⌘K
Change language Switch ThemeSign In
Narrow Mode
暗壳 AI：物理空间设计，需要自己的 Lovart
赛 赛博禅心 @赛博禅心
One Sentence Summary
DarkShell AI is a vertical AI platform dedicated to physical space design. By engineering industry experience and combining it with real supply chain data, it solves the core pain points of "low fault tolerance" and "difficult implementation" that AI faces in space design.
Summary
This article provides an in-depth introduction to "DarkShell AI," an AI product vertical to the space design field. Founded by a team with deep architectural design background, DarkShell positions itself as "the Lovart for space design." Unlike Midjourney and other planar AI tools that focus on visual presentation, DarkShell emphasizes the "hard constraints" of the physical world—accuracy, realism, and implementability in design. Its core product architecture includes AI Agent, collaborative canvas, industry private data, and a supply chain database of 300,000+ real SKUs. The article points out that pure technology teams struggle to cross the threshold of industry experience. Through infusing over a decade of design logic into the model, DarkShell achieves a closed loop from vague requirements to precise procurement lists. In the future, DarkShell will not only serve human designers but also serve as a space-understanding Agent接入 AI ecosystem, while providing high-quality 3D training data for embodied intelligence.
Main Points
* 1. Physical space design has extremely low fault tolerance for AI, requiring "serious and accurate" output.Unlike planar design, space design involves hard constraints such as dimensions, materials, and construction. Through deep training on SKUs, DarkShell ensures that generated schemes have controllable form, lighting, and materials, avoiding the common AI "hallucination" that makes schemes unimplementable. * 2. Engineering encapsulation of industry experience is the core moat of vertical AI.Pure technology teams struggle to understand the subtle differences between "Eastern design" and "New Chinese style." DarkShell leverages the million-level private data from parent company Matrix Shares and feedback from hundreds of designers to transform non-standardized design experience into capabilities that Agents can dispatch. * 3. Bridging the closed loop between "infinite creativity" and "finite physical world" is the product's commercial value.DarkShell not only provides an infinite canvas for creative brainstorming but also integrates an ecosystem library of 300,000+ real products for sale, enabling one-click generation of procurement lists and budgets after scheme confirmation, greatly reducing communication and modification costs. * 4. Space design Agents will become important infrastructure for embodied intelligence and the AI ecosystem in the future.The 3D space data accumulated by DarkShell and its deep understanding of physical environments can serve as training material for embodied intelligence in the future. Additionally, its professional capabilities can be invoked by other Agents through API, becoming an irreplaceable link in the AI ecosystem.
Metadata
AI Score
83
Website mp.weixin.qq.com
Published At Today
Length 2557 words (about 11 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
原创 金色传说大聪明 2026-03-20 19:41 北京
