Three-dimensional modeling method and system based on spatial hand-drawing and AI intelligent completion

Abstract

This application relates to the technical field of 3D modeling methods and systems based on spatial hand-drawing and AI intelligent completion. Specifically, it discloses a 3D modeling method and system based on spatial hand-drawing and AI intelligent completion, aiming to solve the problems of complex operation, high learning threshold, and limited creative expression in traditional 3D modeling. The method includes: freely drawing spatial curves on a 3D canvas and then using AI to optimize smoothness; selecting a closed wireframe and triggering AI to intelligently repair gaps and generate quadrilateral mesh surfaces, providing multiple topology options for the user; based on geometric features and design context, AI automatically matches and generates suitable materials; finally, a lightweight neural rendering engine is called to quickly generate a preview image. By adopting the above technical solution, this application can achieve high-quality 3D modeling with minimal interaction, significantly reducing the usage threshold, improving the freedom of creative expression and modeling efficiency, while ensuring the topological regularity and rendering realism of the output model.

Description

Technical Field

[0001] This application belongs to the field of computer graphics and artificial intelligence, specifically involving a 3D modeling method and system based on spatial hand-drawing and AI intelligent completion. Background Technology

[0002] With the widespread application of 3D modeling technology in industrial design, film and animation, game development, and other fields, users have placed higher demands on the intuitiveness, flexibility, and intelligence of modeling tools. While traditional 3D modeling software such as Maya, 3ds Max, Blender, and SolidWorks are powerful, their interactive logic heavily relies on menu commands, parameter panels, and precise coordinate input. This results in a severe disconnect between the operational flow and natural hand-drawn expression, making it difficult to meet the needs of rapid conceptualization in the early stages. Especially in the concept design phase, designers often rely on freehand sketching to capture inspiration, but existing tools lack direct support for unstructured line drawings in 3D space, making the conversion process from 2D sketches to 3D models lengthy and technically challenging.

[0003] Among these approaches, spatial hand-drawn 3D modeling attempts to bridge the gap between creative expression and digital realization. The core of this direction lies in allowing users to draw free-form curves in a 3D environment, analogous to drawing with pen and paper, with the system automatically inferring the geometric intent and generating surfaces and structures that conform to the design semantics. However, current technologies still have significant limitations: most solutions only support sketch stretching under 2D projection, unable to handle truly 3D spatial line art; even some systems that support spatial drawing lack intelligent repair capabilities for issues such as line art jitter, gap breaks, and topological blur, leading to subsequent surface generation failures or low quality.

