Three-dimensional model generation method and system based on multi-scale autoregression model

Abstract

The invention discloses a three-dimensional model generation method and system based on a multi-scale autoregression model. The system comprises a three-dimensional shape expression module, an efficient autoregression model construction module based on a window attention mechanism and a three-dimensional model generation module. Multi-scale binary serialization three-dimensional shape expression based on an octree is established, and a binary vector quantization variational automatic encoder based on the octree is constructed; the method comprises the following steps: constructing an efficient autoregression model based on a window attention mechanism, and designing three-dimensional rotation position coding and a multi-Token prediction strategy by using a Transform model based on an octree and a hierarchical perception mask mechanism; training an autoregression model, and predicting the obtained binary coding sequence; and generating a three-dimensional shape by using the trained autoregression model and a binary vector quantization variational automatic encoder based on an octree. According to the method, high-quality, diversified and multi-mode shape generation can be realized, and the efficiency and quality of three-dimensional shape modeling generation are improved.

Description

Technical Field

[0001] The present invention belongs to the fields of computer graphics technology and artificial intelligence technology, and relates to a three-dimensional model generation technology based on a multi-scale autoregressive model, which can generate high-quality and diversified three-dimensional shapes unconditionally or under given multiple input conditions such as sketches, pictures, texts, etc. Background Art

[0002] The task goal of 3D shape generation is to automatically create or complete complex 3D geometric structures through advanced technologies such as deep learning. Especially in the context of the current booming development of personalized customization and intelligent manufacturing, the application of this technology is not limited to traditional fields, but will also play an important role in customized product development, digital twins and even metaverse content generation. In recent years, due to the breakthrough progress of diffusion models, 3D content generation technology has developed rapidly. At the same time, autoregressive models have become one of the core technologies in the field of general artificial intelligence due to their outstanding performance in natural language processing, 2D image generation and multimodal understanding. However, in terms of 3D shape generation, the performance of existing autoregressive models still lags behind that of diffusion models. Therefore, exploring how to use autoregressive models to efficiently generate 3D shapes has become the focus of this study.

[0003] Traditional autoregressive models face two major challenges when dealing with 3D generation tasks: On the one hand, since 3D shapes lack a natural sequential structure, the existing method of constructing a one-dimensional sequence according to the 3D scan order will destroy the hierarchical characteristics and spatial locality of the shape, resulting in slow model convergence and limited generation quality; on the other hand, high-resolution 3D shapes require tens of thousands of tokens to describe the details, and directly using the standard Transformer architecture will cause the computational complexity to grow quadratically, resulting in a sharp increase in resource consumption during training and inference. Although some studies have attempted to reduce the computational load by directly generating 3D meshes or using low-resolution token solutions, these methods have significant bottlenecks in the ability to express geometric details and the generation resolution, making it difficult to meet the needs of high-quality 3D model generation, and unable to build a multi-modal 3D shape generation general model, and their application in industrial design, virtual reality, digital twins and other fields is limited. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention proposes a three-dimensional shape generation model method and system based on a multi-scale autoregressive model, which uses a three-dimensional shape data set for model training, can achieve high-quality, diverse, and multi-modal shape generation, and improve the efficiency and quality of three-dimensional shape modeling.

[0005] For the convenience of explanation, the present invention agrees on the following definitions of terms: Token: The smallest unit processed by the attention mechanism.

[0006] VQVAE: Vector Quantized Variational Autoencoder, a model architecture proposed in the literature (Van Den Oord, Aaron, and Oriol Vinyals. "Neural discrete representation learning." Advances in neural information processing systems 30(2017).) for compressing input data into discrete tokens.

[0007] BSQ: Binary Spherical Quantization, a method proposed in the paper (Zhao, Yue, Yuanjun Xiong, and Philipp Krähenbühl. "Image and Video Tokenization with Binary Spherical Quantization." arXiv preprint arXiv:2406.07548 (2024).) to discretize the feature vector of VQVAE into binary tokens.

[0008] Transformer: The model architecture based on the attention mechanism proposed in the literature (Vaswani, A. "Attention is all you need." Advances in Neural Information Processing Systems (2017).) The literature (Wang, Peng-Shuai. "Octformer:Octree-based transformers for 3d pointclouds." ACM Transactions on Graphics (TOG) 42.4 (2023): 1-11.) designed a more efficient Transformer structure for octrees.