AGI Bar 的空间设计，就是暗壳团队做的
PRODUCT INSIGHT
AGI Bar 整个空间的设计，就是暗壳团队做的
暗壳是老朋友，AGI Bar 从选址到落地，他们出了大力，帮了大忙，纯友情的
我们是相互看着一步步往前走的。在这里，让我一方面以行业观察者的角度，也以真实朋友的角度，来介绍下「暗壳 AI」
暗壳，是空间设计的 Lovart
--- 我给的定义
TL;DR
Lovart 在平面设计领域已经是行业典范。那空间设计领域呢？暗壳要做的就是这件事——空间设计的 Lovart
不同的是，物理世界有硬约束。方案里的沙发不只要好看，还得买得到、搬得进去、尺寸严丝合缝
01 / PRODUCT
暗壳 Agent 2.0
先介绍下两个朋友：
黄燕虹，创始人兼CEO，矩阵股份合伙人，独立家居品牌创始人，阿里巴巴设计趋势专家顾问，连续多年发布设计行业趋势白皮书，做过家居品牌目所 Msol Studio
黄政民，联合创始人兼COO，矩阵股份合伙人，职业经理人，曾主导开发参数化设计平台及矩阵全员AI设计转型
两个人 2022 年底从矩阵(301365)出来，做了暗壳
产品架构五个模块：AI Agent、协同自由画布、行业数据、供应链生态库、数字资产管理
说核心逻辑——
空间设计师的日常是这样的。客户说「我想要有呼吸感的客厅」，设计师花好几轮揣摩这句话到底什么意思
出方案，改方案，再改。确认之后找供应商，对接采购，落地施工
中间要跳转若干个软件：建模工具、渲染工具、PPT、微信
这个过程中，真正贵的是沟通和修改
暗壳的 Agent 试图接管这整个流程
用户输入「有呼吸感的客厅」，它不会立刻出图。它会反问：你说的「呼吸感」，是侧重大面积留白，还是材质的通透性？
这种对话能力，靠的是团队把十几年的行业经验做了工程化封装，灌进了 Agent 的工作流逻辑里
出图之后它还会主动提醒：这个产品尺寸可能不对，要不要换一个生态库里能买到的
画布是工作台，所有操作在一个界面里完成
生态库里目前有超百家品牌、30 万+ SKU真实在售产品，设计师可以直接调用。方案确定后一键生成采购清单和预算
从一句模糊需求，到一卡车家具搬进你家，尽可能多地跑在一个系统里
02 / CONSTRAINT
物理世界不接受近似值
说一件我自己经历的事
前段时间租了邻居家当办公房，让宜家上门量尺做软装方案
量完之后工作人员在后台一顿倒腾，从尺寸到布局，忙活了好一阵
最后怎么交付的？挨个截图，贴进 PPT，保存，发过来
这大概是目前大部分软装提案的现状——一张张需要你看完自己脑补的贴图
但这还不是最痛的
最痛的是你看着效果图觉得挺好，买回来之后发现差一厘米，塞不进去
PPT 的元素左移三厘米没人在意
沙发宽了三厘米，就塞不进你家客厅
过去两年出来的 AI 设计工具不少
Disco Diffusion、Leonardo、Pika，火一阵就没了。活下来的 Midjourney、即梦，都在平面和视觉领域站住了脚 空间设计这个赛道，至今没有一个产品真正接住过
我跟燕虹聊到这个问题，她说了一个词：容错率
PPT 这种形式容错空间很大，元素往左往右，你又能怎么样呢
但空间设计需要的是「严肃且准确」的效果图
暗壳最早服务家居品牌方做电商图的时候，商家有 1000 个 SKU，他们对每一个 SKU 单独训练。形体可控、光影可控、材质可控
一个螺丝钉都不能有幻觉
三层约束，逐层加码
01精度——生成的图形体可控、光影可控、材质可控
02真实性——方案里那张沙发，能不能买到？买到的是不是图里那个？
03落地——上门量房、安装家具、调试灯光，AI 做不了
暗壳自己也坦诚，目前还没完全到达「严肃且准确」的百分百阶段
很多功能还在排期。但他们从 day one 就知道方向在哪
03 / MOAT
为什么纯技术团队做不了
燕虹 2022 年底看到生成式 AI 爆发，觉得想象空间极大
第一反应是去投一些技术团队
找了一圈——没有人分得清「东方设计」和「新中式」
能训模型，能调参数
但不知道一个 100 平的家怎么设计才合理，不知道 20 万预算做北欧风应该怎么分配，更不知道飘窗是什么时候出现在中国家庭里的
这些东西不在任何公开数据集里
最终矩阵决定自己做
政民牵头搞了一个「AI 先锋学会」，征集了 66 个人，分 10 个队，每周测试全球最新的 AI 工具
2023 年 5 月矩阵开始全员 AI 培训，所有设计师必须通过考核
考核「科三」是用 Stable Diffusion 出图。不学，不发年终奖
过程中他们发现了核心痛点：当时的 SD WebUI 加 ControlNet 对设计师太难用了
于是找了技术合伙人，加上燕虹在家居品牌审美和商业落地上的经验，暗壳就这么出来了
矩阵给暗壳的资源
→百万级的行业私有数据
→600+设计师作为测试和调优团队
→超百家家居建材品牌的供应链关系
直到 2025 年年中，暗壳才有自己的 Marketing 团队
之前客户都是基于矩阵的口碑自己找上来的
我跟燕虹聊到 Skills 的时候，她举了一个例子
去年暗壳服务阿里巴巴的大家居板块做生活方式白皮书，要决定扫地机器人到底放在家里哪里
这个问题扫地机器人制造商回答不了
放在阳台家政间的底下，前提是阳台的设计没有门槛，扫地机器人才能过去
这类经验无时无刻在产生，最终变成了 Agent 能调度的能力
空间设计的行业经验不是静态的
设计风向在变，户型在变，生活方式在变
!Image 13 这些变化只有在行业最前沿的人才能感知到
04 / FUTURE
服务人，也服务 Agent
聊到暗壳未来要服务谁，政民说：
不管是有界面的用户操作，还是其他 Agent 通过 API 来调用我们
我们都要做市面上最专业的空间设计 Agent
暗壳的服务对象不只是人类设计师，也包括未来的 Agent
从我自己的判断来看，这个方向大概率是对的
当 Agent 之间开始相互调用的时候，一个专精于物理空间理解的 Agent，会成为整个 AI 生态里很难被替代的一环
物理空间的经验和约束，通用大模型从互联网数据里学不到
面向 C 端的最终形态，可能就是手机上的一个对话界面
不需要复杂的画布，Agent 自动调取能力和资源完成设计推荐
什么时候能真正纯 ToC？燕虹的回答很坦诚：取决于 AI 的能力有多大
现阶段空间设计的容错率太低，还需要专业设计师在回路里做判断
万一真到 AGI 那一天，我们也是最早实现产业落地的
因为我们现在就开始做了
暗壳也在探索打通数字和物理接口的技术方向——NeRF、3D 高斯、空间深度测量
还有一个很少被提到的角度：暗壳积累的 3D 空间数据，未来可以作为具身智能的训练素材
这件事平面工具做不了
聊天快结束的时候，政民做了一个总结——
我们画布叫无限画布，满足无限创意。但落地必须用有限的产品来约束
在 AI 的世界里，空间和时间是无限的
在人类的现实世界里，时间和空间是有限的、受约束的
暗壳站在这两个世界的交界处
无限画布那一头，是创意和想象
有限物理世界这一头，是尺寸、材质、施工、一厘米的误差 两头都得接住，这个事才算成 跳转微信打开
赛 赛博禅心 @赛博禅心
One Sentence Summary
DarkShell AI is a vertical AI platform dedicated to physical space design. By engineering industry experience and combining it with real supply chain data, it solves the core pain points of "low fault tolerance" and "difficult implementation" that AI faces in space design.