[0004] Existing technologies also face multiple challenges in areas such as surface reconstruction, material assignment, and rendering preview. On the one hand, Boolean operations and mesh merging in traditional modeling workflows are destructive and difficult to reverse once executed, limiting the flexibility of design iteration. On the other hand, the generation of high-quality quadrilateral meshes and the configuration of reasonable material parameters are highly dependent on user experience, making it difficult for non-professional users to complete independently. In addition, existing AI-assisted tools mostly focus on single tasks (such as optimizing curves or generating materials only), lacking a collaborative intelligent mechanism that runs through the entire "hand-drawing—surface creation—material assignment—rendering" chain, and cannot achieve context-aware completion from local wireframes to the overall model. Therefore, there is an urgent need for a new modeling method and system that supports freehand drawing in 3D space, has full-process AI intelligent completion capabilities, and maintains non-destructive editing characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a 3D modeling method and system based on spatial hand-drawing and AI intelligent completion, which can effectively solve the problems in the background technology mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A 3D modeling method based on spatial hand-drawing and AI intelligent completion includes the following specific steps: Step 1: Spatial hand-drawn line drawing input and intelligent optimization: In response to the user's free spatial hand-drawing operation using a stylus on the 3D canvas, an original 3D spatial curve is generated; the user's hand-drawing experience is analogous to 2D drawing, and can be drawn directly on any plane in 3D space without switching modes or inputting coordinates; the original curve is smoothed and optimized by a pre-trained AI curve optimization model, recognizing and preserving its original design features, while reducing noise caused by hand-drawing jitter, generating an optimized spatial line drawing; Step 2: Selecting closed wireframes and triggering AI surface generation: In response to the user selecting at least one closed wireframe in the 3D canvas using the selection tool... The process involves merging wireframes and the user clicking the "Generate Surface" command in the AI ​​assistant panel. The system then obtains the geometric information of the selected wireframe. Users can choose to use the AI ​​to generate surfaces for the entire model via the overall generation panel on the right side of the interface, or directly select or outline a local wireframe in the 3D viewport. Step 3: AI generates surfaces. The AI ​​surface generation model analyzes the topology of the selected wireframe, automatically identifies minute gaps or incompletely closed regions between wireframes, and performs intelligent repair based on context. Based on the identified and repaired wireframes, a 3D surface composed of quadrilateral meshes that conforms to the wireframe boundaries is generated, and the number of control points on adjacent boundaries is automatically matched to ensure a clean topology. When multiple possible surfaces exist... When generating surfaces, several alternative topology schemes are previewed in thumbnail form in the overall generation panel on the right for the user to select and confirm. In response to the user's selection command for a particular alternative scheme, the selected 3D surface is added to the current layer. Step 4: Intelligent material addition: In response to the user's command to assign materials to at least a portion of the surfaces in the model, the AI ​​material generation model is invoked to perform local or overall selection. The user can select a single surface in the 3D viewport by point selection, or select multiple surfaces by frame selection for local material editing; alternatively, the user can select the entire model in the layer panel on the right for overall material generation and assignment. The AI ​​model analyzes the geometric features and design context of the selected area, automatically matching and generating initial material properties that meet the requirements. The right panel provides a material library with multiple categories. Users can select different categories, and the AI ​​will automatically adjust the corresponding parameters to achieve intelligent replacement. For the same material category, multiple material variations are provided for users to preview and select. Step 5: AI generates and renders the image: In response to the user's selected specific viewpoint and the "Generate Render Preview" command, the AI ​​rendering engine is invoked to determine the viewpoint. Users can adjust the viewing angle through the viewpoint switching tool or select from preset viewpoints. The AI ​​model quickly calculates lighting, reflection, refraction, shadow, and other effects based on the geometric information of the current model, the assigned material properties, and the default or adjustable lighting conditions. A preview-level rendered image is generated and displayed in the interface to help users evaluate the final design effect.

[0007] Preferably, the AI ​​curve optimization model in step 1 adopts a sequence-to-sequence neural network architecture based on an attention mechanism. During the training phase, the model uses a large dataset of hand-drawn curves and corresponding ideal curves labeled by professional designers for supervised learning. This model can distinguish between key inflection points in the design intent and hand-drawn jitter noise, retain the style features of the original sketch, and reduce the standard deviation of the curve curvature change to below a predetermined proportion of the original hand-drawn curve.

[0008] Preferably, the logic for determining closed wireframes in step 2 not only relies on geometric closure but also combines topological connectivity and visual closure perception. The system determines whether a closure is constituted by calculating the Euclidean distance threshold between the endpoints of the wireframe (set as a predetermined multiple of the average stroke width in the current view) and introduces a deep learning-based visual closure prediction module to intelligently identify wireframes that the human eye subjectively believes should be closed but are not fully connected, thereby improving the fault tolerance of user interaction.

[0009] Preferably, the AI ​​surface generation model in step 3 adopts a hybrid architecture combining graph neural networks and implicit fields. The model first converts the input wireframe into a weighted graph structure, where nodes represent key control points and edges represent connection relationships. Then, it extracts global topological semantics through multi-layer graph convolution and drives the implicit neural radiation field to generate continuous surfaces. The generated quadrilateral mesh satisfies boundary consistency constraints, the number error of vertices on the shared boundary between adjacent surface patches does not exceed a preset threshold, and the internal triangulation satisfies the Delaunay criterion.

[0010] Preferably, the several alternative topology schemes provided in step 3 are generated based on different surface extension strategies, including minimum curvature extension, symmetry-priority extension, and volume-filling-priority extension; each strategy corresponds to a different combination of loss function weights, and the system dynamically adjusts the activation probability of each strategy according to the number of symmetry axes of the wireframe, the bounding box ratio, and historical user selection preferences, to ensure that the alternative schemes are both diverse and in line with the design context.

