Evaluation on AI-Generative Emotional Design Approach for Urban Vitality Spaces: A LoRA-Driven Framework and Empirical Research

Abstract

Recent advancements in urban vitality space design reflect increasing academic attention to emotional experience dimensions, paralleled by the emergence of AI-based generative technology as a transformative tool for systematically exploring the emotional attachment potential in preliminary designs. To effectively utilize AI-generative design results for spatial vitality creation and evaluation, exploring whether generated spaces respond to people’s emotional demands is necessary. This study establishes a comparative framework analyzing emotional attachment characteristics between LoRA-generated spatial designs and the real urban vitality space, using the representative case of THE BOX in Chaoyang, Beijing. Empirical data were collected through structured on-site surveys with 115 validated participants, enabling a comprehensive emotional attachment evaluation. SPSS 26.0 was employed for multi-dimensional analyses, encompassing aggregate attachment intensity, dimensional differentiation, and correlation mapping. Key findings reveal that while both generative and original spatial representations elicit measurable positive responses, AI-generated designs demonstrate a limited capacity to replicate the authentic three-dimensional experiential qualities inherent to physical environments, particularly regarding structural articulation and material tactility. Furthermore, significant deficiencies persist in the generative design’s cultural semiotic expression and visual-interactive spatial legibility, resulting in diminished user satisfaction. The analysis reveals that LoRA-generated spatial solutions require strategic enhancements in dynamic visual hierarchy, interactive integration, chromatic optimization, and material fidelity to bridge this experiential gap. These insights suggest viable pathways for integrating generative AI methodologies with conventional urban design practices, potentially enabling more sophisticated hybrid approaches that synergize digital innovation with built environment realities to cultivate enriched multisensory spatial experiences.

1. Introduction

Accelerating global urbanization has elevated the design of urban vitality spaces to a central concern in contemporary urban planning and architectural practice [1,2]. While conventional approaches prioritize functional efficiency and visual appeal, they frequently overlook people’s psychological needs and profound emotional attachment with built environments [3,4]. Empirical evidence demonstrates that human–space interactions and their associated emotional responses constitute fundamental determinants of spatial desirability and place attachment [5,6,7,8]. Research underscores that emotional experiences generated through these interactions significantly enhance people’s belongingness while shaping a space’s perceived attractiveness [9]. Nevertheless, current methodologies lack systematic frameworks for incorporating emotional science into spatial design practices or implementing technological solutions for targeted emotional design methodology [10,11,12]. This gap persists despite emerging capabilities in emotion-aware technologies, limiting designers’ capacity to create psychologically responsive built environments [13]. The complexity and interdisciplinary nature of emotional attachment hinder its quantification and systematic study, let alone the translation of relevant findings into actionable design methods [3,4,14,15,16,17]. While the big data methodology now enables multimodal data analysis, converting two-dimensional results into three-dimensional spatial designs remains challenging. The recent maturity of AI-generative design presents an opportunity to address this limitation [18,19].

Meanwhile, generative artificial intelligence (AI) models, such as Stable Diffusion and Midjourney, have shown significant potential in automating visual content creation [20,21]. AI-generative design is currently employed by practitioners for early scheme comparison and final rendering production in architectural and urban design [13,22,23]. However, their integration with psychological principles remains underdeveloped, often yielding outputs with a limited scientific validation of emotional relevance. This disconnect underscores the critical need to establish intelligent, emotion-driven design methodologies [23]. Current emotional design research predominantly employs subjective evaluations or qualitative analyses, lacking systematic, quantifiable, and replicable technical frameworks, which can be explored with the help of AI’s generative design capabilities based on massive data [13,20]. Conversely, while generative AI has advanced image synthesis capabilities, existing models prioritize formal aesthetics over psychological resonance, rendering them inadequate for addressing the nuanced emotional requirements of multifunctional urban environments. Therefore, this study further aims to address limitations in applying AI-generative technology to urban design, enabling its meaningful application for advancing human-centered spatial design rather than mere image generation. Bridging these interdisciplinary and technological divides to create a unified design and evaluation framework has emerged as a critical challenge for fostering emotionally intelligent urban vitality spaces.

To address these challenges, this study introduces an emotionally responsive design methodology leveraging low-rank adaptation (LoRA) technology, establishing a novel integration of generative AI with psychological evaluation systems through a dual-phase “intelligent generation–emotional quantification” framework. Taking a representative urban vitality complex, THE BOX in Beijing’s Chaoyang District, as an example, this study intends to explore the following to operate this framework:

(1)

The way to fine-tuning the Stable Diffusion model via LoRA to optimize output fidelity for context-specific urban scenarios is explored.

(2)

The possibility to automate text-to-panoramic design generation, combining ChatGPT 4.0’s semantic parsing with CycleGAN’s cross-domain style, is explored.

(3)

The construction and implementation of emotional attachment scales to quantitatively assess people’s attachment to original and AI-generated environments are explored.

(4)

The similarities and differences between multi-dimensional features of people’s emotional attachment to original and AI-generated environments are explored.

In this way, this study systematically integrates LoRA fine-tuning, multimodal text-image synthesis, and cross-domain style transfer techniques to address the limitations of single-model architectures in generating urban environments aligned with people’s emotional needs. By developing a dual-phase framework that synergizes generative AI with emotional attachment theory, it establishes an empirically verifiable technical framework for affectively optimized urban design. The results advance human-centric urban vitality design strategies while providing actionable insights for computationally augmented, psychologically attuned urban development in the AI era. Furthermore, the methodology offers urban planners and designers a systematic workflow to strategically implement AI tools for evidence-based built environment design and evaluation.

1.1. Design and Research of Urban Vitality Spaces in AI Era

The conceptual evolution of urban vitality space remains intrinsically tied to the interpretation of the core value of “vibrancy.” Initially framed by Jane Jacobs through the lens of built environment design as spatial attractiveness and activity driving force [24], this notion was expanded by Montgomery through emphasizing its experiential dimension of creating “liveliness” [25] perceived by people. After over half a century of interdisciplinary research, scholarly consensus acknowledges that urban vitality universally correlates with three elements across various perspectives and scales: people density, diverse behavior, and mixed function [14,26,27,28]. This understanding has directed the recent research focus toward urban vibrant complexes [29], which has also provided the contextual framework and reference for the research object selection for this study [14]. Though lacking a universal definition, the core concept of “vibrancy” inherently prioritizes the creation of life-sustaining built environments and human-centered spatial design methodology. In the digital era, achieving such vitality increasingly demands attention to advanced human needs, particularly emotional dimensions, during spatial design processes. The emergence and evolution of vibrant spaces fundamentally stem from interactions between people and built environments, where fulfilling emotional spatial needs constitutes the essence of vibrancy.