Summary
This article provides an in-depth introduction to "DarkShell AI," an AI product vertical to the space design field. Founded by a team with deep architectural design background, DarkShell positions itself as "the Lovart for space design." Unlike Midjourney and other planar AI tools that focus on visual presentation, DarkShell emphasizes the "hard constraints" of the physical world—accuracy, realism, and implementability in design. Its core product architecture includes AI Agent, collaborative canvas, industry private data, and a supply chain database of 300,000+ real SKUs. The article points out that pure technology teams struggle to cross the threshold of industry experience. Through infusing over a decade of design logic into the model, DarkShell achieves a closed loop from vague requirements to precise procurement lists. In the future, DarkShell will not only serve human designers but also serve as a space-understanding Agent接入 AI ecosystem, while providing high-quality 3D training data for embodied intelligence.
Main Points
* 1. Physical space design has extremely low fault tolerance for AI, requiring "serious and accurate" output.
Unlike planar design, space design involves hard constraints such as dimensions, materials, and construction. Through deep training on SKUs, DarkShell ensures that generated schemes have controllable form, lighting, and materials, avoiding the common AI "hallucination" that makes schemes unimplementable.
* 2. Engineering encapsulation of industry experience is the core moat of vertical AI.
Pure technology teams struggle to understand the subtle differences between "Eastern design" and "New Chinese style." DarkShell leverages the million-level private data from parent company Matrix Shares and feedback from hundreds of designers to transform non-standardized design experience into capabilities that Agents can dispatch.
* 3. Bridging the closed loop between "infinite creativity" and "finite physical world" is the product's commercial value.
DarkShell not only provides an infinite canvas for creative brainstorming but also integrates an ecosystem library of 300,000+ real products for sale, enabling one-click generation of procurement lists and budgets after scheme confirmation, greatly reducing communication and modification costs.
* 4. Space design Agents will become important infrastructure for embodied intelligence and the AI ecosystem in the future.
The 3D space data accumulated by DarkShell and its deep understanding of physical environments can serve as training material for embodied intelligence in the future. Additionally, its professional capabilities can be invoked by other Agents through API, becoming an irreplaceable link in the AI ecosystem.
Key Quotes
* The physical world does not accept approximations. If a sofa is three centimeters too wide, it won't fit in your living room. * In the space design track, no product has truly delivered on this. When I discussed this issue with Yanhong, she used a term: fault tolerance. * Even a single screw cannot have hallucinations. * Our canvas is called infinite canvas, satisfying infinite creativity. But implementation must be constrained by finite products. DarkShell stands at the intersection of these two worlds. * When Agents start invoking each other, an Agent specialized in understanding physical space will become a part of the entire AI ecosystem that is very difficult to replace.
AI Score
83
Website mp.weixin.qq.com
Published At Today
Length 2557 words (about 11 min)
Tags
Space Design
Vertical AI
AI Agent
Supply Chain Integration
Embodied Intelligence
Related Articles
* GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * GLM-5 Technical Report Deep Dive | a16z: "The Best Open-Source Model" * 128. Manus Co-founder's Final Interview Before Sale: The Fantastical Drift to 2025... * 2025 Review: AI Is No Longer Theology, But Physics, the practical implementation of Agent technology, and the engineering ethos driving industry transformation.") * How to Design an AI Agent System * Deconstructing Clawdbot: Local Architecture, Memory Management, Agent Orchestration, and Context Assembly Principles * What You Don't Know About Claude Code: Architecture, Governance, and Engineering Practices - Tw93 * Claude Sonnet 4.6 Released * Gemini 3.1 Pro Released and showcasing multiple application demos ranging from code generation to end-to-end products.") * MiniMax M2.1: Enhanced Polyglot Programming Capabilities HomeArticlesPodcastsVideosTweets