[0009] The core of the present invention is to propose a multi-scale three-dimensional serialized shape expression with the octree structure as the core. The octree naturally expresses the multi-scale hierarchical characteristics of the three-dimensional shape through a recursive splitting mechanism, and its Z-curve sorting strategy effectively retains the spatial locality, providing an ideal serialization basis for autoregressive prediction. Compared with the method of directly predicting three-dimensional coordinates, this method decomposes the three-dimensional shape generation task into a binary sequence of octree node split states and binary vector quantization, and draws on the progressive reasoning paradigm of the Chain-of-Thought to significantly improve the model convergence speed and generation quality. In terms of computational efficiency, the present invention reduces the time complexity to a linear level through an improved window attention mechanism, and combines the parallel Token generation strategy of hierarchical perceptual masks to improve the training speed.

[0010] The technical solution provided by the present invention is as follows: The method proposed in the present invention constructs an octree-based binary vector quantized variational autoencoder VQVAE (binary vector quantized VQVAE), which constructs an octree according to whether the three-dimensional shape is in the octree node, and is used to compress the input three-dimensional shape into a binary sequence; and an efficient autoregressive model based on the window attention mechanism is used to generate a three-dimensional shape according to the input conditions. The steps of the method of the present invention include: a first step, constructing an octree according to an input three-dimensional shape; a second step, constructing a VQVAE based on the octree, using the VQVAE to compress the deep leaf nodes of the octree obtained in the previous step to the shallow layer, obtaining the feature vectors on the shallow leaf nodes, and then using the binary vector quantization formula to discretize the continuous feature vector into a binary code (i.e., a binary feature vector); a third step, constructing an efficient autoregressive model based on a window attention mechanism; splicing the octree split state sequence and the binary feature vector obtained in the previous step to construct a binary sequence of the input three-dimensional shape, using the constructed binary sequence to train the efficient autoregressive model based on the window attention mechanism according to the deep learning training mode; a fourth step, using the trained efficient autoregressive model based on the window attention mechanism and the binary vector quantization variational autoencoder based on the octree to generate a three-dimensional shape.

[0011] First, the present invention designs a binary VQVAE to encode three-dimensional shapes into multi-scale binary sequences. Then, the present invention trains an efficient autoregressive model based on a window attention mechanism to predict the sequence obtained in the previous step in an autoregressive manner. Hierarchical-aware masks are used in the model to maintain the hierarchical dependency mechanism of the octree. In order to improve the efficiency and effect of the model, the present invention also introduces a window attention mechanism, a multi-Token parallel prediction strategy, and three-dimensional position encoding. Finally, the present invention can generate three-dimensional shapes step by step from the overall to the details according to a variety of input conditions, and surpasses other existing methods in efficiency and quality.

[0012] In the first part, a multi-scale binary serialized 3D shape expression based on octree is established, and the 3D shape generation task is decomposed into a binary sequence of octree node splitting states and binary quantization encoding.

[0013] The present invention proposes an octree-based multi-scale binary serialized three-dimensional shape expression, which is composed of the octree node split state and the binary Token generated by VQVAE. Given an input three-dimensional shape, it is first constructed as an octree, and a continuous feature vector is obtained through the octree node split state to build a binary sequence containing multi-scale geometric information. The first three levels of the octree are forced to be filled with full voxel grids. Starting from the fourth level, the octree recursively subdivides at non-empty voxels and does not operate at empty voxels until the specified maximum number of levels, that is, the maximum depth D. The octree node split state is used Indicates that is the number of the octree node, 0 and 1 represent the node not split and split respectively. Arrange the node splitting status of the dth layer into a sequence according to the Z-shaped curve , the node splitting states of each layer are concatenated in increasing order of depth to obtain the sequence , and finally construct a binary sequence containing multi-scale geometric information. This binary sequence can be used to reconstruct the original octree structure and fully preserve the geometric features of the three-dimensional shape.