However, the advancement of digital technologies, particularly the maturation of data mining and analytical methods, has driven interdisciplinary scholars in urban vitality research to prioritize data-driven identification of vitality-influencing factors. Their focus predominantly centers on non-material dimensions such as socioeconomic and cultural dimensions [1,30,31,32], while spatial studies often reduce people’s three-dimensional behavioral patterns to simplified two-dimensional heatmaps or one-dimensional metrics through diverse data mining and analysis approaches [33]. Current research and design in the field of built environments rarely extend beyond street-scale analyses [32,34,35,36], with minimal attention to architectural spatial characteristics that are close to human scale and help shape people’s real emotional experiences. Existing discussions predominantly offer either qualitative descriptions of built environments or quantified indicators, failing to capture the interactive essence of vitality. This deficiency may stem from two critical gaps: first, the challenge of representing human-scale three-dimensional spatial interactions within conventional two-dimensional mapping frameworks common in geographical and urban planning studies; second, the absence of interdisciplinary theoretical frameworks capable of objectively evaluating human perceptions and emotional responses in urban vitality studies.

While urban vitality research has yielded substantial insights, it necessitates further methodological advancements in quantifying and evaluating emotional dimensions—the core essence of vitality. Meanwhile, the rapid development of AI-generative design tools has prompted some urban planners and architects to experiment with algorithmically generated vibrant urban complexes. However, the critical question of whether such AI-generated spaces authentically align with people’s emotional needs for urban vitality spaces remains unanswered, directly determining whether these tools genuinely enhance design quality or merely perpetuate technical formalism. Consequently, systematically evaluating the emotional attachment characteristics of both existing and AI-generated vibrant spaces holds dual significance: methodologically, it advances urban vitality research by integrating humanized evaluations into design frameworks; ethically, it establishes value-oriented guidelines for redefining design logic in an era where AI increasingly mediates the creation of urban environments [13,23]. This interdisciplinary inquiry bridges the gap between computational design efficiency and human experiential depth, ultimately steering AI’s role in urban design and research toward emotionally intelligent spatial solutions.

In this case, the low-rank adaptation (LoRA) model, a parameter-efficient fine-tuning technique, offers enhanced computational efficiency and output precision in generative AI systems. This capability positions generative AI as a transformative tool for urban vitality space design, enabling the data-driven optimization of complex built environments while addressing their emotional dimensions. Designers and researchers globally have employed this methodology across diverse spatial design dimensions [37,38,39]. Recent practices and findings have yielded particularly notable results, proving its effectiveness and offering foundational reference for this study. However, current research remains confined to generated forms and technical workflows of LoRA, largely unexplored in AI-generative design’s impact on real users’ spatial perception, which is the precise gap this study addresses.

1.2. Development and Application of AI-Generated Design in Urban Design and Related Fields

To evaluate the commonalities and distinctions in emotional attachment characteristics between real urban vitality spaces and those generated through AI-driven design methodologies, thereby optimizing the application of generative AI in urban vitality space design, it is imperative to systematically examine existing research and practical applications of this methodology in urban design contexts. AI-assisted design has successfully promoted the creation of various types of works in the field of art, including images, videos, music, and so on, and has received positive feedback from the public [40,41,42]. However, prior to deployment, generative AI systems require substantial model training to translate human design intent and expertise into computationally tractable representations [43,44]. The low-rank adaptation (LoRA) technique addresses this challenge through the parameter-efficient fine-tuning of foundation models, employing a trainable low-rank matrix decomposition to update network weights [37,38,39,45]. This approach achieves targeted model specialization while preserving baseline performance, drastically reducing computational and memory overhead compared to full-parameter retraining. By decoupling adaptation complexity from original model architecture, LoRA enables the resource-efficient optimization of generative systems [46]—a critical advancement for scalable applications in design domains where balancing computational pragmatism with output fidelity remains paramount [39].This LoRA-driven spatial image generation is widely adopted across disciplines such as industrial design, game development, and animation production [47,48]. Its rapid iteration capability offers a novel methodological approach to design exploration, which also influences urban design field [13].

Recent advancements demonstrate the growing efficacy of low-rank adaptation (LoRA) in generative urban design applications. Some researchers have employed LoRA to enhance text-guided architectural façade generation, minimizing dependence on extensive training datasets while integrating ControlNet for constraint-based refinement, thereby improving output precision and controllability [49,50,51]. Parallel research has systematically compared fine-tuning strategies for Stable Diffusion, validating LoRA’s superiority in achieving parameter-efficient model adaptation without compromising generative fidelity [52,53,54]. These studies collectively underscore LoRA’s dual strengths: (1) enabling high-quality domain-specific synthesis through low-rank matrix decomposition that preserves baseline model integrity, and (2) drastically reducing computational overhead via selective weight updates—a critical advantage for iterative design workflows. While LoRA’s technical merits are well-established in image generation, its application to specific urban vitality spaces remains underexplored, particularly in reconciling computational efficiency with the psychosocial complexity of public environments. The methodology’s capacity for rapid, targeted adjustments positions it as a transformative tool for AI-generated urban design, yet the absence of empirical implementations necessitates further research to optimize its integration with vitality space design and research, as well as human-centered emotional evaluation frameworks.

1.3. Definition and Measurement of People–Environment Emotional Attachment

The study of people–environment emotional attachment has evolved through interdisciplinary contributions spanning psychology, human geography, architecture, and social ecology. Pioneering concepts such as Wright and Tuan’s Topophilia (1960s–70s) [55] and Ralph’s sense of place (1970s) [56] have laid the foundation for later expansions by scholars including Krupat, Hidalgo, and Norberg-Schulz, who articulated multi-dimensional frameworks encompassing place identity [57], place dependence, place attachment [5], and genius loci [58]. Defined as the dynamic interplay of emotional, cognitive, and behavioral bonds between people and specific environments (both built and social), emotional attachment theory has continuously gained traction in urban design field as cities prioritize spaces that enhance public well-being [6,8]. Urban vitality spaces—critical loci for promoting activity and fostering happiness—exemplify this shift, serving as perfect research objects for adapting attachment principles (e.g., place-based rehabilitation strategies), while addressing complex socio-cognitive needs. This theoretical progression directly informs the focus of this study on vitality spaces as catalysts for emotionally responsive urban design. Meanwhile, this study further advocates integrating affective considerations into assessment frameworks for vital urban spaces.

