A generative semantic communication system for 3D content generation

Abstract

This application discloses a generative semantic communication system for 3D content generation, comprising: a transmitter deployed on an edge device, a receiver deployed on a cloud server, and a wireless channel connecting the transmitter and receiver; the transmitter extracts visual semantic tokens from a single input 2D image to form a visual semantic token sequence; the visual semantic token sequence and the signal-to-noise ratio of the current wireless channel are used as input and mapped to a channel input symbol sequence for transmission; the receiver recovers the semantic token sequence from the received signal; and a 3D mesh model is generated based on the recovered semantic token sequence; thereby, by transmitting compact visual semantic tokens instead of raw data, the wireless bandwidth requirement is significantly reduced; an efficient adaptive transmission architecture based on structured state space duality is adopted, which improves parallel efficiency while maintaining linear computational complexity, and integrates a channel adaptive state adjustment module to dynamically adapt to time-varying channels.

Description

Technical Field

[0001] This application relates to the technical field of data transmission, and more particularly to a generative semantic communication system for 3D content generation. Background Technology

[0002] With the rapid development of extended reality (XR) applications, users have an increasingly urgent need to generate high-fidelity 3D content instantly on mobile edge devices.

[0003] However, existing 3D content generation methods face multiple challenges: on the one hand, high-quality 3D modeling usually relies on complex volume rendering or diffusion processes, and the computational and memory overhead far exceeds the capabilities of mobile devices, forcing the system to adopt an "edge acquisition - cloud generation" architecture; on the other hand, traditional wireless transmission solutions require uploading high-resolution multi-view images or original features, resulting in severe bandwidth pressure and latency.

[0004] While semantic communication can reduce transmission load by extracting task-relevant features, current solutions are mostly geared towards 2D images or text, making it difficult to preserve the fine geometric and texture priors required for 3D reconstruction. Furthermore, mainstream attention-based coding architectures have quadratic computational complexity, making it difficult to support real-time inference at the edge, and lack effective adaptive mechanisms for dynamic wireless channels, resulting in a significant deterioration in 3D reconstruction quality under low signal-to-noise ratio or limited bandwidth conditions. Summary of the Invention

[0005] This application aims to at least partially address one of the technical problems in the related art.

[0006] Therefore, one objective of this application is to propose a generative semantic communication system for 3D content generation. This system significantly reduces wireless bandwidth requirements by transmitting compact visual semantic tokens instead of raw data. It employs an efficient adaptive transmission architecture based on structured state-space duality, improving parallel efficiency while maintaining linear computational complexity, and integrates a channel adaptive state adjustment module to dynamically adapt to time-varying channels. Combined with a stream-matching generation model, the system can robustly reconstruct high-quality 3D content from sparse semantic tokens. Under conditions of low signal-to-noise ratio and limited bandwidth, it exhibits significant advantages over existing solutions in terms of generation quality, inference latency, and energy consumption.

[0007] To achieve the above objectives, a first aspect of this application proposes a generative semantic communication system for 3D content generation, comprising: a transmitter deployed on an edge device, a receiver deployed on a cloud server, and a wireless channel connecting the transmitter and the receiver; the transmitter is used to extract semantic information from an input image and encode it into a transmission sequence; the receiver is used to decode the semantic information from the received signal and reconstruct a 3D mesh model; the transmitter includes a semantic extractor and a joint semantic channel encoder; the semantic extractor is used to extract visual semantic tokens from a single input 2D image to form a visual semantic token sequence; the joint semantic channel encoder is used to take the visual semantic token sequence and the signal-to-noise ratio of the current wireless channel as input and map them into a channel input symbol sequence for transmission; the receiver includes a joint semantic channel decoder and a generative reconstructor; the joint semantic channel decoder is used to recover the semantic token sequence from the received signal; the generative reconstructor is a generator based on a flow matching model, which generates a 3D mesh model based on the recovered semantic token sequence.

[0008] This application discloses a generative semantic communication system for 3D content generation. By transmitting compact visual semantic tokens instead of raw data, it significantly reduces wireless bandwidth requirements. It employs a highly efficient adaptive transmission architecture based on structured state-space duality, improving parallel efficiency while maintaining linear computational complexity, and integrates a channel adaptive state adjustment module to dynamically adapt to time-varying channels. Combined with a stream-matching generation model, the system can robustly reconstruct high-quality 3D content from sparse semantic tokens. Under conditions of low signal-to-noise ratio and limited bandwidth, it exhibits significant advantages over existing solutions in terms of generation quality, inference latency, and energy consumption.

