Architectural rendering recommendation method and electronic device

Abstract

This invention provides a method and electronic device for recommending architectural renderings, belonging to the field of artificial intelligence technology. The method includes: extracting multimodal features and latent representation vectors for each architectural rendering; obtaining a user's query intent vector and matching the query intent vector with the dimension key matrix corresponding to the multimodal features to generate weight coefficients for each feature dimension; calculating an initial recommendation score based on the multimodal features and weight coefficients, and filtering a candidate set from the architectural renderings based on the initial recommendation score; the candidate set includes multiple architectural renderings whose initial recommendation scores meet preset conditions; selecting and sorting architectural renderings from the candidate set based on the similarity values ​​of the latent representation vectors between the architectural renderings in the candidate set, and generating a final recommendation list. This invention improves the adaptability and diversity of architectural rendering recommendations.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and electronic device for generating architectural renderings. Background Technology

[0002] In recent years, generative artificial intelligence, especially text-to-image generation technology based on diffusion models, has been widely applied in architectural design, urban planning, and interior rendering. Diffusion models, such as Stable Diffusion, can generate highly realistic architectural renderings within seconds based on natural language prompts, greatly expanding the boundaries of architectural concept design. AI-powered architectural rendering sharing communities (or "community squares") built on these models have emerged, allowing users to publicly publish generated images with complete metadata, forming an open-source knowledge ecosystem. With the exponential growth of daily image data in these communities, providing accurate, personalized, and diverse image recommendations from a massive and inconsistent quality library to users with different backgrounds has become a pressing technical challenge.

[0003] For image recommendation and quality assessment, various solutions have been developed. In the field of general recommendation systems, the industry widely adopts architectures based on multi-stage funnel filtering models, including candidate set recall, coarse ranking, and fine ranking. Among them, the dual-tower model predicts click-through rates by encoding user behavior sequences and image content features respectively. Some systems also introduce dynamic reinforcement learning to extract stylistic and sentiment features of images. In terms of image quality and aesthetic assessment, traditional methods mainly use structural similarity indicators and peak signal-to-noise ratio to quantify distortion and noise. In recent years, there has been a shift towards using indicators such as FID and InceptionScore to evaluate the realism of generated images and text-image alignment. Some aesthetic assessment models predict the average opinion score of images by constructing large-scale labeled datasets and training convolutional neural networks. In addition, in the fields of multi-objective evaluation and data classification, existing technologies utilize Pearson correlation coefficients to identify indicator conflicts, conduct sensitivity analysis through weight perturbation and linear regression to establish baseline weights, or use long short-term memory networks to calculate time decay and supplement it with the PageRank algorithm to correct multi-dimensional feature weights.

[0004] However, the aforementioned existing technologies have significant drawbacks when directly applied to AI-powered architectural rendering sharing communities. First, general image evaluation models and aesthetic evaluation methods focus only on pixel-level realism, color harmony, or compositional rules, failing to grasp professional prior knowledge such as structural load-bearing capacity, spatial proportions, and perspective logic. This results in numerous "visual illusions"—images lacking structural overhangs or contradictory vanishing perspectives—being assigned high scores, severely damaging the community's professional reference value due to a lack of architectural rationality in the recommendations. Second, the lack of adaptive learning capabilities to adapt to users' instantaneous search intentions prevents the capture of user intent within a single session, leading to a misalignment between recommendations and users' actual needs. Finally, recommendation algorithms based on collaborative filtering or homogeneous node aggregation are prone to positive feedback loops, exacerbating aesthetic homogenization and information cocoons. Existing diversity control strategies, typically based on text labels or pixel differences, cannot accurately measure the essential convergence of the underlying style of generative models. Summary of the Invention

[0005] In view of this, it is necessary to provide a method and electronic device for recommending architectural renderings to solve the technical problems existing in the prior art, such as lack of architectural professional assessment, inability to adapt to the dynamic intentions of users in real time, and serious homogenization of recommendation results.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for recommending architectural renderings, comprising: Extract multimodal features and latent representation vectors from each architectural rendering; Obtain the user's query intent vector, and match the query intent vector with the dimension key matrix corresponding to the multimodal features to generate weight coefficients corresponding to each feature dimension; An initial recommendation score is calculated based on the multimodal features and the weighting coefficients, and a candidate set is obtained by filtering the architectural renderings based on the initial recommendation score; the candidate set includes multiple architectural renderings whose initial recommendation scores meet preset conditions. Based on the similarity value of the latent representation vectors among the architectural rendering images in the candidate set, the architectural rendering images are selected and sorted from the candidate set to generate a final recommendation list.

[0007] In one possible implementation, the extraction of multimodal features and latent representation vectors for each architectural rendering includes: A deep geometric consistency analysis was performed on the architectural renderings to obtain the visual rationality characteristics of the building. A multi-dimensional aesthetic decoupling evaluation was performed on the architectural renderings to obtain the architectural aesthetic quality characteristics; Cross-modal semantic matching was performed on the architectural renderings and their accompanying text prompts to obtain text-image alignment features; The deep interaction heat characteristics are obtained based on the cumulative number of various user interaction behaviors of the architectural renderings. A latent spatial encoder based on an image generation model extracts the latent representation vector of the architectural rendering.

[0008] In one possible implementation, the step of performing a deep geometric consistency analysis on the architectural renderings to obtain architectural visual rationality features includes: Extract the depth information from the architectural renderings to generate a relative depth map; Extract the geometric outline information of the architectural renderings to generate a structural wireframe diagram; The architectural rendering, the relative depth map, and the structural wireframe are fused together to form a multi-channel feature tensor. The multi-channel feature tensor is input into a pre-trained classification network, which outputs a probability value representing whether the architectural rendering satisfies architectural geometric consistency. The probability value is used as the visual rationality feature of the building.

[0009] In one possible implementation, the multi-dimensional aesthetic decoupling evaluation of the architectural renderings to obtain architectural aesthetic quality characteristics includes: Extract the first sub-feature related to light and shadow layers, the second sub-feature related to material representation, and the third sub-feature related to spatial composition from the architectural rendering; Based on the first sub-feature, the second sub-feature, and the third sub-feature, a comprehensive aesthetic score is calculated according to preset weight values. The comprehensive aesthetic score is normalized, and the normalization result is used as the architectural aesthetic quality characteristic.

[0010] In one possible implementation, the step of performing cross-modal semantic matching on the architectural rendering and its prompt text to obtain text-image alignment features includes: Extract the image feature vector of the architectural rendering; The prompt text is semantically decomposed to obtain multiple semantic feature vectors, including at least the building subject, architectural style, and environmental context; Calculate the cosine similarity between the image feature vector and each of the semantic feature vectors respectively; The text image alignment feature is calculated based on the cosine similarity and the magnitude of the image feature vector.

[0011] In one possible implementation, obtaining the deep interaction heat characteristic based on the cumulative number of various user interaction behaviors related to the architectural rendering includes: Obtain the cumulative number of various user interactions that occur with the architectural renderings on the platform; The total number of interactions is calculated based on the preset weights of different types of user interaction behaviors and their cumulative number of interactions. The exponential decay factor is calculated based on the release time of the architectural rendering and the current time. The deep interaction heat characteristic is calculated based on the total number of interactions and the exponential decay factor.