Based on this, scholars have developed multi-dimensional measurement frameworks for emotional attachment, among which the Positive and Negative Affect Scale [59] (PANAS) for measuring emotional orientation is often integrated with the new problem-oriented scale for measuring emotional degrees. For the former, the reliability, validity, and rationality for PANAS’ application to architecture, landscape, and urban-related research have been confirmed in relevant studies [60,61]. For the latter, some researchers have constructed emotional attachment measurement scales for specific characteristics of the built environment based on relevant theoretical methods, and have explored their application in various types of cities and landscape spaces at home and abroad [3,14]. Among them, the characteristics are divided into material ones such as materials, structure, form, etc., and immaterial ones such as privacy describing the spatial experience, territoriality describing the culture and history, and changeability-related characteristics emerging with the digital age. Their results have demonstrated the effectiveness of the measurement methods. This methodological evolution not only confirms the efficacy of hybrid emotional attachment evaluation approaches but directly informs the current study’s development of semi-structured evaluation tools, where emotional attachment theory operates dually as a conceptual foundation and operational framework for systematizing people–environment emotional interactions. This approach not only addresses the existing demand for quantitative emotional analysis in urban vitality space research, but also establishes theoretical frameworks and value-oriented guidelines for AI-generated urban space design and evaluation, thereby bridging methodological gaps while steering technological applications toward emotion-oriented urban innovation.

2. Materials and Methods

2.1. Research Object: THE BOX, a Representative Urban Vitality Complex in Beijing

2.1.1. Background and Characteristics of THE BOX Chaowai, Chaoyang, Beijing

The THE BOX Chaowai project, situated in Beijing’s Chaoyang District, exemplifies a vibrant urban renewal initiative through the adaptive reuse of a 36,000-square-meter traditional commercial structure (Figure 1). It is now a contemporary urban complex blending commercial, cultural, and recreational programs to catalyze community engagement. Employing strategic spatial interventions rooted in curation-oriented, scenario-based, and gamified design principles, it reconfigures traditional relationships between consumers, commodities, and commercial environments. Functioning as a youth-centric hub and dynamic catalyst for urban vitality, the project has gained recognition for its capacity to stimulate social engagement and emotional attachment since its completion. This combination of demonstrated vibrancy and documented spatial challenges positions it as an exemplary case for investigating emotional attachment mechanisms in urban vitality spaces. Since opening nearly two years ago, THE BOX has gained significant public popularity and emerged as a government-recognized model for vibrant urban renewal complexes in Beijing. This can also be seen through the text data crawling and text sentiment research of well-known Chinese social platforms and review websites such as Weibo, Dianping, and Xiaohongshu. Therefore, THE BOX performs as a suitable research object, which enables the critical examination of whether AI-generated alternatives can replicate the original’s emotional attachment value while potentially addressing existing spatial challenges through this emotional-oriented AI-generation design methodology.

2.1.2. Original and Generated Images of Representative Spaces of THE BOX

The research selected representative spatial configurations within THE BOX as photographic stimuli representing “original spaces.” These visual materials were synergistically combined with bottom-up semantic mining outcomes from preliminary investigations—specifically spatial descriptors demonstrating positive emotions—to formulate input parameters for AI-driven generative design. This methodological integration yielded paired visual datasets, i.e., original vs. generated spatial images, for the latter’s emotional attachment scale evaluations, enabling a comparative analysis of attachment between existing and computationally derived spatial solutions.

Specifically, this study first employed the LoRA-based fine-tuning of the Stable Diffusion model to enhance its spatial design generation capabilities. Emotional-oriented key parameters were strategically adjusted to align with THE BOX’s specific characteristics, prioritizing the optimization of visual components, spatial organization, color coordination, and material configuration. The adapted model demonstrated improved alignment with predefined emotional design objectives in generated spatial outputs.

Then, using ChatGPT to analyze site photographs of THE BOX, this study generated architectural design briefs through semantic image analysis. These text-based descriptions systematically encoded both spatial attributes (structural configurations, functional zoning) and psychosocial requirements (emotional attachment criteria), combining with bottom-up semantic mining outcomes from preliminary investigations to establish a grounded framework for subsequent generative iterations. By embedding emotional design principles into programmatic specifications, the methodology ensured synthesized outputs addressed people–environment emotional attachment while maintaining spatial coherence.

The process leveraged LoRA technology’s text-to-image and image-to-image capabilities to generate comprehensive renovation proposals for THE BOX, including 2D renderings, panoramic visualizations, and targeted nodal perspectives. To ensure stylistic uniformity across spatial nodes, CycleGAN was integrated to synchronize design aesthetics, achieving both visual consistency and amplified continuity in emotional interaction. The comparative analysis of pre- and post-fine-tuning outputs validated LoRA’s efficacy in enhancing generative design quality. The optimized schemes exhibited superior emotional expression and spatial interactivity, demonstrating heightened visual appeal and emotional attachment.

The process consequently produced the comparative image series employed for the attachment measurement in this study. The generated spatial images and the corresponding original real space images are shown in Figure 2.

2.2. Emotional-Oriented LoRA-Driven Space Generation Workflow

2.2.1. Step 1: Model Pretraining

This study implements a pretrained Stable Diffusion model to ensure high-fidelity design image generation. The model employs a progressive denoising diffusion mechanism, iteratively reconstructing coherent images from random noise while learning hierarchical feature representations of spatial compositions. Through crossmodal alignment between visual data and textual descriptors, the framework synthesizes design-accurate images while preserving latent emotional cues embedded in language inputs, thereby operationalizing emotional design principles.

To optimize detail fidelity, a multi-resolution training regimen enhances the model’s capacity to learn hierarchical feature representations, spanning global compositional logic and textural details. This multi-scale strategy ensures emotionally attachment outputs that balance macro-level spatial legibility with micro-level material authenticity. The pretraining phase thus optimizes both output fidelity and emotional alignment, establishing a robust computational foundation for subsequent LoRA-based specialization and emotional-oriented urban vitality space synthesis.

