no navigation through menus. The photos appear automatically, and the assistant responds to simple voice commands.
The technology here isn’t bleeding edge — ESP32-S3, LVGL, and MQTT are all well-established. What sets this project apart is the consideration for the user experience of someone who doesn’t want to “figure it out.”
Have you ever built a gadget for your parents? I’d love to hear about your experiences in the comments.
想复刻的看这里
Cost is about 200 yuan (including screen, 3D printed case extra). Includes schematics, PCB, BOM, 3D model, firmware source, and host software — all open under GPL 3.0.
三个提醒:
1. Don’t forget the NTC pit — I already mentioned it.
2. The author’s test server isn’t guaranteed to stay up; best to deploy your own Flask + MQTT on a cheap cloud server.
3. Use the author’s case model as-is — it fits the PCB perfectly; modifying it risks poor fit.
最后说点别的
While writing this, I recalled the elderly lady in the hospital.
Our field often assumes users will learn — register, click, find entrances. But for my parents’ generation, “no learning required” is the ultimate feature.
A frame on the desk, photos appear without unlock or search, and it responds when spoken to.
No cutting-edge tech here — ESP32-S3, LVGL, MQTT are all well-established.
The challenge is caring enough to consider: what would a 70-year-old reach for first time seeing this?
Have you made anything like this for your parents? Share in comments — I’d love to see.
If you found this helpful, tap “Like” to spread the word among embedded peers.