[0011] Preferably, the AI ​​material generation model in step 4 incorporates a multi-scale geometric feature encoder, which extracts curvature, normal distribution entropy, concavity / convexity, and orientation consistency features at three scales: local (radius is a predetermined proportion of the model's maximum size), mesoscale (predetermined proportion), and global (predetermined proportion). These features are input into a conditional generative adversarial network, which outputs a set of material parameters that conform to physical reality, including diffuse reflection coefficient, specular reflection intensity, roughness, metallicity, and subsurface scattering coefficient.

[0012] Preferably, the material library in step 4 is divided into multiple categories according to function, and each category contains multiple subcategories; for example, the metal category includes aluminum, copper, stainless steel, brushed gold, etc.; the organic material category includes wood, leather, fabric, rubber, etc.; each subcategory is associated with a set of learnable material prototype vectors. When the user selects a subcategory, the AI ​​model fine-tunes the prototype vectors based on the current geometric features to generate a personalized material instance that fits the current surface.

[0013] Preferably, the AI ​​rendering engine in step 5 adopts a lightweight neural radiation field approximation model. This model is pre-trained offline on a dataset containing a large number of high-quality rendering images to learn the mapping relationship from geometry-material combination to final pixel color. During online rendering, a preview image of a predetermined resolution can be generated within a predetermined time period. Its lighting effect and the physically based path tracing results reach a preset threshold in terms of structural similarity index.

[0014] Preferably, the method supports a non-destructive editing workflow, where all AI-generated surfaces, materials, and rendering results are stored in the scene graph as parametric nodes. Users can return to any historical node at any time to modify input conditions (such as adjusting the original hand-drawn line art or changing the material category), and all subsequent dependent nodes will be automatically recalculated, ensuring that the entire process is traceable, iterative, and reversible.

[0015] Preferably, the system integrates a real-time feedback mechanism in the 3D canvas. When the user pauses for more than a predetermined time period during the drawing process, the system preloads the AI ​​optimization and surface generation module in the background to predict the user's next operation intention. If it detects that the user is drawing a wireframe that is close to a closed shape, it automatically highlights the potential closed area and displays a miniature "Generate Surface" prompt icon to achieve intelligent guidance with zero delay.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, an extremely simple learning curve. By allowing users to directly hand-draw line art in 3D space, with an interaction method similar to 2D drawing, and without the need to learn complex commands, the learning threshold for 3D modeling is greatly reduced, allowing anyone to get started quickly.

[0017] Secondly, it offers freedom of creative expression. Users can freely draw any curve in three-dimensional space, without being limited by a two-dimensional view. AI optimizes in real time to ensure the smoothness of hand-drawn art and the integrity of creative expression.

[0018] Third, intelligent completion and optimization. AI automatically completes tedious tasks such as curve optimization, gap repair, surface generation, topology optimization, material matching, and rendering calculations, significantly improving the conversion efficiency from spatial hand-drawn sketches to high-quality 3D models.

[0019] Fourth, flexible workflow. All AI generation steps support local selection and adjustment, allowing users to make fine-tuning modifications to any part of the model at any time; combined with non-destructive editing features, it gives users a great deal of creative freedom and flexibility to rework.

[0020] Fifth, ensure output quality. By generating industry-standard quadrilateral meshes and reasonable material parameters using AI, the compatibility and quality of the model are guaranteed in subsequent subdivision sculpting, UV unwrapping, animation rigging, rendering, and other stages.

