夜雨聆风AI 一次可以生成很多图，却常常不说明为什么这样生成，也未必理解建筑表达的层级，更不会主动在发散与收敛之间切换。
AI can generate many images in one pass, yet the reasoning is often hidden, architectural abstraction may be misread, and the response does not deliberately switch between divergence and convergence.
01 · THREE AI LIMITATIONS
三项控制，对应三个 AI 问题
Three AI limitations, three corresponding controls
ConvoAI 的探索模式（Exploration Mode）针对的是传统生成式 AI 与通用大语言模型（LLM）在建筑设计中的三项局限：1. 单次图像生成缺少可追踪的推理过程；2. 通用模型对建筑表达的抽象层级认识不足；3. 客观中立的默认回答又不符合设计过程对发散与收敛的需要。
EN
ConvoAI's Exploration Mode addresses three limitations of conventional generative AI and general-purpose LLMs in architectural design: single-pass image generation hides the reasoning, architectural abstraction is not interpreted consistently, and neutral default responses do not match a design process that needs deliberate divergence and convergence.
01｜单次图像生成只有结果，缺少可追踪的推理过程
Single-pass generation hides the reasoning
传统图像生成通常把一段输入直接变成一张图。设计者看得到结果，却难以追踪 AI 为什么这样判断、哪些设计建议进入了下一轮。这就是可解释性不足（lack of explainability）。
EN
Conventional image generation turns an input into an output in a single step. The image is visible, but the reasoning behind the decision and what carries into the next round are difficult to trace. This is the explainability gap.
自我迭代（Self-iteration）要求 ConvoAI 按每一轮组织相关的设计分析、建议、文字与图像输入，以及生成结果。相邻轮次可以直接比较，推理过程与设计变化也更容易追踪。
EN
Self-iteration asks ConvoAI to organise the relevant design analysis, advice, text and image inputs, and generated outputs around each round. Successive rounds can be compared directly, making the reasoning and design changes easier to trace.
02｜通用大语言模型对建筑抽象层级认识不足
General-purpose LLMs lack architectural awareness
通用大语言模型（LLM）对体量（Massing）、深化（Development）、渲染（Render）这类建筑常用词及其具体程度，未必有稳定认识。你要求它继续探索白模，它可能直接补上材质、氛围和细节，把仍在讨论的体量过早推成渲染图。
EN
General-purpose LLMs do not always interpret architectural terms such as Massing, Development and Render at the intended level of abstraction. Asked to explore a white massing model, the AI may add material, atmosphere and detail, pushing an unresolved study into a render too early.
探索层级（Exploration Level）让设计者直接指定这一轮的建筑抽象层级。
EN
Exploration Level lets the designer specify the architectural abstraction for the current round.
01
体量（Massing）：保持抽象，讨论概念、空间关系与构成。
Massing stays conceptual and relational.
02
深化（Development）：把概念体量推进到建筑构成。
Development moves from massing towards architectural composition.
03
渲染（Render）：进入材质、功能与视觉表达。
Render addresses materiality, function and visual expression.
这三项描述的是建筑表达的具体程度，不是必须依次完成的工作流。
EN
These settings describe the degree of architectural resolution. They are not a prescribed sequence of workflows.
03｜客观中立，不等于设计过程需要的行为
Neutral answers do not provide the behaviour design needs
通用大语言模型倾向给出客观、中立、不过度下结论的回答。但设计过程并不总需要中立：概念阶段可能要刻意发散，方向明确后则需要收敛深化。缺少行为控制时，AI 容易停在通用建议，也可能过早锁定一个答案。
EN
General-purpose LLMs tend to produce objective, neutral answers. Design does not always need neutrality: an early concept may require deliberate divergence, while a clearer direction needs convergence and development. Without behavioural control, AI can remain generic or lock onto one answer too early.
设计行为（Design Behaviour）把选择交给设计者。第四版原型（Prototype 4）不再让系统自行判断这一轮该发散还是收敛。
EN
Design Behaviour gives that choice to the designer. Prototype 4 no longer asks the system to decide by itself whether the round should diverge or converge.
01
发散（Divergent）：打开不同路径。
Divergent opens alternatives.
02
收敛（Convergent）：围绕较明确的方向继续发展。
Convergent develops a clearer direction.
03
平衡（Balanced）：一边澄清问题，一边保留探索空间。
Balanced clarifies the problem while keeping room for exploration.
02 · FIRST EXPLORATION
第一次使用，怎样设定探索模式？
How to set up a first exploration
01
提供设计基础：上传草图、模型画面或当前设计图。
Provide a design basis: upload a sketch, model view or current design image.
02
补充背景与问题：说明项目条件、希望保留什么，以及这一轮想探索什么。
Add context and a question: describe the conditions, what should remain and what you want to explore.
03
设定三项控制：决定迭代承接、抽象层级与设计行为。
Set the three controls: decide the iteration context, abstraction level and design behaviour.
04
比较结果：检查哪些关系值得保留，再决定下一轮继续、转向或收敛。
Compare the outputs: identify what should remain, then decide whether the next round should continue, shift or converge.
03 · NEXT CASE
下一篇：从探索模式走进真实项目
Next: from Exploration Mode to a real project
下一篇，我们会看 PF Architects 如何把 ConvoAI 用在真实设计项目中。
EN
Next, we will look at how PF Architects used ConvoAI in a real design project.
REGISTER · EXPLORE
欢迎注册探索
中国版MAINLAND CHINA
convoai.cn
国际版INTERNATIONAL
convoai.design
ABOUT HAD STUDIO
关于 HAD Studio
浑然设计研发，前身为香港天成设计工作室有限公司，成立于 2016 年，是一家以建筑设计为主干的跨专业AI设计研发工作室，我们的项目分布于香港、内地及伦敦，客户涵盖发展商、建筑师、承建商、银行、大学，以及来自不同行业的机构与企业。
Founded in Hong Kong in 2016 as AND Studio Limited, HAD Studio is an interdisciplinary AI design and research practice rooted in architecture. Our projects span Hong Kong, Mainland China and London, serving developers, architects, contractors, banks, universities and organisations across industries.
