Objective As China’s urbanization construction enters the stage of quality improvement, landscape design practice presents the development trend of segmentation, customization and specialization. Designers need to study the design experience in segmented scenarios and come up with customized design strategies. The rapid development of the Internet industry and artificial intelligence technology is transforming the traditional working modes of landscape design, demonstrating immense potential in the fields of design evaluation research and computer-aided design. Against this backdrop, the research aims to explore the path of AI-aided design for customized landscape architecture scenarios and, by combining network big data and fine-tuning model technology of artificial intelligence generated content (AIGC), bridge the gap between big data analysis research and scenario generation in design practice, and construct a lightweight AI-aided landscape design workflow in response to the trend of segmentation, specialization, and customization in landscape design practice.
Methods The research adopts the Research for Design (RfD) methodology to build a workflow framework for integrating research and design practice. The workflow can be divided into four major processes: Network big data collection, intelligent evaluation of network big data, AIGC image fine-tuning model construction, and AI-aided design generation. 1) Network big data collection: Obtain datasets related to the required landscape architecture segmentation scenarios through online social platforms for evaluation and fine-tuning model training. 2) Intelligent evaluation of network big data: Analyze and categorize image data, and filter out the scenario images with excellent user evaluation based on text sentiment evaluation and subsidiary information analysis. 3) AIGC image fine-tuning model construction: Utilize the high-quality image dataset obtained in the previous stage to conduct fine-tuning model training based on a mature pre-trained general model, and inject relevant knowledge and experience from the sub-scenarios of landscape architecture in a cost-effective manner, thereby enhancing the model’s generative capabilities. 4) AI-aided design generation: Employ the fine-tuning model obtained through training to assist in generating scenario images according to the needs of design practice, and based on the intensity of control over the generated content, divide the aided scenario generation into “weakly controlled” and “strongly controlled” aided design scenarios. The research takes greenways as a typical example to verify the performance of the workflow framework. Image and text data related to Beijing greenways from 2013 to 2022 are collected from Weibo platform as the original dataset. Image features are extracted using the pre-trained convolutional neural network model InceptionResNetV2 and image data is clustered by K-means clustering algorithm. Through the process of image recognition and clustering, a total of 11 categories of images are obtained, including signage systems, pathways (cycling lanes, pedestrian walkways, and waterside promenades), recreational facilities (children’s playgrounds, elderly activity areas, fitness areas, and rest benches), flora (flowers), forested hills, trees, water bodies (hard and soft waterside spaces), sculptures, plazas, and pavilions/towers/halls. Among these, the image categories of pathways, water bodies, and plazas are selected as they are representative of greenway spatial scenarios. Consequently, these three categories of images are chosen as data samples for subsequent analysis and model training. SnowNLP is employed to analyze the sentiment of texts associated with images, obtain the sentiment score of images, and correct image score in combination with the microblog interaction data, with the corrected sentiment score being taken as the basis for image quality judgment. The image data collected is gradually cleaned and filtered into a training set for LoRA fine-tuning model training.
Results The LoRA model for the greenway scenario is successfully obtained by applying the workflow. The model can accurately reproduce the spatial details of the greenway scenario and the fitting degree is appropriate. Based on the two major scenarios of “weak control” and “strong control”, with the aid of prompt words and the ControlNet, two effective paradigms for generation control are proposed: Rapid generation of design intentions and redrawing of existing scenarios. In addition, the use of various fine-tuning models can realize the tasks of generating error control and drawing style migration.
Conclusion The workflow proposed in the research has some limitations. First, in terms of data collection, the image data from social platforms are not specialized enough, there are user preferences in the data, and the ownership of the dataset is unknown. Second, in terms of design-aided generation, there is a lack of relevant mature models, a lack of specialization control methods for landscape gardens, and a poor interpretation of the generated results. In future research, new technological tools should be combined to gradually improve workflow performance and continuously reduce the cost of workflow deployment. However, the workflow also shows potential for adaptation in landscape design, which is mainly reflected in two aspects: relevance and extensibility. The workflow provides a targeted path to real datasets based on real feedback from social media users on similar built cases. With the help of fine-tuned modeling techniques, it can be trained according to the specific needs of design practice. In addition, the workflow is scalable and can be quickly deployed for other tasks besides the greenway scenario practiced in the research.
DownLoad: