Platform for generating AIGC data by converting 2D (two-dimensional) to 3D (three-dimensional) graphs based on AI (artificial intelligence) computing power
By constructing a computing power demand prediction module and a dynamic scheduling module, and combining edge nodes and cloud clusters, the resource waste and supply-demand imbalance problems of the 2D to 3D graphics generation platform were solved, achieving efficient and economical computing power resource management and generation quality assurance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU DASSAULT INTELLIGENT ENERGY TECHNOLOGY CO LTD
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-19
AI Technical Summary
The existing AIGC data platform for 2D to 3D graphics generation has a fixed computing power allocation method, which leads to low resource utilization, limited generation efficiency and energy waste, and cannot meet the different computing power requirements of different generation stages.
The system constructs a computing power demand prediction module, a dynamic computing power scheduling module, a computing power cache reuse module, and an energy consumption quality balancing module. Through a two-stage prediction model and reinforcement learning, it achieves dynamic adaptation of computing power resources and energy consumption control. Combined with the flexible scheduling of edge nodes and cloud clusters, it reduces redundant calculations and optimizes resource utilization.
It improves resource utilization, optimizes generation efficiency, reduces energy consumption costs, ensures the adaptability of generation quality and efficiency across multiple scenarios, and realizes on-demand allocation and dynamic adaptation of computing resources.
Smart Images
Figure CN122064488A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of 2D to 3D technology, specifically to an AIGC data platform for generating 2D to 3D graphics based on AI computing power. Background Technology
[0002] With the rapid development of AIGC technology, 2D to 3D graphics generation platforms based on AI computing power have been widely used in game development, industrial design, film and television production and other fields. The core requirement of such platforms is to quickly convert 2D images into high-quality 3D models through AI algorithms, and the rationality of computing power allocation directly affects the generation efficiency, model quality and resource utilization.
[0003] Existing AIGC data platforms for 2D to 3D graphics generation mostly use a fixed configuration for computing power allocation, that is, pre-setting computing power resources according to task type. For example, high-precision generation tasks are fixedly allocated 8 GPUs. This allocation method does not take into account the different computing power requirements of each stage in the 2D to 3D generation process. For example, the pre-processing of materials and format conversion stages only require low computing power support, while the 3D geometry generation stage is the peak stage of computing power demand, which requires the distributed computing support of high-performance GPUs. The fixed computing power allocation mode leads to the overabundance of computing power in stages with low computing power demand and the underabundance of computing power in stages with high computing power demand. This results in problems such as low resource utilization, limited generation efficiency, and energy waste, which seriously affect the practicality and economy of the platform. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an AIGC data platform for 2D-to-3D graphics generation based on AI computing power, which solves the problems mentioned in the background technology.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: an AIGC data platform for 2D to 3D graphics generation based on AI computing power, including a computing power demand prediction module, a dynamic computing power scheduling module, a computing power cache reuse module, an energy consumption quality balance module, and a real-time monitoring and adjustment module; The computing power demand prediction module is used to construct a two-stage prediction model and output the computing power demand values for each generation stage of 2D to 3D conversion. The dynamic computing power scheduling module achieves accurate allocation of computing power for each generation stage based on the computing power demand values and the real-time computing power resource status of the platform. The computing power cache reuse module caches and reuses intermediate features for high-frequency similar tasks to reduce redundant calculations. The energy consumption and quality balancing module adjusts the computing power configuration according to user scenario requirements to balance generation quality and energy consumption. The real-time monitoring and adjustment module dynamically corrects the computing power allocation scheme by collecting multi-dimensional operating indicators. The above modules work together to achieve dynamic adaptation of computing power resources and generation stages, thereby improving resource utilization.
[0006] Furthermore, the dual-stage prediction model includes an initial prediction submodule and a dynamic correction prediction submodule. The initial prediction submodule uses an LSTM network with an attention mechanism, inputting standardized input material feature vectors and outputting the initial computing power requirements for each generation stage. The dynamic correction prediction submodule uses a GRU network to collect the computing power utilization, memory usage, and generation progress of the current generation stage in real time, and dynamically corrects the initial computing power requirements using the following formula: The initial computing power requirement is dynamically adjusted, whereby... The correction factor is between 0.1 and 0.3. The output is a correction factor between -0.2 and 0.2; This represents the current computing power utilization rate. This represents the current video memory usage. This represents the current generation progress.
[0007] Furthermore, the input material feature vector includes image resolution normalization value, edge density, texture complexity, and one-hot encoding of the input type; wherein the image resolution normalization value is calculated using the following formula: in, ; .
[0008] Furthermore, the aforementioned This represents the largest unit of resource that can be allocated between the two-tiered computing resource pools. This value can meet the peak computing power requirements for 3D graphics generation in ultra-high-definition and complex scenes, ensuring smooth operation in large projects or high-concurrency tasks. With the smallest resource allocation granularity, it aims to ensure the computing power supply for basic graphics processing tasks while avoiding excessive resource fragmentation. Combined with a two-stage prediction model to accurately output computing power requirements, it enables the cache reuse mechanism and the balancing adjustment model to operate more efficiently, improving resource utilization while taking into account generation efficiency and energy consumption costs.
[0009] Furthermore, the edge density is calculated using the following formula: in, Number of pixels at the edge; This represents the total number of pixels.
[0010] Furthermore, the texture complexity is calculated using the following formula: in, These are elements of the gray-level co-occurrence matrix.
[0011] Furthermore, the dynamic computing power scheduling module constructs a two-tiered computing power resource pool consisting of edge nodes and a cloud cluster. Lightweight GPUs are deployed at the edge nodes for low-computing-power-demand tasks such as material pre-processing and format conversion. High-performance GPUs are deployed in the cloud cluster, supporting multi-GPU distributed computing for high-peak-power-demand tasks such as 3D geometry generation. The dynamic computing power scheduling module collects real-time data on the load, remaining video memory, and task queue length of each node through a resource status matrix, and allocates corresponding computing power resources according to the computing power requirements of each generation stage. The number of GPUs required for the 3D generation stage is calculated using the following formula: in The total computing power required to generate the corrected graphics; This represents the computing power of a single GPU; the A100 single-card computing power is set to 100 TFLOPS.
[0012] Furthermore, the computing power cache reuse module performs similarity calculation and cache determination through the following steps: First, extract feature vectors. Specifically, for new tasks, extract the key features of their input data and construct multi-dimensional feature vectors; for historical tasks, retrieve the stored feature vectors from the cache records. These features cover image color distribution, texture structure, and object shape and contour information. Next, cosine similarity is calculated using the cosine similarity formula: in, For the new task feature vector; For historical task feature vectors; It is the dot product of two vectors; These are the magnitudes of the vectors; this formula quantifies the similarity between the feature vectors of two tasks, outputting a similarity value between -1 and 1, with the value closer to 1 indicating a higher similarity. The calculated similarity value is compared with a preset threshold of 0.85. When the similarity is not lower than 0.85, it is determined that the new task and the historical task have a high similarity, triggering the caching mechanism, reading the cached intermediate features, and skipping the repeated calculation step. The intermediate features include the semantic segmentation map and depth map output from the material preprocessing, as well as the point cloud data and implicit field parameters from the 3D generation step. The computing power cache reuse module is set to a 72-hour cache validity period and uses the least recent use principle to clean up excess cache.
[0013] Furthermore, the energy consumption quality balancing module trains a balancing adjustment model through reinforcement learning, taking the system state as input and the computing power adjustment amount as output. The system state includes, but is not limited to, the current computing power allocation and task load. The computing power configuration is optimized based on a reward function, the formula of which is as follows: in, For quality weights; Energy consumption weighting; To generate a quality score, ; For energy consumption, This represents the energy consumption value at full computing power. For computing power utilization; This function balances the generation quality, energy consumption, and computing power utilization, incentivizing the model to find the optimal configuration. During the reinforcement learning iteration process, the model parameters are updated through temporal difference learning, as shown in the formula: ,in The learning rate; Discount factor; The updated state value; The value function of the state before the update represents the state value at time step [step]. System state Value; Indicates at time step The system status; In time step The immediate reward obtained after performing an action; It is a time step System state The value of.
[0014] Furthermore, the real-time monitoring and adjustment module collects the computing power utilization, generation quality score, execution time, energy consumption, and video memory usage of each generation stage once per second; when the computing power utilization is below 0.4 and remains below 0.4 for 3 seconds, it adjusts accordingly. Reduce computing power The reduction ratio is 0.1-0.2; when the computing power utilization rate is higher than 0.85 and lasts for 3 seconds, it is reduced by... Additional computing power, An additional ratio of 0.2-0.3 is applied; the generated quality score is calculated by weighting three dimensions: geometric accuracy, texture consistency, and topological rationality, using the following formula: ,in Scoring for geometric accuracy; Scoring for texture consistency; Score the topological rationality.
[0015] This invention provides an AIGC data platform for generating 2D-to-3D graphics based on AI computing power, which has the following beneficial effects: 1. This AIGC data platform for 2D-to-3D graphics generation, based on AI computing power, completely solves the problems of resource waste and supply-demand imbalance in the traditional fixed computing power allocation model by constructing a dynamic computing power scheduling system with multi-module linkage and cooperation. Relying on a two-stage prediction model, it accurately captures the differences in computing power requirements of each generation stage. Combined with the flexible scheduling of two-level resource pools of edge nodes and cloud clusters, it realizes on-demand allocation and dynamic adaptation of computing power resources, avoiding the overabundance of computing power in low computing power demand stages and the underabundance of computing power in high computing power demand stages. At the same time, the computing power caching and reuse mechanism reuses intermediate features for high-frequency similar tasks, greatly reducing the ineffective consumption caused by repeated calculations and further improving resource turnover efficiency. The real-time monitoring and dynamic correction closed loop of the whole process ensures that the computing power configuration always matches the task execution status, which qualitatively improves the utilization rate of platform computing power resources and significantly optimizes the overall efficiency of the generation process.