2.2.2. Step 2: LoRA-Driven Fine-Tuning

Stable Diffusion, a leading diffusion model, excels in image synthesis through iterative denoising but exhibits constrained efficacy in emotional-oriented spatial design—particularly in regulating emotional spatial attributes and contextual detail fidelity. To address this, this study implements low-rank adaptation (LoRA) for the targeted enhancement of Stable Diffusion’s emotional design capacities. This approach enables the precise modulation of design-critical parameters—spatial topology, chromatic relationships, and material semantics—while maintaining computational efficiency. Applied to the case of THE BOX, the refined model demonstrated marked improvements across three emotional attachment dimensions: (1) Visual Legibility, achieved through chromatic intensity and material-textural coherence; (2) Interaction Potential, via spatial layouts amplifying people–environment reciprocity; and (3) Emotional Personalization, accommodating heterogeneous people’s emotion profiles through adaptive design variants.

2.2.3. Step 3: Design Description Generation

Based on Step 2, this study then synthesizes natural language processing (NLP) tools with generative design workflows by analyzing site-specific photographic data of THE BOX through a ChatGPT-enabled emotional spatial analysis. The framework systematically encodes visual-environmental node attributes into structured design briefs that integrate morphological characteristics and psychosocial qualities. By automating design brief generation through NLP, this methodology aligns textual outputs with evidence-based emotional profiles, embedding psychological principles into spatial syntax.

Building on this, LoRA-optimized text-to-image (T2I) and image-to-image (I2I) pipelines translate textual inputs into multimodal design proposals. The T2I framework generates orthographic projections, immersive panoramas, and contextually framed nodal perspectives, while the I2I process iteratively optimizes existing spatial representations through constraint-based style transfers. This generation strategy balances design novelty with attachment characteristics, ensuring generated environments amplify people–environment emotional interaction.

2.2.4. Step 4: CycleGAN Model for Style Consistency

Meanwhile, to ensure design coherence, this study implements a CycleGAN (Cycle-Consistent Generative Adversarial Network), enabling unsupervised style transfers across spatial nodes to harmonize aesthetic expression while preserving emotional continuity. The CycleGAN’s unpaired image translation capability resolves a critical challenge in multifunctional spatial design: reconciling stylistic diversity (e.g., commercial vigor vs. cultural serenity) with unified visual identity. By enforcing cyclic consistency in adversarial training, the framework mitigates stylistic fragmentation across functionally heterogeneous zones—a common pitfall in conventional approaches—while retaining site-specific emotional semantics. Applied to THE BOX case study, CycleGAN-mediated style unification elevates emotional expressiveness against baseline models (Figure 3), demonstrating that computational style regulation is not merely a matter of aesthetic optimization but a critical mechanism for emotional spatial design. This approach establishes a replicable paradigm for balancing programmatic diversity with emotional coherence in complex urban environments.

2.3. Emotional Attachment Scale Construction and Application

Emotional attachment evaluation is critical for assessing the effectiveness of LoRA-generated designs. By quantifying user emotional experiences, this study systematically evaluates the impact of spatial design on emotional responses. A multi-dimensional emotional attachment scale is developed to analyze user emotional engagement and compare LoRA-generated designs with original space of THE BOX.

Considering that this study intends to explore the dimension and degree of people’s emotional attachment to the real and LoRA-generated vitality space, key evaluation instruments include the Positive and Negative Affect Scale (PANAS), which evaluates emotional fluctuations induced by spatial designs, quantifying positive (e.g., pleasure, excitement) and negative (e.g., anxiety, discomfort) emotional responses using a Likert scoring method, as shown in

Table 1. Positive and Negative Affect Scale (The Likert scale was used as measurement method, where 1 represents very slightly or not at all and 5 represents extremely).

	Very Slightly or Not at All				Extremely
1. interested	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
2. distressed	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
3. excited	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
4. upset	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
5. strong	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
6. guilty	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
7. scared	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
8. hostile	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
9. enthusiastic	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
10. proud	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
11. irritable	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
12. alert	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
13. ashamed	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
14. inspired	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
15. nervous	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
16. determined	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
17.attentive	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
18. jittery	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
19. active	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]
20. afraid	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]

.

Additionally, a feature-specific emotional attachment scale, which examines the impact of visual and sensory elements (e.g., color schemes, material choices, and spatial layouts) on emotional experiences, is constructed, as shown in

Table 2. Specific spatial characteristics emotional attachment scale (The Likert scale was used as measurement method, where 1 represents very slightly or not at all and 7 represents extremely).

	Very Slightly or Not at All						Extremely
1. material	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
2. color	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
3. natural-related feature	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
4. form and structure	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
5. privacy	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
6. diversity	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
7. sociability	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
8. territoriality	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
9. playability	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
10. uniqueness	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]
11. changeability	1 [ ]	2 [ ]	3 [ ]	4 [ ]	5 [ ]	6 [ ]	7 [ ]

. Combined with the appeal of this study to explore the emotional attachment intensity of specific vitality space characteristics, a scale that describes the typical characteristics of THE BOX space, including material, color, natural-related features, form and structure, privacy, diversity, sociability, territoriality, playability, uniqueness, and changeability, was established. It is also measured on a Likert scale, with a score of 1 indicating that there is almost no emotional experience from this dimension, and a score of 7 indicating that a strong emotional experience is obtained from this dimension. The scale has been validated to be of good reliability in previous research [3,14,62].

SPSS 26.0 software was used to perform the data analysis process, including overall features, specific characteristics, and the dominant dimension of attachment to the original and generated vitality spaces. Specifically, overall emotional attachment features were analyzed according to descriptive statistical analysis. Specific attachment characteristics were revealed through the correlation (Pearson) analysis between positive affect, negative affect, and specific metrics describing different spatial characteristics. The dominant dimension of attachment was explored using exploratory factor analysis.

2.4. Participants

The emotional attachment assessment was conducted through on-site surveys involving 115 valid participants (37 males, 78 females). The sample size determination integrated the minimum requirement for reliable results using the emotional attachment measurement scale with precedents from prior studies. Specifically, the psychometric research-derived requirement stipulates a minimum sample size ≥ 35 for valid measurement results. Concurrently, existing research on place attachment and emotional computation in the built environment indicates that sample sizes > 80 yield valid and referenceable outcomes [3,4,5,14,15]. Participants sequentially viewed two matched vitality space image sets of identical scene types: Group A (original THE BOX spaces) and Group B (LoRA-generated variations), as illustrated in Figure 2, followed by the completion of emotional attachment scales. To ensure visual clarity, each image set was displayed for 1 min prior to the corresponding scale administration.