[0014] In order to supplement the detail expression ability of the octree, the present invention innovatively introduces the octree-based VQVAE architecture for efficient encoding and reconstruction of the signed distance field (SDF) and geometric details. The model adopts an asymmetric encoder-decoder design to enhance the reconstruction ability of the decoder. The encoder uses an octree-based sparse convolutional neural network (O-CNN), which can compress the input octree by 2 layers and represent the detail information as a feature vector on the leaf node. The encoder then uses binary spherical quantization (BSQ) technology to convert the feature vector into a binary Token. The decoder adopts a dual octree graph convolutional network architecture to decode the binary Token into a local signed distance field SDF, which is integrated into a global SDF through a multi-level unit decomposition method (MPU) to achieve high-precision surface reconstruction. The present invention uses the loss function for the reconstruction of the signed distance field SDF:

[0015] in, and are the predicted and true SDF values, is the coefficient, is the sampled point cloud set, is the number of point clouds. Vector quantization loss, SDF reconstruction loss and octree splitting loss are used in the model training process to ensure accurate restoration of geometric details and structural integrity.

[0016] The binary quantization coding loss used in the present invention is as follows, where is the entropy function, is the continuous feature vector of the input, is the discrete feature vector after binary quantization:

[0017] The present invention also uses octree splitting loss to supervise the octree structure predicted by VQVAE, where is the predicted splitting state of the i-th octree node, The corresponding true value is:

[0018] The overall loss function for training the octree-based binary vector quantization variational autoencoder proposed in the present invention is as follows, where is the SDF reconstruction loss, is the binary quantization coding loss, The octree splitting loss is:

[0019] The binary tokens generated by quantization are serialized in the order of the Z-curve and combined with the binary split sequence of the octree to form the final serialized octree three-dimensional expression. This representation has multiple advantages: all data elements are binary tokens, which greatly simplifies the prediction process of the autoregressive model; the multi-level octree structure realizes the organic integration of multi-scale features of three-dimensional shapes; the Z-sorting mechanism effectively retains the spatial locality characteristics, which significantly improves the efficiency of autoregressive prediction compared to traditional scanning sorting. This representation method provides efficient underlying data support for applications such as the generation, editing and compression of three-dimensional shapes.

[0020] The second part is an efficient autoregressive model based on the window attention mechanism. This model uses an efficient Transformer model based on an octree structure and a hierarchical perceptual mask mechanism, and designs a three-dimensional rotation position encoding and multi-Token prediction strategy based on the characteristics of three-dimensional data to improve generation quality and efficiency.

[0021] First, the present invention uses an efficient Transformer architecture for a three-dimensional octree structure. In order to solve the problem of high computational complexity of traditional autoregressive models when processing long sequences, the present invention adopts an octree-based window attention mechanism (OctFormer), which divides the binary vector quantization Token sequence based on octree serialization into fixed-size attention windows for local self-attention calculation, and realizes cross-window Token interaction by alternating between the expansion attention and shift window attention modules. This design breaks through the window size limit of the prior art on the Token, allowing Tokens from different windows to interact globally, effectively capturing the local geometric features and global structure of the three-dimensional shape while maintaining computational efficiency.

[0022] Secondly, the present invention proposes a position information encoding method that combines three-dimensional rotation position encoding (RoPE3D) with a learnable hierarchical embedding vector. In response to the problem of lack of three-dimensional position sensitivity in the traditional Transformer architecture, the present invention extends the two-dimensional rotation position encoding to three-dimensional space, and establishes the spatial position relationship of the Token through the rotation matrix operation of the three-dimensional space coordinates. At the same time, in order to distinguish the semantic information of different octree depth levels, a learnable embedding vector is introduced to encode the octree hierarchical features. This composite encoding mechanism enables the model to accurately identify the spatial coordinates and the depth of the level to which the Token belongs in the three-dimensional space, providing key position prior information for autoregressive prediction.

[0023] Furthermore, the present invention develops a hierarchically constrained multi-token parallel generation strategy. In view of the low efficiency of token-by-token prediction of traditional autoregressive models, the present invention adopts a masked autoregressive model (MAR) framework to achieve multi-token parallel prediction. The present invention designs a mask mechanism with hierarchical perception capabilities, which locally shuffles and masks tokens of the same depth level without changing the hierarchical dependencies of the octree. Specifically, in the sequence generation process, the depth level is predicted sequentially from shallow to deep, and tokens are allowed to be generated in parallel in a random order within the same octree layer, while ensuring that the prediction of deep tokens can obtain the generated shallow token information. This strategy maintains the hierarchical generation logic of the octree structure and increases the generation speed several times by dynamically controlling the information flow path.

[0024] In the specific implementation process, the model generates node splitting signals layer by layer from the third layer of the octree to obtain an octree structure until the maximum depth D layer, at which a quantized binary Token representing geometric details is generated. The octree and binary Token generated in the above steps are input into the trained octree-based binary vector quantization VQVAE to reconstruct the three-dimensional shape. This multi-scale generation mechanism realizes the progressive synthesis of three-dimensional shapes from basic voxel structures to fine surface features through a hierarchical decoding strategy. Experiments show that the autoregressive model of the present invention increases the sequence generation speed by more than 69 times compared with traditional methods while maintaining the generation quality, and significantly enhances the modeling ability of complex geometric structures.

[0025] The present invention also implements a three-dimensional model generation system based on a multi-scale autoregressive model, including a three-dimensional shape expression module based on octree binary vector quantization and an efficient autoregressive model construction module based on a window attention mechanism; wherein: The three-dimensional shape expression module based on octree binary vector quantization is used to encode and decode the input three-dimensional shape to the binary vector quantized Token sequence; the efficient autoregressive model construction module based on the window attention mechanism is used to autoregressively generate the binary vector quantized Token sequence.

[0026] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a three-dimensional model generation method and system based on a multi-scale autoregressive model, and designs a multi-scale three-dimensional serialized shape expression with an octree structure as the core. Compared with the traditional solution of directly predicting three-dimensional coordinates, this method decomposes the generation task into a binary sequence of node splitting states and binary quantization encoding, and significantly improves the model convergence speed and generation quality based on the progressive reasoning paradigm of the Chain-of-Thought. In terms of computational efficiency, the present invention reduces the time complexity to a linear level through an improved window attention mechanism, and combines a parallel Token generation strategy of hierarchical perceptual masks. Compared with the baseline method, the training speed of the present invention is increased by 13 times, the generation efficiency is increased by 69 times, and the generation quality and scene scale support exceed the existing diffusion model. The present invention constructs a multimodal general model, which has broad application prospects in industrial design, virtual reality, digital twins and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is an overall model diagram of the three-dimensional shape generation method provided by the present invention.

[0028] Figure 2 It is a flowchart of the training and reasoning process of the present invention.

[0029] Figure 3 It is a model diagram of the VQVAE part of the present invention.

[0030] Figure 4 It is a diagram of the hierarchical perceptual mask, window attention and shifted window attention mechanism of the autoregressive model part of the present invention.

[0031] Figure 5 It is the octree schematic diagram and Z-type sorting schematic diagram of the present invention.

[0032] Figure 6 These are the unconditional generation experimental results of the present invention on the ShapeNet and Objaverse datasets.

[0033] Figure 7 This is a comparison between the present invention and the existing autoregressive model on the ShapeNet dataset.

[0034] Figure 8 This is the experimental result of the sketch conditional generation based on the ShapeNet dataset.

[0035] Fig. 9 This is the experimental result of image conditional generation on the ShapeNet dataset of the present invention.

[0036] Fig.10 This is the scene generation experiment result of the present invention on the Room dataset.

[0037] Fig.11 is a comparison of the present invention with a single-scale baseline method.

[0038] Fig.12 The diversity of the three-dimensional shapes generated by the present invention is demonstrated.

[0039] Fig.13 The time efficiency of the present invention is demonstrated. DETAILED DESCRIPTION

[0040] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0041] The present invention provides a multi-scale autoregressive three-dimensional generation model, which consists of two parts. The overall process and network structure are shown in the attached figure. Figure 1 As shown, the first part is an octree-based binary serialization three-dimensional shape expression module, and the second part is an efficient autoregressive model based on the window attention mechanism.

[0042] The specific steps include: 1) Construct an octree based on the input 3D shape, as shown in the following figure. Figure 5 According to the existing octree construction method, the first three levels of the octree are forced to be full voxels. Starting from the fourth level, the octree is recursively subdivided at non-empty voxels until the sixth level. Figure 5As shown in (b) and (c), the present invention proposes to arrange the node split states 0 and 1 of the dth layer according to a Z-shaped curve to form a binary sequence , the node splitting states of each layer are concatenated in ascending order of depth to obtain the octree splitting state sequence, which is expressed as .

[0043] 2) This paper proposes a binary vector quantization VQVAE based on octree sparse convolution and graph convolution, which compresses the input octree into feature vectors on leaf nodes, as shown in the attached figure. Figure 3 The encoder part is composed of a residual connection block based on octree sparse convolution, which compresses the input octree from the 8th layer to the 6th layer, and then discretizes the feature vectors on the leaf nodes of the 6th layer into a binary Token sequence after binary spherical quantization. The decoder is an asymmetric structure. The residual block of the octree graph convolution first downsamples the discretized binary token to the third layer, and then upsamples it to the original eighth layer. From the sixth layer to the eighth layer, VQVAE predicts the SDF value of the sampling point in the three-dimensional space for each layer and calculates the loss function.

[0044] 3) The present invention proposes to split the state sequence of the octree And VQVAE's Token Sequence Splice into binary sequences, as shown in Figure 1 In the training process, the model randomly shuffles and masks the same-layer tokens of the binary sequence, inputs them into the octree transformer, and predicts the masked tokens. In order to keep the hierarchical dependency of the octree unchanged, the present invention proposes a hierarchical-aware mask, as shown in the attached figure. Figure 4 (a) in the figure. Under the effect of this mask, deep tokens can perceive shallow information during training, while shallow tokens cannot perceive deep information, so that the model can be generated layer by layer from shallow to deep. To solve the quadratic complexity problem of traditional Transformer, this paper adopts a window attention mechanism and alternately uses expansion windows and shift windows, as shown in the attached figure. Figure 4 (b) and (c) in the figure. Tokens in the same window can perceive each other through the self-attention mechanism, and by expanding or shifting the window, tokens can also perceive the global information, allowing the model to predict the masked token more accurately.

[0045] 4) During inference, the model generates the split states of the octree nodes from the 3rd layer to the 5th layer according to the conditional information provided by the user, such as text, sketches, and pictures, and grows the octree nodes of the next layer according to the split states of the previous layer. At the 6th layer node, the model generates a binary token on each leaf node. Finally, the octree structure and the token on the leaf node are input into the VQVAE decoder to predict the SDF field in the three-dimensional space, and then the mesh surface is extracted through the MarchingCubes algorithm, as shown in the attached figure. Figure 1 shown.

[0046] Attached Figure 6 These are the results obtained by the unconditional generation experiment of the present invention. The left half is the result on the ShapeNet dataset, and the right half is the result on the Objaverse dataset. Figure 8 It is demonstrated that the present invention can generate corresponding three-dimensional shapes based on input sketches. Fig. 9 It shows that the present invention can generate corresponding three-dimensional shapes according to the input image. Fig.10 The above results show that the model of the present invention can generate scenes with a resolution of up to , three-dimensional shapes with rich details and good diversity, and can complete various conditional generation tasks such as text, sketches, pictures and the generation of complex scenes.

[0047] Attached Figure 7 The present invention is compared with the existing autoregressive model, wherein Figure 7 (a) in the figure is AutoSDF (Mittal et al. 2022), Figure 7 (b) in the figure is 3DILG (Zhang et al. 2022), Figure 7 (c) in the figure is MeshGPT (Siddiqui et al. 2024). The figure shows the token sequence lengths of various methods. The present invention can process token sequence lengths far exceeding those of other methods and generate higher quality three-dimensional shapes.

[0048] Attached Fig.11 This is a comparison between the present invention and the single-scale baseline method based on coordinate prediction. The progressive generation of multiple scales makes the model converge faster, and it can converge initially in only 10 cycles, while the baseline method still cannot generate good quality results after 100 cycles.

[0049] Attached Fig.12The top part shows the histogram of the chamfer distance between the shape generated by the present invention and all the three-dimensional shapes in the data set. The generated result of the present invention is not exactly the same as the data in the data set. The bottom part shows the generated result of the present invention (left) and the most similar three-dimensional shape in the data set (right). It can be seen that the present invention can generate three-dimensional shapes that have never appeared in the data set.

[0050] Attached Fig.13 The left figure compares the speed of the present invention when using window attention or not during training. This module can speed up the model by 13 times when the length of the token sequence exceeds 160k. The right figure compares the speed of the present invention when using window attention or not during reasoning. The above module can speed up the model by 69 times when the length of the token sequence exceeds 40k.

[0051] In summary, the present invention provides an efficient 3D shape generation system of a multi-scale autoregressive model. Given input conditions such as text, sketches, and pictures, the model can generate high-quality and diverse 3D shapes that meet the conditions. The system consists of two parts. The first part is an octree-based serialized binary encoding, which achieves efficient compression of 3D shapes through the multi-level serialized expression of the octree structure and the VQVAE model. The second part is efficient autoregressive generation, which uses an improved Transformer architecture and hierarchical-aware masks, combined with 3D position encoding and parallel generation strategies, which can significantly improve the generation speed and geometric detail restoration capabilities.

[0052] The above is a clear and complete description of the technical solutions in the embodiments of the present invention in combination with the drawings in the embodiments of the present invention. It is understandable that the examples described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

Claims

1. A three-dimensional model generation method based on a multi-scale autoregressive model, characterized in that: include: Establish a multi-scale binary serialized 3D shape expression based on octree, construct a binary vector quantized variational autoencoder based on octree, decompose the input 3D shape generation task into octree node splitting states, and obtain a continuous feature vector containing multi-scale geometric information; Compress the continuous feature vector into a discrete binary code sequence through binary vector quantization; Build an efficient autoregressive model based on window attention mechanism, use the octree-based Transformer model and hierarchical perception mask mechanism, and design three-dimensional rotation position encoding and multi-token prediction strategy; The obtained binary code sequence of the three-dimensional shape is used to train an efficient autoregressive model based on the window attention mechanism, and the obtained binary code sequence is predicted to obtain a trained autoregressive model; Generate 3D shapes using the trained autoregressive model and an octree-based binary vector quantized variational autoencoder.

2. The three-dimensional model generation method based on the multi-scale autoregressive model as claimed in claim 1, characterized in that: The specific steps include: In the first step, an octree is constructed based on the input 3D shape; nodes are arranged into sequences based on the splitting status of the octree nodes to construct a binary sequence containing multi-scale geometric information; The second step is to build an octree-based vector quantization variational autoencoder, and use the octree-based vector quantization variational autoencoder to compress the three-dimensional octree to obtain the feature vector on the leaf node; The third step is to concatenate the octree split state sequence and the discretized feature vector into a binary sequence of the input three-dimensional shape; Build an efficient autoregressive model based on window attention mechanism; train the autoregressive model using binary sequences; Convert the feature vectors on the octave leaf nodes into binary Token sequences; decode the binary Tokens into local signed distance fields and integrate them into global signed distance fields to achieve high-precision surface reconstruction; OctFormer, an octree-based window attention mechanism, is used to divide the token sequence into fixed-size attention windows for local self-attention calculation, and cross-window token interaction is achieved by alternating between the expansion attention and shift window attention modules. Design a position information encoding method that combines three-dimensional rotation position encoding with learnable hierarchical embedding vectors; extend the two-dimensional rotation position encoding to three-dimensional space, and establish the spatial position relationship of the token through the rotation matrix operation of the three-dimensional space coordinates; Encode octree hierarchical features through learnable embedding vectors; In the fourth step, the trained autoregressive model and the octree-based vector quantized variational autoencoder are used to generate 3D shapes.

3. The three-dimensional model generation method based on the multi-scale autoregressive model as claimed in claim 2, characterized in that: The first step is to construct a binary sequence containing multi-scale geometric information, including: The first three levels of the octree are forced to be filled with full voxel grids; starting from the fourth level, the octree recursively subdivides at non-empty voxels and does nothing at empty voxels until the specified maximum number of levels, i.e. the maximum depth D; Octree node split status Indicates that is the number of the octree node, 0 and 1 represent the node not split and split respectively; the node splitting status of the dth layer is arranged into a sequence according to the Z-shaped curve ; The node splitting states of each layer are concatenated in increasing order of depth to obtain a sequence , and finally constructs a binary sequence containing multi-scale geometric information.

4. The method for generating a three-dimensional model based on a multi-scale autoregressive model as claimed in claim 2, characterized in that: Loss function used to train octree-based binary vector quantized variational autoencoders It is expressed as: in, is the signed distance field reconstruction loss, is the binary quantization coding loss, is the octree splitting loss.

5. A three-dimensional model generation system based on a multi-scale autoregressive model, used to implement the three-dimensional model generation method based on a multi-scale autoregressive model as claimed in claim 1, characterized in that: include: A 3D shape expression module based on octree binary vector quantization and an efficient autoregressive model building module based on window attention mechanism; in: The three-dimensional shape expression module based on octree binary vector quantization is used for encoding and decoding the input three-dimensional shape into a binary vector quantized Token sequence; The efficient autoregressive model building module based on the window attention mechanism is used to autoregressively generate a binary vector quantized Token sequence.

Images