After conducting in-depth tests on various floor plan generation methods, I have been reconsidering the future development of this field and the paradigm of human-computer interaction. Ultimately, architectural elements like walls and floors are merely physical means to express space. In the underlying logic of design, the conceptualization of space should always precede the physical boundaries that make it implementable. However, looking at current floor plan generation tools, their underlying logic and interaction models seem to deviate from the true essence of architectural design.
Generating a floor plan layout involves an incredibly complex, tightly coupled system. A convincing architectural scheme must process a massive number of deeply intertwined variables, such as circulation flows, room dimensions, door placements and swing directions, furniture arrangements, natural daylighting, vertical circulation, and MEP systems. Given such intricate variables, expecting an algorithm to prove that its generated scheme is a comprehensive "global optimum" is entirely unrealistic. Most floor plan generation products currently on the market tend to offer a "magical button." With a single click, regardless of whether the underlying technology relies on specific algorithms or machine learning, the system instantly spits out a complete design. The credibility of this black-box model is remarkably low: even if the result is mathematically optimal, when architects are directly presented with a final outcome lacking a deductive process and concrete proof, their professional instinct is often to doubt it.
Furthermore, because it is impossible to one hundred percent exhaust all relevant factors influencing a floor plan within such a coupled system, this highly automated approach is particularly fragile. The moment a generated layout encounters a challenge from an unforeseen factor, the entire scheme may need to be completely overhauled, and the system's own logic might require reconstruction. If that happens, the very purpose of one-click generation is nullified. Even if the system generates a massive number of schemes at once for the designer to choose from, or allows the designer to pick one from numerous options at a certain stage, it fundamentally fails to circumvent this pain point.
Therefore, I believe the future form of such products should not be a generator that simply hands the "best design" to the architect, but rather a real-time feedback and guidance system. It should be deeply integrated into the design tools that architects use daily, providing natural, real-time interaction and evaluation as the designer iteratively arranges the space. This is much like performance-based analysis: just as moving a window instantly updates metrics for building energy consumption, indoor daylighting, and thermal radiation, the quality of a floor plan layout should also serve as a quantifiable decision-making indicator, working alongside other performance parameters to guide designers toward better outcomes. Concurrently, this must be a two-way feedback system. During practical use, if a designer utilizes their professional intuition to identify a glaring error made by the system, they should be able to intervene and correct it, feeding this correction back into the system for continuous learning and self-evolution.
To achieve this convincing, real-time feedback, I suggest fundamentally deconstructing the traditional scoring and generation mechanisms. When facing a massive coupled system like architectural space, we cannot accurately detect the boundaries and correlations of all scoring dimensions early on, making it unfeasible to evaluate a scheme from a singular, global perspective. The solution lies in decoupling the system, breaking down the floor plan evaluation framework into several independent modular components. During the design phase, designers can freely select the specific metrics they are most concerned with at that moment and evaluate the layout from those targeted perspectives.
For these deconstructed, independent modules, the evaluation methods should be built upon a dual-track verification mechanism: on one hand, relying on hard-coded traditional algorithms, and on the other, utilizing machine learning and human tagging based on extensive real-world samples. These two technological approaches should be highly complementary. Take circulation walking distance as an example: calculating the exact physical distance from one room to another can be accurately resolved using hard-coded algorithms. However, the subjective, design-intent-driven judgment of which specific rooms actually require the shortest walking relationship relies on machine learning combined with manual review for intelligent decision-making. Through this highly modular architecture, not only can the research and development work be efficiently divided, but the product itself retains immense flexibility. We can add or remove spatial quality evaluation modules at any time, thereby creating a new generation of spatial decision-support tools that truly aligns with the thought processes of architects.
- Junren Tan 2026.08.18
- Junren Tan 2026.08.18
在深入测试了多种平面图生成的方法后，我对于这一领域的未来发展以及人机交互模式产生了一些重新的思考。建筑的墙体、楼板说到底只是表达空间的一种物理手段，在设计的底层逻辑中，理应先有对空间的构想，而后才有这些让空间得以落地实施的物理边界。然而，审视目前的平面图生成工具，其底层逻辑和交互方式似乎偏离了建筑设计的本质。
平面图布局的生成面对的是一个极度复杂的耦合系统。一个具有信服力的建筑方案需要处理海量且相互深度影响的变量，例如流线组织、房间尺寸、门的位置乃至开启方向、家具布置、自然采光、垂直交通以及机电系统等。面对如此错综复杂的变量，要求一个算法去证明其生成的方案是囊括方方面面的“全局最优解”，在现实中是不切实际的。目前市面上大多数平面生成产品都倾向于提供一个“魔法按钮”，用户只需轻轻点击，不论其底层是基于何种算法或机器学习，都会瞬间吐出一个完整方案。这种黑盒模式的信服力极低：即使结果在数学逻辑上是最优的，但当建筑师直接面对一个缺乏推导过程和具体证明的最终结果时，往往会本能地对其产生怀疑。
此外，由于我们无法在系统中百分之百地穷尽所有影响平面布局的相关因素，这种高度自动化的系统显得尤为脆弱。一旦生成的布局遇到了某个未被预先考虑到的因素挑战，整个方案可能面临推倒重来，系统自身的逻辑也可能需要被重构，那么一键生成的意义就不复存在了。哪怕系统一次性生成海量的方案供设计师挑选，或者在某个阶段让设计师从众多盲盒中挑选一个，也无法从根本上规避这一痛点。
因此，我认为这类产品的未来形态不应该是一个直接施舍给设计师“最好设计”的生成器，而更应该演变为一个实时反馈与引导系统。它应当被深度内置于设计师日常使用的设计工具中，随着设计师对空间排布的推敲，提供如同呼吸般自然的实时交互与评价。这就如同在做性能化分析时，移动一扇窗户的位置就能实时看到建筑能耗、室内采光和辐射热的评测数据一样；平面布局的优劣也应该作为一种可量化的决策指标，与其他性能参数一同辅助设计师，引导他们走向更好的设计。同时，这必须是一个双向的反馈系统。在实际使用过程中，如果设计师凭借专业直觉发现了系统给出的明显错误，可以随时进行人为纠偏，并将这种纠偏反馈给系统进行持续的学习与自我进化。
为了实现这种实时且有说服力的反馈，我建议对传统的评分与生成机制进行彻底的解构化处理。面对建筑空间这样庞大的耦合系统，我们无法在初期精准探测到系统内所有评分维度的边界与关联，因此从全局单一维度来评价一个方案是不可行的。破局的思路在于将系统“解耦”，把平面图的评估体系拆解为若干个独立的评价模块。在设计阶段，设计师可以自由选择他们当下最关心的指标，从特定的模块切入进行评测。
对于这些解构后的独立模块，其衡量方法的建立应当基于双轨并行的验证机制：一方面是基于硬解码的传统算法，另一方面是基于大量实际样本的机器学习与人工标记。这两种技术路线应当是高度互补的。以流线步行距离为例，计算从一个房间走到另一个房间的具体物理距离，完全可以通过硬解码的算法来精准解决；但究竟“哪两个房间之间需要具备最短的步行关系”这种带有主观设计意图的评判，则需要依靠机器学习结合人工审查来进行智能决策。通过这种高度模块化的架构，不仅研发工作可以被高效拆解，产品也能保持极大的灵活性——我们可以随时在系统中添加或删减评估空间质量的模块，从而打造一个真正契合建筑师思维脉络的新一代空间辅助决策工具。
- Junren Tan 2026.08.18
- Junren Tan 2026.08.18