2.5. Research Framework

This study examines emotional attachment differences between LoRA-generated vitality space and the original vitality space based on the representative urban complex, THE BOX in Beijing, aiming to advance emotion-driven generative design methodologies. The research framework comprises four phases: (1) theory construction and case selection, (2) baseline emotional assessment through spatial evaluation metrics, (3) LoRA-enhanced design generation, and (4) emotional attachment evaluation (Figure 4).

Methodologically, this study first established emotional design criteria through a preliminary analysis of people’s responses to urban vitality spaces, taking Beijing’s THE BOX as the example. Subsequently, LoRA-driven generative techniques were implemented to optimize spatial configurations based on identified vitality characteristics. Finally, emotional attachment metrics were developed, and series scales were constructed and applied to quantitatively evaluate design outcomes, systematically validating LoRA’s efficacy in emotional spatial design.

3. Results and Discussion

3.1. Overall Emotional Attachment to Generative Design and Original Space

In general, as shown in

Table 3. Descriptive statistical results of the overall features of emotional attachment for original space images and generated space images.

	α	Original Space Images		Generated Space Images
		Mean	SD	Mean	SD
Positive affect	0.775	2.65	0.76	2.67	0.81
Negative affect	0.790	1.48	0.66	1.55	0.75
Overall attachment to specific characteristics	0.843	4.29	1.10	4.10	1.10

, there was no significant difference between the influence of images of the LoRA-generated space and the original space on people’s emotional attachment. Specifically, the performance of the two types of spatial images on people’s positive affect is almost the same (2.65 and 2.67), as shown in Table 3, where the difference is only 0.02, and the standard deviation indicates that different people have little difference in their experience. As for the negative affect, it can be seen from the table that the generated spatial images bring more negative experiences to people (1.55 > 1.48). Combined with the value of the emotional attachment intensity of specific spatial characteristics (4.10 < 4.29), it can be found that the score of generated ones is also low, from which it may be inferred that the environmental characteristics shown in the real THE BOX images can effectively enhance people’s emotional attachment and inhibit the generation of negative emotions to a certain extent. The correlation analysis and exploratory factor analysis were used to further explore the reasons for this result, especially the characteristics that influence people’s reaction towards them.

3.2. Specific Emotional Attachment Characteristics of Generative Design and Original Space

The correlation analysis was processed to further explore the extent to which each specific built environment characteristic contributed to the degree and dimension of emotional attachment. The specific results of the original space and generated design are show in

Table 4. The correlation analysis result for original space of THE BOX.

Variable	1	2	3	4	5	6	7	8	9	10	11	12	13	14
1. positive effect	1
2. negative effect	0.32 **	1
3. material	0.56 **	0.03	1
4. color	0.46 **	−0.14	0.64 **	1
5. natural-related feature	0.48 **	0.11	0.51 **	0.46 **	1
6. form and structure	0.59 **	−0.05	0.66 **	0.65 **	0.53 **	1
7. privacy	0.36 **	0.17	0.37 **	0.19*	0.52 **	0.34 **	1
8. diversity	0.44 **	−0.15	0.49 **	0.56 **	0.33 **	0.56 **	0.32 **	1
9. sociability	0.36 **	−0.13	0.43 **	0.45 **	0.21 *	0.51 **	0.20 *	0.52 **	1
10. territoriality	0.45 **	0.17	0.46 **	0.48 **	0.48 **	0.54 **	0.40 **	0.47 **	0.42 **	1
11. playability	0.47 **	−0.04	0.41 **	0.35 **	0.22 *	0.31 **	0.26 **	0.46 **	0.51 **	0.29 **	1
12. uniqueness	0.55 **	−0.01	0.49 **	0.53 **	0.37 **	0.55 **	0.30 **	0.63 **	0.42 **	0.55 **	0.52 **	1
13. changeability	0.55 **	0.03	0.55 **	0.46 **	0.33 **	0.51 **	0.37 **	0.52 **	0.50 **	0.53 **	0.54 **	0.61 **	1
14. Overall attachment to characteristics	0.68 **	0.01	0.77 **	0.74 **	0.65 **	0.79 **	0.55 **	0.75 **	0.65 **	0.73 **	0.61 **	0.77 **	0.76 **	1

*: The correlation is significant at the 0.05 level (two-tailed test); **: the correlation is significant at the 0.01 level (two-tailed test).

and

Table 5. The correlation analysis result for generated design of THE BOX.

Variable	1	2	3	4	5	6	7	8	9	10	11	12	13	14
1. positive effect	1
2. negative effect	0.21 *	1
3. material	0.49 **	0.02	1
4. color	0.46 **	−0.18	0.59 **	1
5. natural-related feature	0.40 **	−0.15	0.55 **	0.56 **	1
6. form and structure	0.43 **	−0.18	0.58 **	0.55 **	0.58 **	1
7. privacy	0.34 **	0.21 *	0.45 **	0.28 **	0.33 **	0.32 **	1
8. diversity	0.47 **	0.07	0.42 **	0.44 **	0.49 **	0.49 **	0.34 **	1
9. sociability	0.39 **	0.01	0.43 **	0.39 **	0.38 **	0.48 **	0.26 **	0.60 **	1
10. territoriality	0.45 **	0.24 *	0.42 **	0.24 **	0.34 **	0.32 **	0.42 **	0.41 **	0.48 **	1
11. playability	0.38 **	0.21 *	0.26 **	0.30 **	0.18	0.29 **	0.16	0.51 **	0.53 **	0.36 **	1
12. uniqueness	0.48 **	0.06	0.34 **	0.40 **	0.53 **	0.50 **	0.24 *	0.66 **	0.56 **	0.41 **	0.55 **	1
13. changeability	0.44 **	0.19 *	0.33 **	0.22 *	0.35 **	0.45 **	0.27 **	0.56 **	0.51 **	0.50 **	0.50 **	0.64 **	1
14. Overall attachment to characteristics	0.62 **	0.07	0.71 **	0.65 **	0.70 **	0.73 **	0.54 **	0.78 **	0.74 **	0.65 **	0.61 **	0.77 **	0.71 **	1

*: The correlation is significant at the 0.05 level (two-tailed test); **: the correlation is significant at the 0.01 level (two-tailed test).

as follows. Simultaneously, the correlation analysis results are displayed via heat maps (Figure 5 and Figure 6), enabling an intuitive comparison of the differential effects on people’s attachment between the original space and generative design of THE BOX.

In general, Table 4 and Table 5 reveal that the specific characteristics of the original space and the generated space were significantly correlated with a positive affect and overall emotional attachment, indicating that they significantly promoted people’s positive emotional experience and enhanced this intensity (each characteristic was significantly positively correlated with positive affect and overall attachment), which is also obvious according to the two heat maps.

Specifically, according to the results of the correlation data and heat map between the positive affect and the material characteristics of space including material, color, form and structure, and nature-related feature, it can be found that the attachment influence of the original space (0.562 **, 0.458 **, 0.478 **, 0.587 **) is stronger than that of the generated space (0.491 **, 0.456 **, 0.399 **, 0.427 **), which, to a certain extent, indicates that the real spatial images are still more emotionally attractive than the spatial images generated by AI, which proves the value of the real built environment and is consistent with previous studies [3,4,14].

Among them, for the four material characteristics, in the real space, the one that most significantly related to a positive affect is form and structure (0.587 **), followed by material (0.562 **). Combined with the correlation data and heat map between characteristics, it can be found that there is a very significant positive correlation between form and structure and material (0.581 **). In the generative space, the material characteristics most significantly associated with a positive affect is material (0.491 **), followed by color (0.456 **), and the positive correlation between material and color is also very significant (0.591 **). Meanwhile, the role of form and structure in generating spatial images is not as strong as that of real spatial images. On the one hand, the results indicate that the characteristics that are more effective in triggering people’s positive emotional attachment in real spatial images may be manifested in the three-dimensional structural relationship of space, while the characteristics that matters for attachment in generated ones are two-dimensional material color and material texture. The former has reinforced classical environmental behavior research while underscoring the critical role of tactile engagement in human interactions with physical space [8,55,56,58]. The latter advances a targeted operational framework for developing emotionally associative AI-generative designs, building directly upon prior research foundations [13,38,54]. On the other hand, it also indicates that the sense of space in the generated images is relatively weak, and it cannot replace the three-dimensional roaming sense given by the built environment, which supplements previous studies [21,40]. The results further demonstrate that THE BOX’s renovation, integrating contemporary materials with historic structures, established a compelling spatial composition. The authentic material characteristics and resulting spatial experiences prove determinative for people’s emotional responses [17]. This likely explains the generative design’s weaker attachment-promotion effect at the material characteristic level.

For the immaterial spatial characteristics that describe privacy, sociality, interaction, etc., there are both similarities and differences in the correlation analysis results between the original real space and the generated space. In terms of similar characteristics, the uniqueness of the original space and the generated space both showed obvious promotion effects on positive emotional attachment (0.552 **, 0.480 **), which are both in red, as shown in the heat maps. In addition to uniqueness, changeability and playability in the original space are more emotionally eliciting characteristics (rankings 2 and 3, 0.546 ** and 0.472 **, respectively), while these characteristics in the generative space are diversity and territoriality (rankings 2 and 3, 0.467 ** and 0.448 **, respectively). The above results reveal that spatial uniqueness has a very positive effect on emotions in both real environments and generative designs, and the real space may allow people to experience the fun of an embodied, operable space in a synesthetic way by looking at images. This is similar to previous utility results of perceptual measures using pictures [9,17], such as eye tracking-related research [3,63,64]. The results also reveal that spatial environments can trigger people’s emotional responses through immaterial characteristics alone via visual interaction, an approach demonstrating greater feasibility than reliance on material ones, which has also been discussed in a previous urban vitality study [65]. These findings offer actionable design implications for AI-generative systems, prioritizing factors that articulate spatial uniqueness and regionality.

In terms of negative emotional attachment, there is a clear difference between the original real space and the generated design space, which could also be inferred from the similar color shown in both heat maps. According to the correlation analysis results, the characteristics of the original space images, whether material or immaterial, are not significantly positively correlated with negative affect. Among them, the correlation coefficient between most of the characteristics and negative affect is negative, indicating that they inhibit the generation of negative affect to a certain extent. However, in the generative design images, there are unexpectedly characteristics that are significantly positively correlated with negative affect, mainly in the categories of immaterial characteristics, including privacy, territoriality, playability, and changeability (0.213 *, 0.236 *, 0.211 *, 0.187 *). As shown in the heat map of the generated design, the correlation results between its negative emotions and spatial characteristics are shown as significantly more “red circles”. The results indicate that these characteristics bring negative emotions to people, and to a certain extent, it also indicates that the LoRA-generated space images lack the basic feeling of privacy and security that people need when they are in a space and may also have a poor ability to show the historical and cultural characteristics of the real environment and the dynamic features that can be interacted with. This is something that has not been explored in existing research on AI-generative methods [38,43,47,48]. These findings suggest a pathway for enhancing future AI-generative design methods, prioritizing spatial generative designs that reflect historical and cultural connotations.

3.3. Dominant Emotional Attachment Dimensions of Generative Design and Original Space

In order to condense and refine the above specific spatial characteristics with minimal information loss and explore the dimensions that play a major role in the initiation and establishment of the attachment of both generative designs and original spaces to reveal their similarities and differences, an exploratory factor analysis has been performed using SPSS 26.0. The KMO and Bartlett sphericity test results of the scale data show that their structural characteristics meet with the conditions for a factor analysis (

Table 6. KMO and Bartlett’s test for scale data of original and generated space.

KMO and Bartlett’s Test		Original Space	Generated Design Space
KMO Measure of Sampling Adequacy		0.891	0.886
Bartlett′s test of Sphericity	Approx. Chi-Square	616.347	584.310
	df	55	55
	Sig.	0.000	0.000

). The analysis (EFA) results of emotional attachment scale data measured across both original space and generated design spaces of THE BOX are summarized in

Table 7. Exploratory factor analysis results of original space.

Specific Spatial Characteristics	Factor 1	Factor 2	Factor 3	Factor 4
Color	0.798
Form and Structure	0.746
Material	0.738
Territoriality		0.769
Uniqueness		0.724
Diversity		0.577
Playability			0.876
Sociability			0.659
Changeability			0.549
Privacy				0.888
Natural-related Feature				0.683

and

Table 8. Exploratory factor analysis results of generated design space.