[0021] Sixth, guided creation experience. When users are confused, AI provides multiple alternatives and material variations, presented in an intuitive preview format, to guide users in making modeling decisions and reduce the difficulty of selection. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical solution architecture of the 3D modeling method and system based on spatial hand-drawing and AI intelligent completion proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the AI ​​intelligent completion core process (covering curve optimization, surface generation, material matching and rendering) in this invention; Figure 3 This is a flowchart illustrating the logical flow of spatial hand-drawn line drawing input, closed frame recognition, and AI-triggered surface generation in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between AI surface generation and multi-scheme topology recommendation in this invention; Figure 5 This is a schematic diagram comparing the core principles of this invention with existing technologies in terms of user interaction fault tolerance and modeling efficiency; Detailed Implementation Example 1

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0024] In the aforementioned 3D modeling method based on spatial hand-drawing and AI intelligent completion, step S1, namely spatial hand-drawn line drawing input and intelligent optimization, is implemented as follows: After system initialization, a global right-handed coordinate system is established in the 3D canvas, and the stylus input monitoring module is activated. This module captures the spatial position data stream of the stylus in real time through the underlying API of the operating system, with a sampling frequency set to 200Hz and a position accuracy better than 0.1 mm. When the user starts drawing at any position in 3D space, the system immediately records the first 3D coordinate point. And continue to track subsequent trajectory point sequences { }, forming the original three-dimensional space curve The curve is stored in a memory buffer in parametric form, and its data structure is defined as containing... An ordered list of vertices, each vertex carrying 3D coordinates, a timestamp, and a pressure value.

[0025] During the user's drawing process, the system synchronously runs a plane fitting algorithm to dynamically determine the current drawing plane. Specifically, for continuous... trajectory points The system uses principal component analysis (PCA) to calculate its best-fit plane. The plane is formed by the normal vector N and a point on the plane. This definition constrains the projection of subsequent points, ensuring users have an intuitive experience similar to 2D drawing. Even if a slight tremor of the user's wrist causes the pen tip to briefly deviate from the plane, the system can maintain a visually planar drawing feel through projection correction, without requiring the user to manually switch views or specify a work plane.

[0026] When a user lifts the pen or pauses drawing for more than 300 milliseconds, the system determines that a complete hand-drawing operation has ended and triggers the AI ​​curve optimization model. This model employs a sequence-to-sequence approach based on an attention mechanism. The neural network architecture consists of bidirectional long short-term memory (Bi-LSTM) networks for both the encoder and decoder, with a multi-head self-attention layer embedded in between. During offline training, the model uses a dataset containing 500,000 pairs of hand-drawn and ideal curves for supervised learning. Each pair of samples is annotated by a professional industrial designer: the original hand-drawn curves retain all jitter and irregular features, while the corresponding ideal curves are manually smoothed to retain key design inflection points but eliminate high-frequency noise.

[0027] During the reasoning phase, the original curve The input sequence is normalized to a unit cube [0,1]³ and fed into the encoder. An attention mechanism dynamically allocates weights, focusing on regions of abrupt curvature changes (i.e., key points of potential design intent) while suppressing high-frequency jitter. The decoder outputs an optimized curve. The number of vertices can be dynamically adjusted, typically ranging from 30% to 80% of the original number of vertices, to balance detail preservation and data simplification. The optimization process satisfies the following mathematical constraint: the standard deviation of the curvature change of the optimized curve. Not exceeding the standard deviation of the curvature change of the original curve 40% of, that is:

[0028] Among them, curvature The calculation is performed at discrete points using a three-point circle fitting method. The optimized spatial line drawing then replaces the original curve and is highlighted in the 3D canvas. Simultaneously, its geometric data is encapsulated as an immutable object and stored in the scene graph's history nodes, providing a foundation for subsequent non-destructive editing.

[0029] In the above method, step S2, namely selecting the closed wireframe and triggering AI face generation, is specifically implemented as follows: The system maintains a dynamic wireframe set L, where each element Representing an independent spatial curve, it includes its vertex list, whether it is closed, and topological connectivity information. Users can perform point selection or bounding box selection in the 3D viewport using mouse or touch operations. Point selection is implemented using a ray casting algorithm: a ray is cast from the camera position to the clicked point on the screen, intersection tests are performed with all wireframes, and the wireframe ID corresponding to the nearest intersection point is returned. Bounding box selection defines a rectangular area in screen space and includes all wireframe projection points within this area into the selection set.

[0030] The system performs a closure test on the selected wireframe, which integrates three criteria: geometric closure, topological connectivity, and visual closure perception. First, the geometric closure is checked by examining the Euclidean distance between the first and last vertices of the wireframe. This distance threshold Set as the average stroke width in the current view 3 times, that is

[0031] .like If the initial condition is correct, it is initially determined to be geometrically closure. Secondly, topological connectivity is verified to ensure that all edges inside the wireframe are connected end-to-end to form a single loop, without branches or dangling endpoints. Finally, a visual closure prediction module based on a convolutional neural network is introduced. This module takes the screen projection image of the wireframe as input and outputs a "subjective closure probability". .like Even if it is not geometrically completely closed, the system will still mark it as "visually closed".

[0032] When a user clicks the "Generate Surface" command in the AI ​​assistant panel, the system collects all wireframes marked as closed, extracts their geometric information, including vertex coordinates, edge connectivity, bounding boxes, and normal vector distribution. This information is then serialized into... The format is then transmitted to the AI ​​surface generation module via an internal message bus. Simultaneously, the system supports two trigger modes: in local mode, only the wireframes explicitly selected by the user are processed; in global mode, the user selects "Generate surfaces for the entire model" through the global generation panel on the right, and the system automatically traverses the entire scene graph, identifies all closed wireframes, and submits processing requests in batches.

[0033] In the above method, step S3, namely AI surface generation, involves calling the AI ​​surface generation model to perform the following sub-steps: Sub-step S31: Intent Recognition and Gap Repair. The model first preprocesses the input wireframe set to construct a weighted graph. In this context, node V corresponds to the key control point of the wireframe (the vertices retained after simplification using the Douglas-Peucker algorithm), edge E represents the connection between vertices, and weight w... e The distance is determined by the ratio of the original curve length to the straight-line distance between two points. Subsequently, a graph neural network (GNN) performs multi-layer message passing on G, aggregating neighborhood information to identify potential gaps. For distances less than a threshold... For non-connected endpoint pairs (set as 0.5% of the diagonal length of the model bounding box), the model predicts the optimal connection path based on contextual semantics (such as local curvature consistency and symmetry) and inserts virtual vertices to complete intelligent patching. The patched wireframe ensures topological closure and geometric continuity.

[0034] Sub-step S32: Topology generation. The repaired wireframe is fed into the Implicit Neural Radiation Field (NeRF) submodule. This submodule consists of an 8-layer fully connected network, with spatial coordinates as input. The output is the signed distance function (SDF) value. Network weights are dynamically modulated by global topological semantic features extracted by the GNN to ensure the generated surface strictly fits the wireframe boundary. Subsequently, isosurfaces are extracted from the SDF field using the Marching Cubes algorithm, and quadrilateral-dominated meshing is performed. During this process, the system enforces boundary consistency constraints: for adjacent surface patches sharing a boundary, the number of boundary vertices is adjusted to be the same through interpolation or merging operations, with an error not exceeding two vertices. The final generated mesh, after internal triangulation, satisfies the Delaunay criterion, avoiding elongated triangles.

[0035] Sub-step S33: Providing Multiple Schemes. The system generates alternative schemes in parallel based on three different surface extension strategies: Strategy A uses minimum curvature extension, with the loss function emphasizing surface smoothness; Strategy B uses symmetry-first extension, with the loss function including a symmetry axis detection term; Strategy C uses volume-filling-first extension, with the loss function encouraging maximization of internal volume. Each strategy corresponds to a different combination of loss function weights. The system dynamically calculates the activation probability of each strategy based on the number of symmetry axes s of the input wireframe, the aspect ratio r of the bounding box, and the current user's historical selection preferences h.

[0036] in These are preset coefficients. Finally, the system selects the 2 to 3 schemes with the highest probability and displays them side by side in the overall generation panel on the right as 128×128 pixel thumbnails. Each thumbnail is accompanied by a strategy label (such as "smoothness priority" or "symmetry priority").

[0037] Sub-step S34: Surface Addition. After the user clicks on a thumbnail, the system parses the complete mesh data for the corresponding scheme, including vertex coordinates, face indices, and UV coordinates. The surface is instantiated as a new geometric object and added to the currently active layer. Its creation parameters (such as the strategy used and the input wireframe ID) are recorded as a new node in the scene graph, establishing a dependency relationship with the upstream wireframe node to ensure the traceability of subsequent edits.

[0038] In the above method, step S4, namely intelligently adding generated materials, involves calling the AI ​​material generation model to perform the following sub-steps: Sub-step S41: Local or global selection. Users can select a single facet in the 3D viewport by clicking, or by dragging a selection box to select multiple consecutive faces. The system maintains a facet selection set. Additionally, users can click on the root node of the model in the right-hand layer panel to trigger a global selection. Includes all faces of the model. After the selection is complete, the system highlights the selected area and activates the material assignment command.

[0039] Sub-step S42: Intelligent Material Matching. The AI ​​material generation model incorporates a multi-scale geometric feature encoder. This encoder extracts features at three scales: local scale (radius of 5% of the model's maximum size), mesoscale (20%), and global scale (100%). At each scale, four core features are calculated: mean curvature H, normal distribution entropy E (measuring surface orientation diversity), concavity / convexity D (standard deviation based on depth variation), and orientation consistency C (cosine of the angle between the principal normal vector and the global Z-axis). These 12-dimensional feature vectors are input into a generator of a conditional generative adversarial network (cGAN). The generator outputs a set of physically realistic material parameters, including diffuse reflectance coefficients. Specular reflection intensity .

[0040] Sub-step S43: Material Library Selection and Replacement. The right panel displays a structured material library, functionally divided into four main categories: glass, metal, plastic, and organic materials. Each category has subcategories; for example, metal includes aluminum, copper, stainless steel, and brushed gold; organic materials include wood, leather, fabric, and rubber. Each subcategory is associated with a 128-dimensional learnable material prototype vector. When a user selects a subclass, the system will use the current geometric feature vector. and The data is then stitched together and fed into a fine-tuning network, which outputs personalized material parameters. For example, when the "wood" subcategory is selected, the fine-tuning network adjusts the grain direction based on curvature characteristics and the annual ring density based on unevenness.

[0041] Sub-step S44: Material Variant Preview. For the selected material category, the system pre-generates 3 to 5 variants. For example, under the glass category, it provides clear glass (refractive index 1.5, roughness 0.01), frosted glass (refractive index 1.5, roughness 0.3), and colored glass (with added blue absorption coefficient). Each variant is arranged horizontally in the panel as a small spherical texture map rendered in real time. The user can hover over it to preview its application effect in the selected area, and click to confirm the application.

[0042] In the above method, step S5, namely AI-generated and rendered image, involves calling the AI ​​rendering engine to execute the following sub-steps: Sub-step S51: Viewpoint Determination. Users can freely rotate, pan, and zoom the view using the orbital ball tool, or select from a preset viewpoint library (such as front view, side view, isometric view, and three-point perspective). The system records current camera parameters, including position. Target point Upward and field of view (FOV).

[0043] Sub-step S52: Intelligent Rendering. The AI ​​rendering engine employs a lightweight neural radiation field approximation model. In the offline phase, this model is pre-trained on a dataset containing 100,000 high-quality rendering images, learning an end-to-end mapping from (geometry + material + viewpoint) to pixel color. In the online phase, the engine encodes the current model's mesh data, material parameters, and camera parameters into feature tensors, which are then input into a U-Net architecture rendering network. The network completes the joint calculation of lighting, reflection, refraction, and shadows within 200 milliseconds, outputting a 512×512 resolution preview image.

[0044] Sub-step S53: Preview image generation. The rendering result is displayed in a pop-up window in the center of the interface, with three operation buttons at the bottom: "Accept," "Regenerate," and "Adjust Lighting." After acceptance, the rendering image is cached and bound to the current scene state, serving as a reference for design review.

[0045] To illustrate the operation process of this invention, an application example is constructed: A user wants to create a modern-style table lamp model. First, the user freely draws the lampshade outline on a 3D canvas, and the system optimizes it into a smooth curve in real time. Next, the user draws the circular outline of the base, which the system automatically recognizes as a closed outline. The user selects these two outlines and clicks "Generate Surface." The AI ​​provides two options: Option 1 generates independent surfaces for the lampshade and base respectively; Option 2 generates a continuous surface. The user chooses Option 1. Subsequently, the user selects the lampshade surface and calls the AI ​​to generate materials. The system analyzes its high curvature characteristics and recommends glass. The user selects "glass" from the material library and chooses "frosted glass" from the variants. Finally, the user switches to a three-point perspective view, generates a rendering preview, and confirms satisfaction with the design effect. Throughout the process, all operations are stored as parametric nodes. The user can return at any time to modify the lampshade line drawing, and the system will automatically re-optimize the curves, generate surfaces, and update materials and rendering. Example 2

[0046] In another implementation, the AI ​​curve optimization model uses a Transformer architecture instead of Seq2Seq. This Transformer comprises a 6-layer encoder, each layer containing a multi-head self-attention mechanism and a feedforward network. Position encoding employs learnable absolute position embeddings instead of traditional sine coding to better adapt to non-uniform sampling in 3D space. The training dataset is expanded to 800,000 pairs, and data augmentation strategies are introduced: random affine transformations (rotation, scaling, shearing) are applied to the original hand-drawn curves to improve the model's generalization ability. The optimized curves not only satisfy the curvature standard deviation constraint but also additionally guarantee that the total length change rate does not exceed 10%.

[0047] In the surface generation stage, the AI ​​surface generation model employs a pure graph neural network architecture, abandoning implicit fields. Specifically, the GNN directly outputs the vertex coordinates and connectivity relationships of the quadrilateral mesh. The loss function includes four terms: boundary fitting loss. Curvature smoothing loss Symmetry loss and volume loss The weights of each element are adjusted online via Bayesian optimization to adapt to different design styles. When generating multiple schemes, the system uses two fixed strategies: minimum curvature and volume filling, but allows users to continuously adjust the weights of these two strategies via sliders to achieve a gradual transition between schemes.

[0048] In terms of material generation, the AI ​​material generation model incorporates lighting conditions as additional input. The system provides three default ambient light presets: indoor soft light, outdoor sunlight, and studio three-point lighting. After the user selects the lighting conditions, The generator adjusts material parameters accordingly to ensure optimal visual effects under specific lighting conditions. For example, under strong direct light, the specular reflection intensity of metallic materials is automatically enhanced. Example 3

[0049] In another implementation, the system integrates a voice interaction module. Users can trigger AI operations via voice commands, such as "Generate a surface for this area," "Apply a metallic material," or "Render the current viewpoint." The voice recognition module uses an end-to-end Conformer model, achieving an accuracy of 98% in quiet environments. The recognition results are bound to the current selection set and converted into corresponding API calls.

[0050] Furthermore, the non-destructive editing workflow has been enhanced. All AI-generated operations are encapsulated as editable nodes, with node attribute panels exposing key parameters for user fine-tuning. For example, the curve optimization node provides a "fidelity" slider (0-100) to control the degree of detail retention; the surface generation node provides dual sliders for "smoothness" and "volume"; and the material node allows direct editing of all physical parameters. After any node parameter is modified, the dependency graph automatically marks downstream nodes as "dirty" and incrementally recalculates them during idle periods to ensure real-time feedback.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional modeling method based on spatial hand-drawing and AI intelligent completion, characterized in that, Includes the following steps: In response to the user's free-space hand-drawing operation using a stylus on a 3D canvas, an original 3D space curve is generated. The user can draw directly on any plane in 3D space without switching modes or inputting coordinates. The original 3D space curve is smoothed and optimized by a pre-trained AI curve optimization model, which identifies and retains the original design features while reducing noise caused by hand-drawing jitter, and generates an optimized space line drawing. In response to the user's selection of at least one closed wireframe in the 3D canvas and triggering the "Generate Surface" command, the system acquires the geometric information of the selected wireframe. The determination of a closed wireframe combines geometric closure, topological connectivity, and visual closure perception. An AI surface generation model is invoked to analyze the topological structure of the selected wireframe, automatically identifying and intelligently repairing minor gaps or incompletely closed areas between wireframes. Based on the repaired wireframe, a 3D surface composed of quadrilateral meshes is generated, and the number of control points on adjacent boundaries is automatically matched to ensure a clean topological structure. When multiple possible surface generation methods exist, several alternative topology schemes are previewed in the interface for the user to select and confirm. When a user selects an alternative, the selected 3D surface is added to the current layer. In response to a user's instruction to assign materials to at least a portion of the surfaces in the model, the AI ​​material generation model is invoked to analyze the geometric features and design context of the selected area, automatically matching and generating initial material properties that meet the requirements. A material library categorized by function and multiple material variants are provided in the interface for the user to preview and select. In response to a user's selected specific viewpoint and the "Generate Render Preview" instruction, the AI ​​rendering engine is invoked to quickly calculate lighting, reflection, refraction, and shadow effects based on the current model's geometric information, assigned material properties, and lighting conditions, generating a preview-level rendered image and displaying it in the interface.

2. The three-dimensional modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, The AI ​​curve optimization model adopts a sequence-to-sequence neural network architecture based on an attention mechanism. During the training phase, the model uses a dataset of hand-drawn curves labeled by professional designers and paired with corresponding ideal curves for supervised learning. It can distinguish between key inflection points in the design intent and hand-drawn jitter noise, and reduce the standard deviation of the curvature change of the optimized curve to below a predetermined proportion of the standard deviation of the curvature change of the original hand-drawn curve.

3. The three-dimensional modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, In the closed wireframe determination logic, geometric closure is achieved by calculating the Euclidean distance between the first and last endpoints of the wireframe and comparing it with a predetermined multiple of the average stroke width in the current view. Visual closure perception is achieved by using a deep learning-based visual closure prediction module to intelligently identify wireframes that the human eye subjectively believes should be closed but are not fully connected.

4. The three-dimensional modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, The AI ​​surface generation model adopts a hybrid architecture combining graph neural networks and implicit fields. It converts the input wireframe into a weighted graph structure, extracts global topological semantics through multi-layer graph convolution, and drives the implicit neural radiation field to generate continuous surfaces. The generated quadrilateral meshes satisfy the boundary consistency constraint. The number error of vertices on the shared boundary between adjacent surface patches does not exceed a preset threshold, and the internal triangulation satisfies the Delaunay criterion.

5. The three-dimensional modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, The alternative topology schemes are generated based on different surface extension strategies, including minimum curvature extension, symmetry-first extension, and volume-filling-first extension. Each strategy corresponds to a different combination of loss function weights. The system dynamically adjusts the activation probability of each strategy according to the number of symmetry axes of the wireframe, the bounding box ratio, and historical user selection preferences.

6. The 3D modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, The AI ​​material generation model incorporates a multi-scale geometric feature encoder, which extracts curvature, normal distribution entropy, concavity / convexity, and orientation consistency features at local, mesoscopic, and global scales, respectively. These features are input into a conditional generative adversarial network to output a set of material parameters including diffuse reflection coefficient, specular reflection intensity, roughness, metallicity, and subsurface scattering coefficient.

7. The 3D modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, The material library is divided into multiple categories according to function, each category contains multiple subcategories, and each subcategory is associated with a set of learnable material prototype vectors. When a user selects a subcategory, the AI ​​model fine-tunes the prototype vectors based on the current geometric features to generate a personalized material instance that fits the current surface.

8. The 3D modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, The AI ​​rendering engine employs a lightweight neural radiation field approximation model. This model is pre-trained offline on a dataset containing high-quality rendering images, learning the mapping relationship from geometry-material combination to pixel color. During online rendering, it generates a preview image of a predetermined resolution within a predetermined time period. Its lighting effects and physically based path tracing results reach a preset threshold in terms of structural similarity index.

9. The 3D modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, All AI-generated surfaces, materials, and rendering results are stored in the scene graph as parametric nodes. Users can return to any historical node at any time to modify the input conditions, and all subsequent dependent nodes will be automatically recalculated, realizing a non-destructive editing workflow.

10. The 3D modeling method based on spatial hand-drawing and AI intelligent completion according to claim 1, characterized in that, When the system pauses for more than a predetermined time period during the user's drawing process, it preloads the AI ​​optimization and surface generation module in the background to predict the user's next operation intention. If it detects that the user is drawing a wireframe that is close to a closed shape, it automatically highlights the potential closed area and displays a mini "Generate Surface" prompt icon.

Images