[0009] In addition, the generative semantic communication system for 3D content generation proposed in this application may also have the following additional technical features:

[0010] In one embodiment of this application, the semantic extractor is based on a self-supervised visual model. The steps of forming a visual semantic token sequence by the semantic extractor are as follows: dividing the single two-dimensional image into multiple non-overlapping patches, performing linear projection on each patch and adding position embedding to form an initial token sequence; inputting the initial token sequence into the Transformer encoder of the self-supervised visual model for processing; and using a binary mask to filter the processed token sequence to obtain the visual semantic token sequence containing only the key semantic features of the foreground object.

[0011] In one embodiment of this application, the joint semantic channel encoder employs an efficient adaptive transmission architecture based on structured state-space duality, comprising: an SSD modeling module for transforming the input visual semantic token sequence into an output sequence through matrix multiplication of semi-separable matrices; a block-parallel scanning module for dividing the input sequence into multiple local blocks, performing parallel matrix multiplication within each block, and updating the state between blocks through a discretized state-space model recursive equation; and a channel adaptive state adjustment module for generating a state adjustment bias through a lightweight MLP network based on the real-time channel signal-to-noise ratio and injecting the bias into the generation process of the time-scale parameters of the state-space model, thereby dynamically adjusting the state update rate of the model to adapt to the time-varying channel.

[0012] In one embodiment of this application, the channel adaptive state adjustment module adjusts the time scale parameter using the following formula: ;in, These are the time-scale parameters of the state-space model; It is a lightweight MLP network; For activation functions; For learnable parameter terms, Represents linear projection layer operation. For the first Input characteristics at any given time.

[0013] In one embodiment of this application, the generative reconstructor is built based on a flow matching generation model and includes: a velocity field network for receiving the recovered semantic token sequence as conditional input and injecting conditional information into the generation process through a cross-attention mechanism; an ordinary differential equation solving module for sampling initial latent variables from the prior noise distribution and defining an evolution equation according to the velocity field network: obtaining the target latent variable through numerical integration; and a 3D rendering module for decoding the target latent variable and rendering it into a final three-dimensional mesh model.

[0014] In one embodiment of this application, the optimization objective of the system is to minimize the following expected weighted total loss: ;in, This represents the optimized set of parameters. and These represent the trainable parameters at the sending and receiving ends, respectively. Represents the mathematical expectation. The input image is sampled from the real data distribution. With the corresponding 3D mesh model true value , This represents the signal-to-noise ratio sampled from the channel signal-to-noise ratio distribution. The mean squared error loss for semantic tokens. For the generative reconstruction loss based on flow matching, The weighting hyperparameters are used to balance the two losses.

[0015] In one embodiment of this application, the wireless channel is modeled as a Rayleigh fading channel, and the received signal (Y) is represented as: ;in, The channel gain vector. This is additive Gaussian noise, and ⊙ represents element-wise multiplication.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0018] Figure 1 This is a schematic diagram of the data flow of a generative semantic communication system for 3D content generation according to this application;

[0019] Figure 2 This is a schematic diagram of the structure of a generative semantic communication system for 3D content generation according to this application. Detailed Implementation

[0020] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0021] The following description, in conjunction with the accompanying drawings, describes a generative semantic communication system for 3D content generation according to an embodiment of this application.

[0022] This application provides a generative semantic communication system for 3D content generation, which can be applied to...

[0023] like Figures 1-2 As shown in the figure, a generative semantic communication system for 3D content generation according to an embodiment of this application may include a transmitter deployed on an edge device, a receiver deployed on a cloud server, and a wireless channel connecting the transmitter and the receiver.