Specific Spatial Characteristics	Factor 1	Factor 2	Factor 3	Factor 4
Color	0.850
Material	0.718
Natural-related Feature	0.716
Form and Structure	0.694
Changeability		0.779
Uniqueness		0.761
Diversity		0.574
Playability			0.861
Sociability			0.617
Privacy				0.838
Territoriality				0.704

.

Factor 1, representing the dimension with the strongest explanatory power for emotional attachment, demonstrated consistency across both the original and generated spaces through material characteristics of the built environment, with color emerging as the dominant sub-factor in both scenarios. This confirms the acknowledged principle that material characteristics of the built environment fundamentally shape human emotional responses [5,11,66], which can be found in both physical and generative design contexts. Key differences manifested in two aspects: First, the generative design incorporated additional nature-related features absent in original spaces, aligning with the increased presence of natural vegetation in Group B’s AI-generated images (Figure 2). Second, while form and structure ranked second in importance for original spaces, material superseded them in outputs of the generated design. These findings suggest that natural elements, particularly greenery, may exhibit immediate attachment-enhancing effects in urban environments through visual interaction mechanisms. This conclusion directly aligns with theoretical propositions on the healing value of natural environments within landscape research [3,7,12,14,64]. Meanwhile, generated spaces demonstrate heightened material clarity and chromatic vibrancy that may amplify visual appeal, yet this hyperreal aesthetic potentially diminishes architectural volumetric presence—a trade-off requiring critical evaluation in AI-generated design applications, which should be paid attention to in its future application.

Factor 2, representing the dimension with the stronger explanatory power for emotional attachment, captured dimensions of personalization across both original and generated spaces. The result revealed structural parallels yet critical compositional divergences. The shared presence of uniqueness and diversity as sub-factors with the same ordering suggests that AI-generated designs can effectively preserve spatial individuality comparable to authentic environments, maintaining equivalent attachment-enhancing capacities. However, the divergence manifests in the dominant sub-factors: territoriality emerged as the strongest contributor in original spaces versus changeability in generated designs. This discrepancy implies two critical insights which have been implemented previous researches [42,43,47,48]: (1) AI-generated environments currently fail to replicate the profound sense of place identity inherent to real physical urban contexts; (2) the heightened prominence of changeability potentially reflects people’s subconscious uncertainty and distrust toward generated “fake” environments, where spatial mutability may subconsciously signal impermanence or artificiality.

Factor 3 demonstrates relatively weaker explanatory utility for attachment formation. Its constituent sub-factors collectively indicate the spatial capacity to enhance human–environment or interpersonal interactions in both original and generated spaces, though with compositional variances. The shared hierarchical ordering of playability and sociality across both conditions suggests that generated designs may retain the ability to foster a playful experience and facilitate social engagement comparable to the original real space. A divergence manifests in the presence of the changeability factor exclusive to original spaces, potentially reflecting people’s stronger perception of environmental adaptability’s value for interactive behaviors in physical contexts. This is consistent with multidisciplinary perceptual research [10,13,64]. This absence in generated spaces may indicate limitations in simulating the perceived benefits of spatial flexibility inherent to real-world environments through current generative design paradigms, which is worth paying attention to in future AI-generated designs and research.

Factor 4 exhibits the weakest explanatory utility for emotional attachment formation. Its constituent sub-factors collectively address the spatial capacity to evoke perceptions of privacy across both original and generated environments. Notably, privacy associations diverged fundamentally between conditions: natural features influence privacy perceptions in original spaces, whereas territoriality influences privacy in generated designs. This result underscores two possible insights: First, the limited explanatory role of privacy in shaping emotional attachment within urban public vitality spaces possibly aligns with their inherent public-oriented atmosphere, consistent with previous research [14,32,35,65]. Second, the close relationship between privacy and nature-related features identified in original spaces to some extent indirectly highlights how THE BOX’s insufficient integration of natural landscape elements may have constrained users’ ability to experience privacy within its public realm, revealing an underexplored dimension in balancing communal vitality with individualized spatial experiences [28,30].

Notwithstanding these contributions to emotion-oriented AI-generated urban design for vibrant space creation, some key limitations warrant acknowledgment. The reliance on the single case study of THE BOX may fundamentally constrain the generalizability (external validity) of the findings. While offering rich contextual detail, this approach inherently lacks comparative data points, rendering it impossible to discern whether the observed phenomena, relationships, or outcomes are idiosyncratic to the specific context or representative of broader patterns. Meanwhile, failing to provide a rigorous justification for the determined sample size raises significant concerns regarding the statistical robustness and reliability of the results. Additionally, the presence of potential measurement biases within the methodologies employed to evaluate emotional attachment constitutes a substantial threat to the construct validity of this core variable.

At the same time, some critical issues in the AI-generated designs should be underlined. For example, cultural biases embedded in training data directly impact the trustworthiness of outputs for diverse users and influence the ethical responsibility inherent in ambiguous authorship. Furthermore, the opacity hinders the detection and correction of cultural misrepresentation. A critical analysis is therefore useful to examine how these factors co-constitute the power dynamics, potential harms, and ethical complexities within AI-generated design and creative production.

4. Conclusions

This study explored an emotion-driven design generation method based on LoRA (low-rank adaptation) technology, comparing LoRA-generated designs with the emotional attachment characteristics of the real built environments of the representative vitality space, THE BOX, in Beijing. The results show that, in general, both images of generative spaces and original spaces can bring positive emotional experience to people, but generative designs still cannot replace the positive emotional impact of the original real space, especially in the characteristics of forming a true three-dimensional experience of space, such as form and structure. At the same time, the ability of generative designs to convey spatial culture and help people perceive the interactivity of the space through vision is still weak, which will cause people to have a negative experience. The findings indicate that, compared to the original space, LoRA-generated designs need to enhance more spatial dynamism and interactivity by optimizing visual elements, spatial formation and structural layouts, attractive color schemes, and material selection. In the future, this approach may possibly hold promise for broader applications in urban vitality space design, further exploring how to integrate deeply with built environments to create richer and more multi-dimensional spatial experiences for users.