[0012] In one possible implementation, obtaining the user's query intent vector and matching the query intent vector with the dimension key matrix corresponding to the multimodal features to generate weight coefficients corresponding to each feature dimension includes: Obtain the text feature vector of the search text entered by the user in the current browsing session; Obtain the semantic tag distribution vector of the architectural renderings that the user clicked within a preset time window; Obtain the user's registration attribute vector; the registration attribute vector is a vector obtained by encoding the user's occupation category; The query intent vector is obtained by fusing and reducing the dimensions of the text feature vector, the semantic tag distribution vector, and the registration attribute vector. Four embedding vectors are constructed, each corresponding to one of the architectural visual rationality features, architectural aesthetic quality features, text-image alignment features, and deep interaction heat features. The four embedding vectors are then stacked to form the dimension key matrix. The query intent vector is used as the query term for cross-attention, and the dimension key matrix is ​​used as the key term for cross-attention. The dot product of the query intent vector and the dimension key matrix is ​​calculated. The dot product result is scaled and then input into a normalized exponential function to output the weight coefficients corresponding to each feature dimension.

[0013] In one possible implementation, after generating the final recommendation list, the process further includes: When a user performs parameter reuse or image regeneration operations on any architectural rendering in the final recommendation list, the interaction popularity feature of the architectural rendering is updated according to the operation type.

[0014] In one possible implementation, the step of selecting and sorting the architectural renderings from the candidate set based on the similarity value of the latent representation vectors between the architectural renderings in the candidate set, and generating a final recommendation list, includes: Calculate the similarity value between the potential representation vectors of every two architectural renderings in the candidate set; A greedy iterative algorithm is used to select one of the architectural renderings from the candidate set and add it to the final recommendation list; wherein the selected architectural rendering satisfies the condition of maximizing the objective function value, and the objective function is calculated based on the initial recommendation score of the architectural rendering and the similarity value of the latent representation vectors between the architectural renderings; Repeat the iterative selection process until the selected number reaches the preset recommendation number, thus obtaining the final recommendation list.

[0015] In a second aspect, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the recommended method for generating architectural renderings as described in any of the above implementations.

[0016] The beneficial effects of this invention are as follows: Firstly, by jointly extracting multimodal features and latent representation vectors, this invention can comprehensively characterize the attributes of architectural renderings across different perceptual dimensions, while simultaneously obtaining a compact representation for deep similarity measurement, providing a rich and decoupled feature foundation for subsequent recommendations. Furthermore, by generating weight coefficients through matching query intent vectors with dimensional key matrices, dynamic weighting of feature dimensions can be achieved, enabling the recommendation system to adaptively adjust the importance of different features according to the user's current needs, improving the targeting and flexibility of recommendations. Furthermore, rapidly filtering the candidate set based on weight coefficients can significantly reduce the computational load in the subsequent reordering stage, improving system response efficiency. Similarity reordering based on latent representation vectors can, while maintaining relevance to user intent, uncover the inherent semantic connections between images at a deep level, enhancing the diversity and coherence of recommendation results, and effectively improving the accuracy, personalization, and computational efficiency of architectural rendering recommendation results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic flowchart of an embodiment of the recommended method for creating architectural renderings provided by the present invention; Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of S101; Figure 3 For the present invention Figure 2A schematic diagram of an embodiment of S201; Figure 4 For the present invention Figure 2 A schematic diagram of an embodiment of S202; Figure 5 For the present invention Figure 2 A schematic diagram of an embodiment of S203; Figure 6 For the present invention Figure 2 A schematic diagram of an embodiment of S204; Figure 7 For the present invention Figure 1 A schematic diagram of an embodiment of S102; Figure 8 For the present invention Figure 1 A schematic diagram of an embodiment of S104; Figure 9 This is a schematic diagram of an embodiment of the multi-scale geometric consistency verification network of the present invention; Figure 10 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0021] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] This invention provides a method for creating architectural renderings and an electronic device, which will be described below.

[0024] Figure 1 A schematic flowchart of an embodiment of the recommended method for creating architectural renderings provided by the present invention is shown below. Figure 1 As shown, recommended methods for creating architectural renderings include: S101. Extract the multimodal features and latent representation vectors of each architectural rendering.

[0025] It should be noted that architectural renderings refer to two-dimensional digital images generated by users using text-to-image generation models (such as diffusion models) to express visual elements such as architectural design concepts, forms, materials, and lighting. Architectural renderings can be commonly used architectural representations such as exterior perspective, interior renderings, elevations, and axonometric drawings. For each architectural rendering in the input data, feature extraction techniques are used to obtain its multimodal features and the latent representation vector derived based on these multimodal features, either in parallel or sequentially. The multimodal features are used to characterize the image's attributes across different information dimensions, while the latent representation vector is used to provide a holistic and dense vector representation of the entire rendering in the implicit feature space.

[0026] S102. Obtain the user's query intent vector, and match the query intent vector with the dimension key matrix corresponding to the multimodal features to generate the weight coefficients corresponding to each feature dimension.

[0027] It's important to note the following steps: First, a vector representation of the user's query intent is obtained, indicating the task or information the user is interested in. This query intent vector is then matched against a pre-defined dimension key matrix of multimodal features. The correlation strength between the query intent vector and each feature dimension is evaluated using similarity calculations or attention mechanisms. Finally, a set of weight coefficients corresponding to each feature dimension is automatically generated based on the matching results. These weight coefficients reflect the importance or contribution of that feature dimension in responding to the current query intent, assigning higher weight coefficients to dimensions more relevant to the query intent and lower weight coefficients to dimensions less relevant, thus achieving differentiated attention to different dimensions.

[0028] S103. Calculate an initial recommendation score based on the multimodal features and the weight coefficients, and select a candidate set from the architectural renderings based on the initial recommendation score; the candidate set includes multiple architectural renderings whose initial recommendation scores meet preset conditions.

[0029] It should be noted that for each architectural rendering with extracted multimodal features and latent representation vectors, its multimodal features are weighted and summed with the generated weight coefficients to obtain the initial recommendation score for that architectural rendering. For example, when the user's intent leans towards structural rationality, the weight coefficient corresponding to structural rationality is larger, and its corresponding initial recommendation score will be significantly boosted. Then, all architectural renderings are sorted in descending order based on the initial recommendation scores, and the top K architectural renderings (e.g., the top 200) are selected as the candidate set. The architectural renderings in the candidate set initially satisfy the user's personalized preferences.

[0030] S104. Based on the similarity value of the latent representation vectors between the architectural rendering images in the candidate set, select and sort the architectural rendering images from the candidate set to generate a final recommendation list.

[0031] It should be noted that, for each architectural rendering in the candidate set, the similarity value between any two architectural renderings in the latent representation space is calculated using the latent representation vector of each previously extracted image. The similarity value reflects the degree of proximity or strong association between the two architectural renderings in terms of deep semantic features. Subsequently, according to preset selection rules (e.g., based on a reference image or the degree of association with user intent), the architectural renderings in the candidate set are filtered and sorted based on their similarity values. Specifically, architectural renderings with similarity below a preset threshold are removed, while those meeting the requirements are retained. Furthermore, the retained architectural renderings are arranged in descending order of similarity value (or descending order, depending on task requirements). Finally, the sorted image sequence is output as the final recommendation list for subsequent display or use.

[0032] In summary, the architectural rendering recommendation method provided in this invention firstly extracts multimodal features and latent representation vectors jointly, which can comprehensively characterize the attributes of architectural renderings in different perceptual dimensions, and simultaneously obtain a compact representation for deep similarity measurement, providing a rich and decoupled feature foundation for subsequent recommendations. Furthermore, by generating weight coefficients through matching query intent vectors with dimensional key matrices, dynamic weighting of feature dimensions can be achieved, enabling the recommendation system to adaptively adjust the importance of different features according to the user's current needs, improving the targeting and flexibility of recommendations. Furthermore, quickly filtering the candidate set based on weight coefficients can significantly reduce the computational load in the subsequent reordering stage, improving system response efficiency. Similarity reordering based on latent representation vectors can, while maintaining relevance to user intent, uncover the inherent semantic connections between images, enhancing the diversity and coherence of recommendation results, and effectively improving the accuracy, personalization, and computational efficiency of architectural rendering recommendation results.

[0033] In some embodiments of the present invention, such as Figure 2 As shown, step S101 includes: S201. Perform a deep geometric consistency analysis on the architectural renderings to obtain the architectural visual rationality characteristics.

[0034] It should be noted that: Deep geometric consistency analysis is performed on architectural renderings to assess whether the spatial structural relationships between the geometric elements in the renderings conform to real-world physical laws and perspective projection principles, thereby extracting feature information that reflects the visual rationality of the architecture. These architectural visual rationality features are used to quantitatively represent the realism and credibility of the rendering in terms of geometric construction, perspective consistency, scale proportions, and spatial coherence.

[0035] S202. Perform a multi-dimensional aesthetic decoupling evaluation on the architectural renderings to obtain the architectural aesthetic quality characteristics.

[0036] It should be noted that performing a multi-dimensional aesthetic decoupling evaluation on architectural renderings involves starting from multiple independent or separable aesthetic dimensions (such as color harmony, compositional balance, light and shadow coordination, material realism, and overall stylistic consistency), and decoupling and measuring the aesthetic attributes of each dimension separately to avoid mutual interference between different dimensions. Through this multi-dimensional decoupling evaluation, architectural aesthetic quality characteristics are ultimately generated that can quantitatively characterize the overall aesthetic level of the architectural rendering. These characteristics, presented as vectors or sets of scores, comprehensively reflect the degree of performance of the architectural rendering across various aesthetic indicators.

[0037] S203. Perform cross-modal semantic matching on the architectural rendering and its prompt text to obtain text-image alignment features.

[0038] It should be noted that a cross-modal semantic matching operation is performed on each architectural rendering and its corresponding cue text (i.e., the natural language description information used to generate or describe the architectural rendering). This cross-modal semantic matching process maps the visual modality of the architectural rendering and the linguistic modality of the text to the same semantic space. By calculating the similarity or correlation between their representations in this space, the semantic consistency between the content of the architectural rendering and the text description is quantitatively evaluated. Based on the results of the above cross-modal matching, text-image alignment features are extracted. The text-image alignment features reflect the degree of semantic consistency between the architectural rendering and its cue text in numerical or vector form. The higher the consistency, the more the architectural scheme presented by the architectural rendering matches the intent or requirements expressed in the text.

[0039] S204. Based on the cumulative number of various user interaction behaviors of the architectural renderings, the deep interaction heat characteristics are obtained.

[0040] It should be noted that for each architectural rendering, the cumulative number of occurrences corresponding to various types of user interaction behaviors is counted. User interaction behaviors include, but are not limited to, recordable user actions such as clicking, browsing, saving, sharing, rating, zooming in, and rotating views. By accumulating or weighting the number of these different types of interaction behaviors, a quantitative indicator that comprehensively reflects the degree of user attention and participation in the architectural rendering is obtained, namely, the deep interaction popularity characteristic. The deep interaction popularity characteristic not only reflects the static popularity of the architectural rendering, but also characterizes the dynamic preferences of users in the long term or short term from multiple dimensions of behavioral patterns, providing a basis for subsequent recommendation ranking or quality assessment based on user feedback.

[0041] For example, for any set of image candidates in a community square The system executes four feature extraction pipelines in parallel, and finally outputs a four-dimensional basic evaluation matrix (i.e., multimodal features) representing the architectural rendering. Parallel execution refers to triggering four independent computation tasks (processes / threads) simultaneously in the background for any candidate image within the community square. These four pipelines are independently responsible for extracting architectural visual rationality features, architectural aesthetic quality features, text-image alignment features, and deep interaction heat features from the image, respectively. This asynchronous parallel processing mechanism significantly reduces the system's inference latency. The four-dimensional basic evaluation matrix is ​​a four-dimensional vector (or one-dimensional matrix) formed by concatenating the four scalar scores calculated by the above four parallel pipelines. For example, for the architectural rendering generated by user A... The constructed multimodal features are =[0.88, 0.94, 0.82, 0.00].

[0042] S205. A latent spatial encoder based on an image generation model extracts the latent representation vector of the architectural rendering.

[0043] It should be noted that before the architectural rendering is generated, the latent spatial tensor is directly extracted from the output of the variational autoencoder (VAE) of the image generation model (such as StableDiffusion). Its shape is typically C×H×W (e.g., 4×64×64). This tensor is flattened into a one-dimensional vector, which serves as the latent representation vector of the image and is stored in a vector database. The latent representation phasor preserves the essential features of the image at the underlying manifold level, such as the compositional skeleton, spatial layout, and core semantics, and is insensitive to surface texture and color changes.

[0044] In this embodiment, deep geometric consistency analysis can quantitatively evaluate the realism and rationality of architectural renderings in terms of three-dimensional geometry, eliminating low-quality images with perspective distortion or scale distortion, and improving the realism of the recommendation results. Furthermore, multi-dimensional aesthetic decoupling evaluation allows for scoring from multiple independent aesthetic dimensions, avoiding interference between aesthetic indicators and thus more precisely characterizing the visual appeal of the image. Furthermore, cross-modal semantic matching ensures a high degree of alignment between the recommendation results and the user's input text description, enhancing the semantic understanding capability of the recommendation system. Furthermore, deep interaction popularity features introduce collective intelligence, utilizing user historical behavior to reflect the actual popularity of the image, making the recommendation results more aligned with public preferences. Finally, based on the latent representation vector extracted by the latent spatial encoder of the generative model, a compact and semantically rich image representation can be obtained, providing an efficient and reliable feature foundation for subsequent similarity re-ranking. Through the joint extraction of multi-dimensional features and latent representation vectors, the geometric rationality, aesthetic quality, semantic alignment, user popularity, and deep semantics of architectural renderings can be comprehensively and accurately characterized, laying a solid data foundation for a high-quality recommendation system.

[0045] In some embodiments of the present invention, such as Figure 3 As shown, step S201 includes: S301. Extract the depth information of the architectural rendering and generate a relative depth map.

[0046] It should be noted that when an architectural rendering is input into a pre-trained monocular depth estimation network, the network can infer the relative depth value of each pixel from a single 2D image and output a dense relative depth map with the same resolution as the original image. Brighter areas in the depth map represent areas spatially closer to the observer, while darker areas represent areas farther away, thus characterizing the spatial depth relationships within the image.

[0047] S302. Extract the geometric outline information of the architectural rendering and generate a structural wireframe diagram.

[0048] It should be noted that the architectural rendering is input into a noise-robust edge detection operator (such as the deep learning-based HED algorithm or an improved version of the traditional Canny operator) to extract the main geometric contour lines in the image. This operator can suppress interference from surface information such as texture, lighting, and material color, retaining only pure geometric features such as structural transitions, door and window openings, and roof outlines on the building facade, and outputting a binarized or grayscale structural wireframe.

[0049] S303. The architectural rendering, the relative depth map, and the structural wireframe are fused together to form a multi-channel feature tensor.

[0050] It should be noted that: First, the spatial resolution of the relative depth map and the structural wireframe map is forcibly aligned to the resolution of the original architectural rendering (e.g., H×W) using a bilinear interpolation algorithm, and then the pixel values ​​of the three are uniformly mapped to the [0, 1] interval using max-min normalization. Subsequently, the original image (three RGB channels), the relative depth map (single channel), and the structural wireframe map (single channel) are stitched together along the channel dimension to form a multi-channel feature tensor (e.g., 5×H×W).

[0051] S304. Input the multi-channel feature tensor into a pre-trained classification network, output a probability value representing whether the architectural rendering satisfies architectural geometric consistency, and use the probability value as the visual rationality feature of the building.

[0052] It should be noted that a dataset containing tens of thousands of labeled samples is pre-constructed. Positive samples are real photos or high-quality generated images that conform to architectural geometry (e.g., continuous structural supports, correct perspective vanishing points), while negative samples are AI-generated images with structural illusions (e.g., unsupported cantilevered structures, multiple vanishing points, disproportionate door and window dimensions). Binary classification transfer learning is performed on a residual network (e.g., ResNet-50 architecture) based on this dataset. After training, the previously generated multi-channel feature tensor is input into the classification network. After global pooling and fully connected layers, a probability value in the (0, 1) interval is output through a Sigmoid activation function. The closer this probability value is to 1, the more consistent the image's structural proportions, support relationships, and perspective logic are with real-world architectural geometry.

[0053] For example, the original architectural renderings Two pre-trained vision models are input in parallel, such as Figure 9 As shown, the two pre-trained visual models are a large monocular depth estimation model (such as MiDaS) and a noise-robust edge detection operator (such as a fine-tuned Canny or HED algorithm). Among them, the architectural rendering... Simultaneously, the data is fed into both the MiDaS branch and the Canny branch. The MiDaS branch is used to extract the dense relative depth map of the image. This characterizes the spatial depth relationship. The Canny branch is used to extract the structural wireframe diagram of the image. This is done to remove the interference of color and light and shadow. The original architectural renderings... Relative depth map With structural wireframe diagram Channel concatenation is performed to construct a multi-channel feature tensor. Specifically, due to the architectural renderings... Relative depth map With structural wireframe diagram The resolutions may vary. Before channel stitching, a bilinear interpolation algorithm is used to extract the relative depth map. With structural wireframe diagram Force spatial resolution alignment and scaling to architectural rendering resolution (e.g.) Simultaneously, Min-Max Normalization is used to uniformly map the pixel values ​​or feature values ​​of the three to the [0, 1] interval, and then the three are concatenated along the channel dimension to form a multi-channel feature tensor. Then, the multi-channel feature tensor is input into a deep residual network (based on the ResNet-50 architecture), which is used to extract high-level semantic features to obtain a two-dimensional feature map. The deep residual network is pre-classified and transferred to a dataset labeled with tens of thousands of "correct buildings" and "AI illusion buildings" (such as missing supports, excessive cantilever, and multiple vanishing point perspective errors). Among them, missing supports refer to the ratio of the length of the cantilever part of the structure to the width of the root support structure being greater than the safety engineering threshold (such as exceeding 1:5), and no obvious suspension or cable structure can be identified in the image. Subsequently, the two-dimensional feature map output by the deep residual network is passed to a global pooling layer, which compresses the two-dimensional feature map into a one-dimensional feature vector. The one-dimensional feature vector passes through a fully connected layer and outputs a logit (denoted as ). Finally, the probability value is calculated using the following formula; this probability value represents the building's visual rationality characteristic. , to any real number field Nonlinear mapping to the (0, 1) interval. Architectural visual rationality characteristics. The higher the value, the more the image's structural proportions, door and window alignment, and viewpoint perspective logic align with real-world engineering and architectural understanding. Among these, multiple vanishing point perspective error refers to the error quantified when extracting linear features from an image using the Hough Transform. If, when the original image (based on the Prompt) is determined to be in single-point or two-point perspective, the extracted structural lines, which should be parallel, converge in the image space or its extended space to more than a expected number (e.g., >2) of independent vanishing points, and the distance between these vanishing points is greater than a set threshold, then this is defined as a perspective error.

[0054] In this embodiment, by extracting the relative depth map and structural wireframe, pseudo-3D spatial relationships and geometric contours can be rapidly reconstructed in the 2D image space. This provides information from two complementary dimensions—spatial depth and pure geometric structure—for subsequent analysis, avoiding material texture interference caused by relying solely on RGB pixels. Furthermore, through channel fusion, the original image, depth map, and wireframe are integrated into a unified multi-channel feature tensor, allowing information from different modalities to mutually enhance each other at the feature level, improving the classification network's comprehensive ability to discriminate complex geometric contradictions. Moreover, by inputting the multi-channel feature tensor into a pre-trained classification network to output geometric consistency probability values, end-to-end automatic scoring of architectural rationality is achieved, significantly reducing computational overhead while improving the credibility of recommended content.

[0055] In some embodiments of the present invention, such as Figure 4 As shown, step S202 includes: S401. Extract the first sub-feature related to the lighting and shadow layer, the second sub-feature related to the material expression, and the third sub-feature related to the spatial composition from the architectural rendering.

[0056] It should be noted that the evaluation of architectural aesthetics is not limited to basic color harmony or the rule of thirds composition, but requires a comprehensive assessment of light and shadow texture and material representation. Therefore, this embodiment deploys an aesthetic scoring model based on the Visual Transformer (ViT) architecture. The overall structure of the aesthetic scoring model includes an input embedding layer (composed of image patch embeddings and learnable classification tokens), multiple stacked Transformer encoding layers (each encoding layer containing multi-head self-attention and a feedforward neural network), and an output layer (linear probe). The aesthetic scoring model performs fine-grained decoupled scoring of architectural renderings through a multi-head self-attention mechanism. The scoring dimensions include three sub-features: light and shadow layering and contrast, material fidelity and expressiveness, and spatial composition tension and balance. Three independent, learnable classification tokens, named [CLS_light], [CLS_mat], and [CLS_comp], are additionally concatenated at the input of the aesthetic scoring model structure. The specific processing flow is as follows: first, the architectural rendering is divided into fixed-size image patches and linearly projected into sequence tokens, which are then fed into the multi-head self-attention module along with the three classification tokens mentioned above. In the multi-layer, multi-head self-attention module, each of the three classification tokens interacts with the image feature sequence, aggregating corresponding lighting, material, and composition features. Finally, three parallel linear probes output corresponding scalar scores. Specifically, during the multi-layer self-attention interaction, the [CLS_light] token actively aggregates features related to lighting contrast, lighting direction, and shadow levels; the [CLS_mat] token aggregates features related to material surface reflection, texture details, and realism; and the [CLS_comp] token aggregates features related to architectural compositional balance, axial relationships, and visual center of gravity distribution. After all encoding layers, the three classification tokens output their corresponding feature vectors, which are then each input into an independent linear probe, outputting the first, second, and third sub-feature scores, respectively.

[0057] S402. Based on the first sub-feature, the second sub-feature, and the third sub-feature, a comprehensive aesthetic score is calculated according to the preset weight values.

[0058] It should be noted that the fusion weights of the three sub-features are pre-trained offline on a large-scale architectural aesthetic annotation dataset (e.g., images containing different architectural styles, rendering qualities, composition methods, and corresponding aesthetic scores) or determined through expert experience, and are denoted as follows: , , And satisfy + + =1. Score the first sub-feature. Second sub-feature score Third sub-feature score The overall aesthetic score is obtained by linear combination calculation using the following formula: ; in, Fixed parameters determined for offline training (and ).

[0059] S403. Normalize the comprehensive aesthetic score and use the normalization result as the architectural aesthetic quality characteristic.

[0060] It should be noted that: due to the calculated comprehensive aesthetic score It may exceed the (0, 1) interval; therefore, the overall aesthetic score will be... Normalization is performed to a uniform (0, 1) interval to facilitate weighted calculations with other features (architectural visual rationality features, text-image alignment features, and deep interaction heat features). Normalization methods can employ min-max scaling (based on upper and lower bounds of scores statistically derived from the training set) or mapping via the Sigmoid function. For example, if the theoretical maximum value of the comprehensive aesthetic score in the dataset is known to be M (empirically 10), then the architectural aesthetic quality feature... = / M, and truncate to the interval [0, 1].

[0061] In this embodiment, by using multiple learnable classification tokens in the visual Transformer model to interact with image features through multi-head self-attention, the light and shadow layers, material representation, and spatial structure can be decoupled and extracted. Figure 3 The three independent aesthetic sub-features make aesthetic assessments interpretable and targeted. Furthermore, by linearly combining and normalizing these three sub-features according to preset weights, the focus of aesthetic evaluation can be flexibly adjusted to meet the needs of architectural design professionals, while ensuring consistency with other feature dimensions. This allows recommended architectural renderings to be finely weighted based on the different importance of light and shadow, materials, or composition, thereby more accurately matching architects' different preferences for visual effects in actual creation.

[0062] In some embodiments of the present invention, such as Figure 5 As shown, step S203 includes: S501. Perform semantic decomposition on the prompt text to obtain multiple semantic feature vectors, including at least the building subject, architectural style, and environmental context.

[0063] It should be noted that the architectural renderings are generated from user-input prompt text using an image generation model (such as a diffusion model). Therefore, there is a semantic correspondence between the prompt text and the architectural renderings. A text encoder using a contrastive language-image pre-trained model (such as the SigLIP or CLIP architecture) is used to convert the semantically decomposed text fragments into corresponding semantic feature vectors. Before feeding the user-input prompt text into the text encoder, the prompt text is first semantically decomposed using natural language processing techniques (such as a lightweight word segmenter based on dependency parsing or keyword extraction). For example, if the prompt text is "A futuristic parametric cultural center, floating over a lake, structural steel diagrid, soft golden hour lighting", the prompt text is divided into three semantic fragments: the main building (e.g., "parametric cultural center"), the architectural style (e.g., "futuristic, structural steel diagrid"), and the environmental context (e.g., "floating over a lake, soft golden hour lighting"). Each semantic fragment is then input into the text encoder to obtain the corresponding semantic feature vector of the main building. Semantic feature vector of architectural style Contextual semantic feature vector .

[0064] S502. Calculate the cosine similarity between the image feature vector and each of the semantic feature vectors respectively.

[0065] It should be noted that the image encoder, which uses a contrastive language-image pre-trained model, extracts the image feature vectors from the architectural renderings. The specific extraction process is as follows: input the architectural rendering into the visual Transformer branch of SigLIP, extract the last layer of features and normalize them to obtain the image feature vector. (Normalized to unit length or retaining the original modulus). For each semantic feature vector obtained in step S501 (also output by the encoder and optionally normalized), calculate its cosine similarity to the image feature vector: ; Among them, the closer the cosine similarity is to 1, the more consistent the architectural rendering is with the prompt text entered by the user in that semantic dimension.

[0066] S503. The text image alignment feature is calculated based on the cosine similarity and the magnitude of the image feature vector.

[0067] It should be noted that, given that architectural cues are typically long and structurally complex (e.g., "subject + environment + medium + renderer"), this embodiment employs a multi-granularity alignment calculation to determine text-image alignment features. First, weights for three sub-dimensions are set based on architectural expertise or offline statistics. , , (For example , , This reflects the highest priority of subject identification in architectural design, followed by style, and lowest priority of environment. Subsequently, the cosine similarity sum is calculated based on the cosine similarity and the magnitude of the image feature vector: ; To eliminate the scaling effect caused by the magnitude of the image feature vector (if not normalized), the sum of cosine similarities is divided by the magnitude of the image feature vector. (like Normalized values ​​have a modulus of 1 (division operations are retained to ensure generality). The result is then used as the text-image alignment feature. : ; The formula is calculated to obtain It can accurately quantify the extent to which the generated image executes the semantic intent of complex architecture. If it is necessary to limit the output range, it can be further mapped to the (0,1) interval through Sigmoid or clamping operations.

[0068] In this embodiment, multiple semantic feature vectors of the building subject, style, and environmental context are obtained by semantically decomposing the prompt text. This achieves fine-grained decomposition of complex architectural design intentions, making semantic alignment evaluation no longer dependent on a single global text vector. It can detect the performance of architectural renderings in different dimensions. By calculating the cosine similarity between the image feature vector and each sub-vector, and performing weighted summation according to preset domain knowledge weights, while dividing by the image feature magnitude for normalization, a stable alignment feature that reflects the degree of multi-granular semantic matching and is not affected by low-order statistical characteristics of the image can be obtained, thus improving the accuracy of ensuring that the recommended architectural renderings meet the user's intentions.

[0069] In some embodiments of the present invention, such as Figure 6 As shown, step S204 includes: S601. Obtain the cumulative number of various user interaction behaviors that occur on the platform when the architectural rendering is displayed.

[0070] It should be noted that the platform maintains an interaction behavior log table for each architectural rendering on the backend. When a user performs actions such as liking, saving, browsing, copying generation parameters, or remixing, the server records the behavior type, occurrence time, and user ID in real time. For the architectural rendering to be evaluated, the cumulative number of various interaction behaviors since its publication is retrieved from the database or cache, for example, N. remix (Number of secondary generation) N copy (Number of copies), N save (Number of collections), N like (Number of likes), N view (Number of times viewed). The cumulative number of views reflects the objective amount of user feedback, at different levels, received by the architectural renderings within the community.

[0071] S602. Calculate the total number of interactions based on the preset weights of different types of user interaction behaviors and their cumulative number of interactions.

[0072] It should be noted that, within the open-source knowledge context of AI-generated architecture, different interactive behaviors are assigned differentiated weights to highlight image interactions with underlying engineering inspiration value. The behavior weights are ranked as follows: Remix > Copy Parameters > Save > Like > View. Specifically, Remix and Copy Parameters represent the highest level of professional recognition and are therefore assigned greater weights: β remix =5.0, β copy =3.0. Superficial interactions such as favorites and likes have a lower weight: β save =1.5, β like =1.0. Browsing on a paused period, as the most basic attention behavior, has a weight of only β. view =0.1. The total interaction count is obtained by multiplying the cumulative number of each behavior by its corresponding weight and then summing the results. RawScore=β remix ·N remix +β copy ·N copy +β save ·N save +β like ·N like +β view ·N view .

[0073] S603. Calculate the exponential decay factor based on the release time of the architectural rendering and the current time.

[0074] It should be noted that, to prevent highly-rated images from dominating the top of the recommendation list for an extended period (Matthew effect), a Newtonian cooling decay model is introduced to penalize popularity over time. Let the publication time of the architectural rendering be... The current calculation time is Time difference Δt = - (Usually in days or hours). If a half-life constant λ is preset (e.g., λ = 0.1 / day), then the exponential decay factor is... The exponential decay factor decreases exponentially with increasing time difference, causing the popularity of historical architectural renderings to naturally decline. When Δt=0, the exponential decay factor is 1; when Δt equals the duration corresponding to the half-life, the exponential decay factor is approximately 0.5; when Δt is much greater than the half-life, the exponential decay factor approaches 0, thus achieving a natural decline in popularity.

[0075] S604. The deep interaction heat characteristic is calculated based on the total interaction amount and the exponential decay factor.

[0076] It should be noted that the calculated total interaction is multiplied by the exponential decay factor to obtain the decayed total interaction: DecayedScore = RawScore. To map the heat value to the same order of magnitude as other multimodal features, typically within the (0, 1) interval, and to suppress extremely high scores, the decayed total interaction value is then input into the tanh function for nonlinear transformation to obtain the deep interaction heat feature. ; The tanh function compresses any real number into the interval (-1, 1), with the output value closer to 1 indicating higher popularity. To ensure non-negativity, the interval (0, 1) can also be used.

[0077] In this embodiment, by assigning significantly higher weight than likes and favorites to deep professional interactive behaviors such as secondary generation and parameter copying, the engineering reference value of architectural renderings at the open-source knowledge level can be accurately captured. This ensures that images with rigorous structure and exquisite parameter configurations, but not the most visually stunning, receive the attention they deserve, encouraging designers to share high-quality, reusable generated assets. Furthermore, by introducing an exponential decay factor based on Newton's law of cooling, the image's popularity naturally declines over time, effectively preventing the solidification of traffic for early highly-rated images and ensuring that newly released high-quality works receive equal exposure, thus maintaining the metabolism of the community ecosystem.

[0078] In some embodiments of the present invention, such as Figure 7 As shown, step S102 includes: S701. Obtain the text feature vector of the search text entered by the user in the current browsing session.

[0079] It should be noted that when a user enters the architectural rendering sharing community and actively inputs search text (e.g., "layout of the facade of a modern wooden pavilion"), the search text is fed into a lightweight pre-trained language model (such as BERT or DistilBERT) to extract the sentence vector corresponding to its [CLS] token, which is then used as the text feature vector. Text feature vectors typically have a dimension of 128 or 256, and can represent the design concerns that the user is currently expressing, such as materials (wood structure, concrete), building type (pavilion), perspective (facade), and level of focus (layout).

[0080] S702. Obtain the semantic tag distribution vector of the architectural renderings that the user clicked within a preset time window.

[0081] It should be noted that this method records every click a user makes to zoom in on architectural renderings during the current browsing session. A short time window (e.g., 5 minutes) is preset, and the semantic tags of all architectural renderings clicked by the user within this window are counted. These semantic tags are automatically annotated by the model during upload or generation, or manually annotated by the user, such as "timber structure," "concrete," "curved roof," and "detailed node." The frequency of each semantic tag within the count window is used to generate a semantic tag distribution vector. If there are M predefined labels, the vector is an M-dimensional sparse or dense representation, where each component represents the relative frequency of the corresponding label in the click history.

[0082] S703. Obtain the user's registration attribute vector; the registration attribute vector is a vector obtained by encoding the user's occupation category.

[0083] It should be noted that user occupational category information (e.g., "architectural designer," "structural engineer," "interior designer," "landscape planner," "student," etc.) is collected during user registration. This occupational category is then converted into a registration attribute vector using one-hot encoding. Register attribute vector The dimension equals the total number of preset occupational categories. For users who have not provided occupational information, it can be set to the "Unknown" category or use an all-zero vector. (Registered attribute vector) It expresses the inherent preferences of user roles; for example, a structural engineer may be more concerned with practicality, while a concept designer may be more concerned with aesthetic expression.

[0084] S704. The query intent vector is obtained by fusing and reducing the dimensions of the text feature vector, the semantic tag distribution vector, and the registration attribute vector.

[0085] It should be noted that: text feature vectors Semantic label distribution vector and registered attribute vector The concatenation vector is formed by concatenating the features along the feature dimensions to create a high-dimensional initial feature representation. Then, the concatenated vector is input into a multilayer perceptron (MLP, which contains one linear mapping layer and one ReLU activation function layer), which will generate a high-dimensional concatenated vector. Compress to a low-dimensional space and output the intent query vector. The intent query vector integrates information from three aspects: user active retrieval, passive click behavior, and long-term career preferences. It represents the user's subconscious emphasis on images across four evaluation dimensions during the current session.

[0086] S705. Construct four embedding vectors that correspond one-to-one with the architectural visual rationality feature, architectural aesthetic quality feature, text-image alignment feature, and deep interaction heat feature, and stack the four embedding vectors to form the dimension key matrix.

[0087] It should be noted that during the network initialization phase, a learnable embedding vector of length d (the same as the intent vector dimension) is randomly initialized for each of the four evaluation dimensions (architectural visual rationality features, architectural aesthetic quality features, text-image alignment features, and deep interaction popularity features), denoted as follows: These four embedding vectors are continuously updated as the model trains to learn the semantic relevance of each dimension to the user's intent. Then, these four embedding vectors are vertically stacked to form a 4×d matrix, resulting in the dimension key matrix K. Each row of the dimension key matrix K corresponds to a learnable representation of a feature dimension.

[0088] S706. Using the query intent vector as the query term for cross-attention and the dimension key matrix as the key term for cross-attention, calculate the dot product between the query intent vector and the dimension key matrix, scale the dot product result and input it into the normalized exponential function, and output the weight coefficients corresponding to each feature dimension.

[0089] It should be noted that performing a dot product between the intent query vector Q (dimension d) and the dimension key matrix K (4×d) yields the original attention score vector with dimension 4. Since the dimension d of the intent query vector Q may be large, to prevent the dot product result from falling into the saturation region of the Softmax function, the original attention score vector... Divide by The vector is then scaled. The scaled vector is then input into the Softmax function for normalization, yielding the weight coefficients for each feature dimension. ; in, These correspond to weight coefficients for architectural visual rationality, architectural aesthetic quality, text-image alignment, and deep interaction popularity features, respectively. The Softmax function ensures that the sum of the four dynamic weights in the output is 1, i.e. To illustrate the effect of dynamic weighting coefficients, the following example is provided: Scenario A (Finding Inspiration): When When a text contains a large number of "conceptual, abstract, and futuristic" features, the attention mechanism responds strongly. The network automatically suppresses structural rationality weights, and the output is as follows: We will proactively push visually striking and disruptive concept images.

[0090] Scenario B (Finding a practical reference): When When features such as "arrangement logic, nodes, and structure" are included, the network experiences a weight mutation, which... To maximize the response, the output is as follows: Images with extremely high rigor and reasonable accuracy will be given significantly increased weight.

[0091] In this embodiment, by fusing the semantic tag distribution of the user's actively retrieved text, short-term click behavior, and registered occupational attributes, and then using a multilayer perceptron for dimensionality reduction to obtain a comprehensive intent query vector, the system achieves accurate capture of changes in the user's intent during the current browsing session. Furthermore, by setting learnable embedding vectors for each of the four architectural professional evaluation dimensions and stacking them into a dimension key matrix, the cross-attention mechanism can automatically learn the semantic matching relationship between intent and dimensions. Then, by normalizing and calculating the weight coefficients corresponding to each feature dimension, the system significantly improves the relevance of the recommendation results to the user's real-time needs.

[0092] In some embodiments of the present invention, after generating the final recommendation list, the method further includes: When a user performs parameter reuse or image regeneration operations on any architectural rendering in the final recommendation list, the interaction popularity feature of the architectural rendering is updated according to the operation type.

[0093] It should be noted that after pushing the final recommendation list to users, subsequent user interactions with each architectural rendering in the list are continuously monitored. When a user performs a "parameter reuse" operation (e.g., clicking the "copy prompt and generate parameters" button) or an "image regeneration" operation (e.g., clicking the "secondary generation / Remix" button to modify the prompt or parameters in the cloud based on the current image and re-render), the type of operation and the corresponding image identifier are immediately captured. For the architectural rendering that has been manipulated, the count of the corresponding type of operation is increased by 1 in its original cumulative number of interactions. Specifically, if the user performs a parameter reuse operation, then N copy Increment by 1; if the operation is image regeneration, then Nremix Increment by 1. Then, recalculate the deep interaction heat features of the image according to the aforementioned calculation method (i.e., multiply the total weighted interaction amount by an exponential decay factor and then map it using a hyperbolic tangent function), and update the storage. Write the updated deep interaction heat features back to the database, and refresh the multimodal features of the architectural rendering in the cache.

[0094] In this embodiment, by capturing and accumulating the number of user-performed parameter reuse or image regeneration operations in real time, derived interactive behaviors are given a much higher weight in terms of popularity increment than ordinary interactions (likes, favorites). This allows architectural renderings with rigorous keyword structure and exquisite parameter configuration, but not the most visually stunning, to receive a popularity boost commensurate with their professional contribution. Furthermore, by recalculating the popularity characteristics using the Newtonian cooling index decay factor, the timeliness of the popularity value is ensured, preventing early highly-rated architectural renderings from monopolizing the recommendation list for a long time. This guarantees that newly released high-quality architectural renderings receive equal exposure opportunities, thereby improving the quality of recommended architectural renderings.

[0095] In some embodiments of the present invention, such as Figure 8 As shown, step S104 includes: S801. Calculate the similarity value between the potential representation vectors of every two architectural renderings in the candidate set.

[0096] It should be noted that after the candidate set C (e.g., containing 200 images) is generated, the latent representation vector Z (a flattened one-dimensional vector) of each image is read from a vector database (such as Milvus). This latent representation vector is obtained by directly calling the output of the variational autoencoder (VAE) encoder in the image generation model (such as Stable Diffusion): truncating the latent space compressed feature tensor of the encoder output. Its shape is usually (like ), then flatten it into a one-dimensional vector This vector contains essential information about the image's underlying manifold, including its compositional skeleton, spatial layout, and core semantics. It can keenly identify highly homogeneous derived images that "apply the same structural skeleton but only replace the surface material." For each pair of images in the candidate set... Calculate the cosine similarity of their latent representation vectors as the similarity value: .

[0097] The similarity value ranges from [-1, 1]. A similarity value closer to 1 indicates that the two images are more similar in underlying manifolds such as compositional skeleton, spatial layout, and core semantics (e.g., derivative images generated by fine-tuning the seed with the same cue word, or images with different cue words but following the same compositional template). All images in the candidate set are calculated. After obtaining the similarity value, it is cached for use in subsequent iterations.

[0098] S802. Using a greedy iterative algorithm, select one of the architectural renderings from the candidate set and add it to the final recommendation list; wherein, the selected architectural rendering satisfies the condition of maximizing the objective function value, and the objective function is calculated based on the initial recommendation score of the architectural rendering and the similarity value of the latent representation vectors between the architectural renderings.

[0099] It should be noted that: First, for each architectural rendering in candidate set C... Its initial recommendation score is highly correlated with the user's immediate intent. Calculate using the following formula: ; After calculating the initial score, to break the information cocoon and aesthetic homogenization, the Maximum Marginal Relevance (MMR) algorithm is introduced for list reconstruction. The final recommendation list L is initialized to empty, and the current candidate set is the candidate set C. Diversity adjustment parameters are then set. (For example =0.5), this parameter is used to balance recommendation quality and diversity: The larger the value, the stronger the diversity penalty. In each iteration, for each architectural rendering in the candidate set C... Calculate the corresponding maximum marginal correlation objective function value: ; in, For the architectural rendering The initial recommendation score (which has been calculated using dynamic weighting). express The similarity to the most similar image selected in the final recommendation list L (if L is empty, the second item is set to 0). Then, select the image that is most similar to the selected image in the final recommendation list L. The architectural rendering with the highest value is removed from the candidate set C and appended to the end of the final recommendation list L.

[0100] S803. Repeat the iterative selection until the number of selections reaches the preset recommendation number to obtain the final recommendation list.

[0101] It should be noted that the selection process in step S802 is repeated, with one selected architectural rendering added to the final recommendation list L in each round. The iteration stops when the number of images in L reaches the system's preset recommendation limit (e.g., 20 images) or the candidate set is empty. At this point, the images in L are arranged in the order of selection (i.e., the order of greedy selection is the recommendation order), and this list is pushed to the user's front end as the final architectural rendering recommendation list.

[0102] In this embodiment, by directly reusing the latent representation vectors of the generative model to calculate the essential similarity between candidate images, it is possible to accurately identify derivative images with isomorphic compositional skeletons and spatial layouts but different surface textures, providing a more reliable metric basis for diversity penalty than pixels or CNN features. Furthermore, by employing the maximum edge relevance algorithm combined with greedy iteration, images that maximize the objective function (balancing initial recommendation scores and difference penalties) are selected at each step, achieving an explicit balance between recommendation quality and diversity: ensuring that high-ranking, high-quality images are included in the list while forcibly inserting long-tail works with styles distinct from the selected images. This iteration is repeated until a preset number of recommendations is reached, resulting in a final recommendation list that presents a rich and diverse appearance, similar to a multi-faceted architectural exhibition, effectively suppressing aesthetic homogenization and information cocoons caused by algorithmic traffic-driven approaches. This mechanism requires no additional training of a similarity network; reusing existing vectors efficiently completes the reordering, significantly improving the user's browsing experience and the efficiency of obtaining design inspiration while maintaining system throughput performance.

[0103] In all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. Additionally, when embodiments of this application require access to sensitive personal information, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0104] To better implement the method for recommending architectural renderings in the embodiments of the present invention, based on the method for recommending architectural renderings, correspondingly, as follows: Figure 10 As shown, the present invention also provides an electronic device 1000. The electronic device 1000 includes a processor 1001, a memory 1002, and a display 1003. Figure 10 Only some components of the electronic device 1000 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0105] In some embodiments, processor 1001 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 1002 or process data, such as the recommended method for architectural renderings in this invention.

[0106] In some embodiments, processor 1001 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 1001 may be local or remote. In some embodiments, processor 1001 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, etc., or any combination thereof.

[0107] In some embodiments, memory 1002 may be an internal storage unit of electronic device 1000, such as a hard disk or memory of electronic device 1000. In other embodiments, memory 1002 may also be an external storage device of electronic device 1000, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 1000.

[0108] Furthermore, the memory 1002 may include both internal storage units of the electronic device 1000 and external storage devices. The memory 1002 is used to store application software and various types of data installed on the electronic device 1000.

[0109] In some embodiments, display 1003 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 1003 is used to display information from electronic device 1000 and to display a visual user interface. Components 1001-1003 of electronic device 1000 communicate with each other via a system bus.

[0110] In one embodiment, when the processor 1001 executes the recommendation program for architectural renderings stored in the memory 1002, the following steps can be implemented: Extract multimodal features and latent representation vectors from each architectural rendering; Obtain the user's query intent vector, and match the query intent vector with the dimension key matrix corresponding to the multimodal features to generate weight coefficients corresponding to each feature dimension; An initial recommendation score is calculated based on the multimodal features and the weighting coefficients, and a candidate set is obtained by filtering the architectural renderings based on the initial recommendation score; the candidate set includes multiple architectural renderings whose initial recommendation scores meet preset conditions. Based on the similarity value of the latent representation vectors among the architectural rendering images in the candidate set, the architectural rendering images are selected and sorted from the candidate set to generate a final recommendation list.

[0111] It should be understood that when the processor 1001 executes the recommended program for the architectural renderings in the memory 1002, in addition to the functions mentioned above, it can also perform other functions, as detailed in the description of the corresponding method embodiments above.

[0112] Furthermore, the embodiments of the present invention do not specifically limit the type of the electronic device 1000 mentioned. The electronic device 1000 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, the electronic device 1000 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0113] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0114] The above provides a detailed description of the recommended method and electronic device for generating architectural renderings provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for recommending architectural renderings, characterized in that, include: Extract multimodal features and latent representation vectors from each architectural rendering; Obtain the user's query intent vector, and match the query intent vector with the dimension key matrix corresponding to the multimodal features to generate weight coefficients corresponding to each feature dimension; An initial recommendation score is calculated based on the multimodal features and the weighting coefficients, and a candidate set is obtained by filtering the architectural renderings based on the initial recommendation score; the candidate set includes multiple architectural renderings whose initial recommendation scores meet preset conditions. Based on the similarity value of the latent representation vectors among the architectural rendering images in the candidate set, the architectural rendering images are selected and sorted from the candidate set to generate a final recommendation list.

2. The method according to claim 1, characterized in that, The extraction of multimodal features and latent representation vectors for each architectural rendering includes: A deep geometric consistency analysis was performed on the architectural renderings to obtain the visual rationality characteristics of the building. A multi-dimensional aesthetic decoupling evaluation was performed on the architectural renderings to obtain the architectural aesthetic quality characteristics; Cross-modal semantic matching was performed on the architectural renderings and their accompanying text prompts to obtain text-image alignment features; The deep interaction heat characteristics are obtained based on the cumulative number of various user interaction behaviors of the architectural renderings. A latent spatial encoder based on an image generation model extracts the latent representation vector of the architectural rendering.

3. The method according to claim 1, characterized in that, The in-depth geometric consistency analysis of the architectural renderings yields the architectural visual rationality characteristics, including: Extract the depth information from the architectural renderings to generate a relative depth map; Extract the geometric outline information from the architectural renderings to generate a structural wireframe diagram; The architectural rendering, the relative depth map, and the structural wireframe are fused together to form a multi-channel feature tensor. The multi-channel feature tensor is input into a pre-trained classification network, which outputs a probability value representing whether the architectural rendering satisfies architectural geometric consistency. The probability value is used as the visual rationality feature of the building.

4. The method according to claim 2, characterized in that, The multi-dimensional aesthetic decoupling evaluation of the architectural renderings yields architectural aesthetic quality characteristics, including: Extract the first sub-feature related to light and shadow layers, the second sub-feature related to material representation, and the third sub-feature related to spatial composition from the architectural rendering; Based on the first sub-feature, the second sub-feature, and the third sub-feature, a comprehensive aesthetic score is calculated according to preset weight values. The comprehensive aesthetic score is normalized, and the normalization result is used as the architectural aesthetic quality characteristic.

5. The method according to claim 2, characterized in that, The step of performing cross-modal semantic matching on the architectural renderings and their accompanying text prompts to obtain text-image alignment features includes: Extract the image feature vector of the architectural rendering; The prompt text is semantically decomposed to obtain multiple semantic feature vectors, including at least the building subject, architectural style, and environmental context; Calculate the cosine similarity between the image feature vector and each of the semantic feature vectors respectively; The text image alignment feature is calculated based on the cosine similarity and the magnitude of the image feature vector.

6. The method according to claim 2, characterized in that, The method of obtaining deep interaction heat characteristics based on the cumulative number of various user interaction behaviors based on the architectural renderings includes: Obtain the cumulative number of various user interactions that occur with the architectural renderings on the platform; The total number of interactions is calculated based on the preset weights of different types of user interaction behaviors and their cumulative number of interactions. The exponential decay factor is calculated based on the release time of the architectural rendering and the current time. The deep interaction heat characteristic is calculated based on the total interaction volume and the exponential decay factor.

7. The method according to claim 2, characterized in that, The step of obtaining the user's query intent vector and matching the query intent vector with the dimension key matrix corresponding to the multimodal features to generate weight coefficients corresponding to each feature dimension includes: Obtain the text feature vector of the search text entered by the user in the current browsing session; Obtain the semantic tag distribution vector of the architectural renderings that the user clicked within a preset time window; Obtain the user's registration attribute vector; the registration attribute vector is a vector obtained by encoding the user's occupation category; The query intent vector is obtained by fusing and reducing the dimensions of the text feature vector, the semantic tag distribution vector, and the registration attribute vector. Four embedding vectors are constructed, each corresponding to one of the architectural visual rationality features, architectural aesthetic quality features, text-image alignment features, and deep interaction heat features. The four embedding vectors are then stacked to form the dimension key matrix. The query intent vector is used as the query term for cross-attention, and the dimension key matrix is ​​used as the key term for cross-attention. The dot product of the query intent vector and the dimension key matrix is ​​calculated. The dot product result is scaled and then input into a normalized exponential function to output the weight coefficients corresponding to each feature dimension.

8. The method according to claim 2, characterized in that, After generating the final recommendation list, the process also includes: When a user performs parameter reuse or image regeneration operations on any architectural rendering in the final recommendation list, the interaction popularity feature of the architectural rendering is updated according to the operation type.

9. The method according to any one of claims 1 to 8, characterized in that, The step of selecting and sorting the architectural renderings from the candidate set based on the similarity value of the latent representation vectors between the architectural renderings in the candidate set, and generating the final recommendation list, includes: Calculate the similarity value between the potential representation vectors of every two architectural renderings in the candidate set; A greedy iterative algorithm is used to select one of the architectural renderings from the candidate set and add it to the final recommendation list; wherein the selected architectural rendering satisfies the condition of maximizing the objective function value, and the objective function is calculated based on the initial recommendation score of the architectural rendering and the similarity value of the latent representation vectors between the architectural renderings; The selection process is repeated iteratively until the number of selections reaches the preset recommendation number, thus obtaining the final recommendation list.

10. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the recommended method for generating architectural renderings according to any one of claims 1 to 9.

Images