[0024] The transmitting end is used to extract semantic information from the input image and encode it into a transmission sequence;

[0025] The receiving end is used to decode semantic information from the received signal and reconstruct a three-dimensional mesh model;

[0026] The transmitting end includes a semantic extractor and a joint semantic channel encoder;

[0027] The semantic extractor is used to extract visual semantic tokens from a single input two-dimensional image to form a sequence of visual semantic tokens;

[0028] It should be noted that the steps described in the above embodiments for extracting visual semantic tokens to form a visual semantic token sequence are as follows:

[0029] First, a pre-trained semantic extractor is used. From the input image We extract key geometric and texture features. We represent the extracted features as discrete sequences of visual semantic tokens: ;

[0030] It should be noted that, Indicates the input image It is a real number tensor (i.e., a multidimensional array of real numbers), where H represents the height of the image, i.e., the number of pixels in the vertical direction; W represents the width of the image, i.e., the number of pixels in the horizontal direction; 3 indicates that the image has 3 color channels, usually corresponding to red, green, and blue, i.e., an RGB color image;

[0031] Where T is the token matrix, L represents the length of the token sequence, and D represents the dimension of the feature vector of each token.

[0032] The extraction process aims to filter out background redundancy and retain only the key semantic features of the foreground object, thereby achieving a high degree of sparsity and compactness of T.

[0033] The joint semantic channel encoder is used to take the visual semantic token sequence and the signal-to-noise ratio (SNR) of the current wireless channel as inputs and map them into a channel input symbol sequence for transmission; it should be noted that the joint semantic channel encoder is used to efficiently transmit the semantic token sequence.

[0034] The joint semantic-channel encoder is denoted as... , where θ is the set of trainable parameters for the encoder. The encoder directly maps the semantic token T to a sequence of channel input symbols in the complex domain. :

[0035] ;

[0036] S is the transmitted symbol vector, K is the total number of transmitted complex symbols, and γ represents the signal-to-noise ratio (SNR) scalar of the current channel. To enhance the robustness of the system under different channel conditions, a joint semantic channel encoder is used. It is designed to use γ as a conditional input, thereby enabling dynamic adjustment of the encoding strategy.

[0037] To meet hardware power constraints, the transmitted symbol sequence S must satisfy the average power constraint. We define bandwidth compression ratio (BCR) as the ratio of the number of transmitted symbols to the number of pixels in the original image, i.e.

[0038] in, The total number of transmitted symbols; For mathematical expectation; The square norm of the transmitted symbol sequence S, The normalized average transmit power constant; This refers to the bandwidth compression ratio. This represents the total number of pixels in the original image.

[0039] It should be noted that the wireless channel is modeled as a Rayleigh fading channel, and the received signal... Represented as:

[0040] ;

[0041] in, The channel gain vector. This is additive Gaussian noise, and ⊙ represents element-wise multiplication.

[0042] Its elements Conforms to independent distribution , Its weight obey ;

[0043] in, It represents the field of complex numbers, meaning that each element in the vector is a complex number; Let be a complex Gaussian random variable with a mean of 0 and a variance of . ;

[0044] vector Each of them ∈ This represents the channel gain experienced by the k-th sub-channel (or the k-th transmitted symbol).

[0045] The receiver includes a joint semantic channel decoder and a generative reconstructor;

[0046] The joint semantic channel decoder is used to recover the semantic token sequence from the received signal; the joint semantic channel decoder Recover the semantic token sequence from the received signal Y. The formula is: ,in, These are decoder parameters; The signal-to-noise ratio of the wireless channel.

[0047] The generative reconstructor is a generator based on a flow matching model. It first infers three-dimensional latent features based on the recovered semantic token sequence, and then reconstructs the three-dimensional latent features into a three-dimensional mesh model through its internally integrated mesh decoder.

[0048] It should be noted that due to the effects of channel noise and fading, the recovered semantic token sequence May contain distortion; in order to reconstruct high-quality 3D content from imperfect semantic cues, the generative reconstructor is used. Utilizing pre-trained generative priors, to Given the conditions, the three-dimensional mesh model is inferred: .

[0049] To achieve end-to-end semantic consistency and geometric fidelity, we model the system optimization objective as minimizing expected distortion. Definition and Joint semantic signal encoder and joint semantic signal decoder The set of trainable parameters. The system's total loss function. Due to semantic consistency loss and generative reconstruction loss constitute.

[0050] First, to ensure the integrity of semantic information during transmission, we use mean squared error (MSE) as a measure of semantic consistency:

[0051]

[0052] in Describing the Frobenius norm, and These are the semantic tokens output by the semantic extractor and the semantic tokens recovered through decoding, respectively. The length of the semantic token sequence; The feature dimension for each semantic token.

[0053] To guide the generative model in outputting high-fidelity 3D content, we introduce a stream-matching-based reconstruction loss. This measures the degree of deviation between the generated trajectory and the actual data distribution.