Generally speaking, AI-generated design methods enhance real-world urban design workflows and participatory processes through accelerated ideation, data-informed iteration, and scalable engagement. In design workflows of urban vitality spaces, generative tools such as LoRA technology rapidly produce prototypes or spatial configurations based on emotional-oriented parametric constraints, augmenting designers’ creativity with combinatorial efficiency. For public engagement, AI promotes the ability to synthesize community input into visualizable space alternatives—democratizing feedback loops while mitigating resource barriers. In participatory planning, these methods enable dynamic scenario modeling techniques that respond to emotional needs and foster inclusive deliberation by translating diverse stakeholder priorities into tangible options. However, critical oversight remains essential to address embedded biases, ensure cultural resonance, and preserve human agency in final decision-making. Specifically, from the theoretical and methodological perspectives, the conclusions of this study are as follows:

4.1. Practical Implementations of AI-Generated Technology in the Field of Urban Design

This study applied AI-generated LoRA technology to the field of emotional-oriented design for urban vitality spaces, significantly aiming at enhancing the emotional attachment effectiveness of LoRA-generated designs. Spatially, LoRA-generated designs improve physical attributes and play a key role in enhancing spatial interactivity and emotional attachment. They not only visually attract participants’ interest and positive responses but also effectively improve interactivity and comfort, significantly boosting urban vitality. In terms of detail, LoRA fine-tuning allows for precise adjustments of emotional elements, optimizing visual components, structural layouts, and appealing color and material selections. This fosters stronger emotional attachments between participants and spaces, greatly enhancing their sense of emotional attachment and psychological satisfaction. Overall, generative AI technology optimizes designs that have the potential to provide new intelligent solutions for emotional spatial designs and promote the application of cross-individual intelligence in the field. Specifically, AI-generated design methods to promote the formation of emotionally attached vibrant spaces can focus specifically on two levels: material and non-material generation methods.

4.1.1. Material-Level Generation: Focusing on Spatial Structural Characteristics That Elicit Embodied Experiences

At the level of the material characteristics of vitality space, based on the above results, it can be inferred that the influencing ability of generative designs on emotional attachment in describing the three-dimensional structural characteristics of space is still weaker than that of real spaces to a certain extent. Such characteristics often enhance people’s positive emotional attachment to space by triggering people’s perception of the embodiment of spatial experience. Therefore, in the future construction and adjustment of emotional-oriented generative design models, further attention can be paid to the generation methods of characteristics such as spatial materials, spatial structure, and spatial forms.

4.1.2. Immaterial-Level Generation: Focusing on Cultural and Interactive Characteristics Generation That Enhance Sense of Place

At the level of the immaterial characteristics of vitality space, the results show that generative designs lead to people’s negative experience to a certain extent due to the lack of the depiction of local cultural and historical characteristics of a certain space, as well as flexible interactive characteristics. This indicates that even when the spatial image is static, the cultural characteristics such as the territoriality and the interactive characteristics such as the playability of the spatial scene depicted are still valuable in enhancing a positive emotional attachment. Therefore, in the improvement of generative design models in the future, more technologies that can help generate characteristics to express spatial culture and interactivity should be added.

4.2. Theoretical Value of AI-Driven Emotional-Oriented Paradigm for Urban Vitality Space

By integrating the data analysis with emotional attachment, this study proposed a new paradigm for urban vitality space design that surpasses the limitations of traditional design approaches. Through the construction of a series of emotional attachment scales, this study compared the emotional attachment characteristics of LoRA-generated designs with those of built environments, further assessing their advantages and disadvantages compared with traditional design methods. Taking THE BOX in Chaoyang, Beijing, as an example, the analysis of 115 valid datasets validated the effectiveness of the “Generative AI—Emotional Attachment” dual-driven model in creating emotionally resonant spatial designs.

In the process of spatial design generation, iterative optimization models serve as a crucial approach to enhancing design quality and ensuring emotional attachment. To continuously refine the design scheme and improve its emotional guidance, this study employed a multi-round iterative optimization process based on LoRA technology. Through a cyclic feedback mechanism, the design scheme and research model underwent repeated optimization and adjustment, gradually achieving precise capture and response to users’ emotional needs.

Specifically, emotional evaluation and feedback were derived from a combination of interviews and surveys, leading to adjustments in LoRA parameters to maximize the generation of positive affect. After each iteration, the design was not only visually optimized but also fine-tuned in terms of emotional elements—such as color, material, and layout—based on emotional assessment results. This approach ensured that the design continuously approached an ideal state aligned with users’ emotional expectations, ultimately maximizing spatial vitality.

Through this iterative optimization mechanism, this study not only enhanced the emotional attachment of the design scheme but also expanded the depth and breadth of LoRA technology in emotion-driven designs. Each iteration provided valuable data and emotional insights for subsequent design generations, forming a positive feedback loop that offers a reliable theoretical and practical foundation for the emotional design of urban vitality spaces. In the future, generative AI can play an active role in more urban design projects, precisely meeting residents’ emotional needs and enhancing spatial vitality and comfort. The introduction of this method provides new perspectives for urban space design, driving it toward greater intelligence and emotional resonance.

4.3. Research Limitations and Future Work

Although this study provides strong empirical results, it is limited by its focus on a single case (THE BOX in Beijing), small number of subjects, and insufficient diversity, lacking discussions on diverse geographic and socio-cultural contexts, as mentioned above. This may restrict the generalizability of the findings. Additionally, the emotional scales used in this study may contain potential cultural biases, as they do not fully account for how different cultural backgrounds influence emotional responses. Emotional expressions toward the same scene may vary under different cultural influences, potentially leading to measurement validity issues in some emotional dimensions within existing scales, which could affect the reliability of conclusions. Furthermore, emotional complexity necessitates multi-dimensional scales and multimodal measurement methodologies.

Therefore, to advance both practical applications and theoretical frameworks, future work should prioritize the following directions. As for the research object, future research should validate findings across diverse urban vibrant spaces of varying scales and cultural contexts. Meanwhile, multimodal emotional measurement methods should be used to explore the differences and commonalities of emotional attachment across various cultural contexts. Human factor engineering, which studies emotional attachment characteristics through physiological feedback measurement, offers significant potential for application in emotion-focused built environment research. At the same time, how to further adjust the path of generating vitality spaces by the AI model to obtain emotional design results that are more conducive to people’s emotional attachment also needs further exploration. Integrating with the immersive environment through VR and AR devices to experience AI-generated spatial design would also help improve the diversity of evaluation perspectives. At the same time, the time dimension should also be included in the emotional evaluation of AI-generative design results in future research. For example, long-term user research methods from the field of industrial design can help achieve this goal.