[0054] In one embodiment of this application, the semantic extractor is based on a self-supervised visual model, and the step of forming a visual semantic token sequence through the semantic extractor is as follows:

[0055] The single two-dimensional image is divided into multiple non-overlapping patches. Linear projection is performed on each patch and position embedding is added to form an initial token sequence.

[0056] The initial token sequence is input into the Transformer encoder of the self-supervised visual model for processing;

[0057] The processed token sequence is filtered using a binary mask to obtain the visual semantic token sequence containing only the key semantic features of the foreground object.

[0058] Specifically, firstly, for the input image The extractor first divides it into There are 10 non-overlapping patches of size PXP, among which... Subsequently, a linear projection layer is used to map each patch to a D-dimensional embedding vector and learnable positional embeddings are added. To preserve spatial structure information: ;in, This indicates a patch embedding operation; Indicates positional encoding;

[0059] Initial token sequence Then it is processed by the Lvit layer Transformer encoder. Layer transport

[0060] out The calculation is as follows: Where MSA(·) represents multi-head self-attention operation, FFN(·) represents feedforward neural network operation, and LN(·) represents layer normalization operation;

[0061] Subsequently, in order to focus on the target object and eliminate background redundancy, a binary mask is introduced. (Obtained by the cue generator). The final output of the extractor is a masked sequence of visual semantic tokens, T:

[0062] ;in, (·) denotes a mask-based token selection operation, L≤ The number of foreground tokens to be reserved. This sequence T will serve as the semantic feature input for the efficient adaptive transport architecture based on SSD.

[0063] In one embodiment of this application, the joint semantic channel encoder employs an efficient adaptive transmission architecture based on structured state-space duality, comprising:

[0064] The SSD modeling module is used to transform the input visual semantic token sequence into an output sequence through matrix multiplication of semi-separable matrices.

[0065] It should be noted that the core of the efficient adaptive transmission architecture based on structured state space duality lies in transforming traditional sequence modeling into matrix transformation of linear operators.

[0066] For the input sequence and output sequence The recursive process of the state-space model (SSM) is reconstructed into a single matrix multiplication: ; where, transformation matrix It is not an arbitrary matrix, but a lower triangular semi-separable matrix;

[0067] According to SSD theory, the element in the j-th row and i-th column of matrix M (when i≤j) is explicitly defined by the state parameters: , here This represents the cumulative effect of the state transition matrix between time steps i and j;

[0068] In this architecture, to achieve efficient computation, The constraints are diagonal in structure, and the cumulative term is defined as follows:

[0069] It should be noted that this structure reveals the duality between SSM and the attention mechanism: In effect, it functions similarly to an Attention Map, controlling the first... The input token pair The influence weight of each output token.

[0070] The block-parallel scanning module is used to divide the input sequence into multiple local blocks, perform parallel matrix multiplication within the blocks, and update the state between blocks using a discretized state-space model recursive equation.

[0071] Specifically, the visual semantic token sequence of length L is divided into several local blocks of length Q. The computation process consists of intra-block computation and inter-block update, where:

[0072] Intra-block computation: Within each block, utilizing The matrix form can be directly subjected to parallel matrix multiplication (fully utilizing GPU Tensor Cores), with a computational complexity of O(n log n). .

[0073] Inter-block updates: Maintain a streamlined recursive state only at block boundaries. Its update follows the discretized SSM equations: Among them, the parameters of the continuous system Discretization is performed using the zero-order preserve (ZOH) method, which is controlled by the time scale parameter ∆t:

[0074] ; ;

[0075] in, Let be the state vector at time t, used to capture historical information of the sequence; Let be the input vector at time t; This is the state transition matrix, controlling the historical states. To the current state The transmission method; The input matrix controls the current input. State The degree of contribution;

[0076] This hybrid computing model maintains While reducing linear inference complexity, it significantly improves parallelism and throughput during the training phase, thus enabling efficient processing of large-scale visual semantic token sequences.

[0077] The channel adaptive state adjustment module is used to generate a state adjustment bias through a lightweight MLP network based on the real-time channel signal-to-noise ratio and inject the bias into the generation process of the time-scale parameters of the state-space model, so as to dynamically adjust the state update rate of the model to adapt to the time-varying channel.

[0078] It should be noted that in wireless semantic communication, channel state information (CSI) directly determines the required coding strategy. Traditional feature scaling methods only adjust the feature amplitude and cannot change the time scale characteristics of the model's information processing. To address this, we propose a channel adaptive state adjustment (channel adaptive state adjustment module) mechanism, which directly modulates the time step ∆t of the state-space model (SSM) through the signal-to-noise ratio (SNR).

[0079] It should be noted that the lightweight MLP network in the above embodiments maps the normalized signal-to-noise ratio γ to a state adjustment bias. (K is the number of heads);

[0080] This bias term is directly injected into the generation process of the time-scale parameter ∆ of the SSM.

[0081] ;

[0082] in, These are the time-scale parameters of the state-space model; Adjust the bias for channel state; It is a lightweight MLP network; Here is the learnable weight matrix of the MLP, ( For the first layer of weights, (for the second layer weights) For Sigmoid weighted linear units; These are the time-scale parameters of the state-space model. This is the activation function, used to ensure that the time scale parameter is positive (generally defined as...). ), For learnable parameter terms, Represents linear projection layer operation. For the first Input characteristics at any given time;

[0083] parameter State under control arrive The update rate.

[0084] When the channel quality is poor (low SNR), the channel adaptive module adaptively adjusts... This allows the model to have a longer "memory range" by using historical states to smooth out the current noisy input;

[0085] When the channel quality is good (high SNR), the channel adaptation module dynamically adjusts to a faster response mode to capture fine local features.

[0086] This mechanism enables the transmission architecture to fundamentally adapt to time-varying channels at the state evolution level without having to retrain the entire network.

[0087] In one embodiment of this application, the generative reconstructor is built based on a flow matching generation model and includes: a velocity field network for receiving the recovered semantic token sequence as conditional input and injecting conditional information into the generation process through a cross-attention mechanism;

[0088] The ordinary differential equation solving module is used to sample initial latent variables from the prior noise distribution and define evolution equations based on the velocity field network:

[0089] The target latent variables are obtained by numerical integration.

[0090] The 3D rendering module is used to decode the target latent variables and render them into the final three-dimensional mesh model.

[0091] Specifically, generative reconstructor Semantic tokens recovered from the joint semantic channel decoder Generate high-quality 3D content. To address the information incompleteness in the inverse problem, a stream-matching generation paradigm is adopted.

[0092] The reconstruction process is modeled as a result of prior noise distribution. To the target 3D latent variable distribution A continuous probability flow. Define the time variable. 3D latent variables The evolution is described by ordinary differential equations (ODEs): ,in, This is a parameterized velocity field network (typically a Transformer architecture). The network uses a cross-attention mechanism to process the received semantic tokens. As a conditional injection generation process:

[0093] ;

[0094] in, The output features of the cross-attention mechanism; This is the normalization function;

[0095] For querying the matrix, use 3D latent variables The set of query vectors obtained through mapping;

[0096] The key matrix is ​​composed of recovered semantic tokens. The set of key vectors obtained by mapping;

[0097] The value matrix is ​​composed of the recovered semantic tokens. The set of value vectors obtained by mapping;

[0098] During the inference phase, the system extracts Gaussian noise. Starting from this point, we solve the above ODE using numerical integration to obtain... Finally, it is rendered into a 3D mesh by the decoder. This process leverages the powerful priors of the generative model, enabling it to infer missing geometric texture details from sparse tokens.

[0099] In one embodiment of this application, the optimization objective of the system is to minimize the following expected weighted total loss:

[0100] ;

[0101] in, Represents the optimized set of parameters; and These represent the trainable parameters of the sender (semantic extractor and encoder) and receiver (decoder and generative reconstructor), respectively. Represents the mathematical expectation. The input image is sampled from the real data distribution. With the corresponding 3D mesh model true value , This represents the signal-to-noise ratio sampled from the channel signal-to-noise ratio distribution. The mean squared error loss for semantic tokens. For the generative reconstruction loss based on flow matching, To balance the weight hyperparameters of the two losses, .

[0102] In summary, the generative semantic communication system for 3D content generation proposed in this application significantly reduces wireless bandwidth requirements by transmitting compact visual semantic tokens instead of raw data. It employs a highly efficient adaptive transmission architecture based on structured state-space duality, improving parallel efficiency while maintaining linear computational complexity, and integrates a channel adaptive state adjustment module to dynamically adapt to time-varying channels. Combined with a stream-matching generation model, the system can robustly reconstruct high-quality 3D content from sparse semantic tokens. Under conditions of low signal-to-noise ratio and limited bandwidth, it exhibits significant advantages over existing solutions in terms of generation quality, inference latency, and energy consumption.

[0103] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0104] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0105] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A generative semantic communication system for 3D content generation, characterized in that, include: A transmitter deployed on an edge device, a receiver deployed on a cloud server, and a wireless channel connecting the transmitter and the receiver; The transmitting end is used to extract semantic information from the input image and encode it into a transmission sequence; The receiving end is used to decode semantic information from the received signal and reconstruct a three-dimensional mesh model; The transmitting end includes a semantic extractor and a joint semantic channel encoder; The semantic extractor is used to extract visual semantic tokens from a single input two-dimensional image to form a sequence of visual semantic tokens; The semantic extractor is based on a self-supervised visual model. The process of forming a visual semantic token sequence using the semantic extractor is as follows: The single two-dimensional image is divided into multiple non-overlapping patches. Linear projection is performed on each patch and position embedding is added to form an initial token sequence. The initial token sequence is input into the Transformer encoder of the self-supervised visual model for processing; The processed token sequence is filtered using a binary mask to obtain the visual semantic token sequence containing only the key semantic features of the foreground object; The joint semantic channel encoder is used to take the visual semantic token sequence and the signal-to-noise ratio of the current wireless channel as input, and map them into a channel input symbol sequence for transmission. The receiver includes a joint semantic channel decoder and a generative reconstructor; The joint semantic channel decoder is used to recover the semantic token sequence from the received signal; The generative reconstructor is built based on a flow matching generation model and generates a three-dimensional mesh model based on the recovered semantic token sequence.

2. The generative semantic communication system for 3D content generation according to claim 1, characterized in that, The joint semantic channel encoder employs an efficient adaptive transmission architecture based on structured state-space duality, including: The SSD modeling module is used to transform the input visual semantic token sequence into an output sequence through matrix multiplication of semi-separable matrices. The block-parallel scanning module is used to divide the input sequence into multiple local blocks, perform parallel matrix multiplication within the blocks, and update the state between blocks using a discretized state-space model recursive equation. The channel adaptive state adjustment module is used to generate a state adjustment bias through a lightweight MLP network based on the real-time channel signal-to-noise ratio and inject the bias into the generation process of the time-scale parameters of the state-space model, so as to dynamically adjust the state update rate of the model to adapt to the time-varying channel.

3. A generative semantic communication system for 3D content generation according to claim 2, characterized in that, The channel adaptive state adjustment module adjusts the time scale parameter using the following formula: ； in, ; These are the time-scale parameters of the state-space model; It is a lightweight MLP network; For activation functions; For learnable parameter terms, Represents linear projection layer operation. For the first Input characteristics at any given time.

4. A generative semantic communication system for 3D content generation according to claim 1, characterized in that, The generative reconstructor is built based on a flow matching generation model and includes: The velocity field network is used to receive the recovered semantic token sequence as conditional input and inject conditional information into the generation process through a cross-attention mechanism; The ordinary differential equation solving module is used to sample initial latent variables from the prior noise distribution and define evolution equations based on the velocity field network: the target latent variables are obtained by numerical integration. The 3D rendering module is used to decode the target latent variables and render them into the final three-dimensional mesh model.

5. A generative semantic communication system for 3D content generation according to claim 1, characterized in that, The optimization objective of the system is to minimize the following expected weighted total loss: ； in, Represents the optimized set of parameters. and These represent the trainable parameters at the sending and receiving ends, respectively. Represents the mathematical expectation. The input image is sampled from the real data distribution. With the corresponding 3D mesh model true value , This represents the signal-to-noise ratio sampled from the channel signal-to-noise ratio distribution. The mean squared error loss for semantic tokens. For the generative reconstruction loss based on flow matching, The weighting hyperparameters are used to balance the two losses.

6. A generative semantic communication system for 3D content generation according to claim 1, characterized in that, The wireless channel is modeled as a Rayleigh fading channel, and the received signal (Y) is represented as: ； in, The channel gain vector. This is additive Gaussian noise, and ⊙ represents element-wise multiplication.

Images