[0016] 2. This AIGC data platform for 2D to 3D graphics generation, based on AI computing power, achieves precise control over generation quality and energy consumption costs through a reinforcement learning-driven energy-quality balance model, meeting the differentiated needs of various scenarios. For high-precision scenarios such as industrial design and film and television production, it ensures high standards of generation quality by optimizing computing power configuration. For efficiency-priority scenarios such as rapid prototyping and sketch preview, it reduces energy consumption while meeting basic quality requirements, balancing practicality and economy. The generation quality scoring system ensures output quality from multiple dimensions, including geometric accuracy, texture consistency, and topological rationality. Combined with a real-time monitoring and adjustment mechanism, it can promptly correct quality defects and ensure the reliability of the generation results. Attached Figure Description
[0017] Figure 1 This is a schematic diagram illustrating the steps of an AIGC data platform for generating 2D to 3D graphics based on AI computing power, as described in this invention. Detailed Implementation
[0018] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.
[0019] like Figure 1 As shown, the present invention provides a technical solution: an AIGC data platform for 2D to 3D graphics generation based on AI computing power, including a computing power demand prediction module, a dynamic computing power scheduling module, a computing power cache reuse module, an energy consumption quality balancing module, and a real-time monitoring and adjustment module. The modules work together to achieve dynamic optimization allocation of computing power resources. Computing power demand prediction module: Constructs a two-stage prediction model with initial prediction in the early stage and dynamic correction in the process, accurately outputting the computing power demand value of each generation stage; First, data preprocessing and feature engineering are performed, collecting 2D to 3D generation task data from the past 12 months of the platform, including input material features, execution parameters and hardware performance data of each stage, and standardizing features such as image resolution, edge density, and texture complexity to form input feature vectors; The input feature vector includes the image resolution normalization value, edge density, texture complexity, and one-hot encoding of the input type; the image resolution normalization value is calculated using the following formula: in, ; ; This represents the largest unit of resource that can be allocated between the two-tiered computing resource pools. This value can meet the peak computing power requirements for 3D graphics generation in ultra-high-definition and complex scenes, ensuring smooth operation in large projects or high-concurrency tasks. The minimum resource allocation granularity aims to ensure the computing power supply for basic graphics processing tasks while avoiding excessive resource fragmentation. Combined with a two-stage prediction model, it accurately outputs computing power requirements, enabling the cache reuse mechanism and balancing adjustment model to operate more efficiently. This improves resource utilization while taking into account both generation efficiency and energy consumption costs. Edge density is calculated using the following formula: in, Number of pixels at the edge; This represents the total number of pixels. Texture complexity is calculated using the following formula: in, Elements of the gray-level co-occurrence matrix; The initial prediction submodule uses an LSTM network with an attention mechanism. After inputting the feature vector, it outputs the initial computational requirements for four stages: material preprocessing, 3D generation, optimization, and format conversion through the collaborative operation of the input gate, forget gate, cell state, and output gate. The hidden layer state update formula of the LSTM network is as follows: in, , , These are the input gate, forget gate, and output gate, respectively. In cellular state; Output for hidden layer; It is input To the input gate The weight matrix; It is the output of the hidden layer at the previous moment. To the input gate The weight matrix; It represents the cell state at the previous moment. To the input gate The weight matrix; For input gate The bias term; It is input To the Gate of Oblivion The weight matrix; It is the output of the hidden layer at the previous moment. To the Gate of Oblivion The weight matrix; It represents the cell state at the previous moment. To the Gate of Oblivion The weight matrix; Forgotten Gate The bias term; It is input To the weight matrix used for calculating the cell state update; It is the output of the hidden layer at the previous moment. To the weight matrix used for calculating the cell state update; Bias terms calculated for cell state updates; It is input To the output gate The weight matrix; It is the output of the hidden layer at the previous moment. To the output gate The weight matrix; Cellular state To the output gate The weight matrix; For output gate Bias term; The Sigmoid activation function maps the input to between 0 and 1, and is used to calculate the gate signal; The element-wise multiplication operation multiplies corresponding elements of two matrices of the same dimension. The attention mechanism improves prediction accuracy by assigning weights based on the cosine similarity between input features and historical task features. The formula for the attention-weighted feature output is as follows: in, The cosine similarity function; This is the set of hidden layer features for historical tasks; Attention weights; The output of the attention-weighted features; The dynamic correction prediction submodule uses a GRU network to collect real-time data on computing power utilization, memory usage, and generation progress during the current generation stage, and then corrects these data using a formula. The initial computing power requirement is dynamically adjusted, whereby... The correction factor is between 0.1 and 0.3. The output is a correction factor between -0.2 and 0.2; This represents the current computing power utilization rate. This represents the current video memory usage. This represents the current generation progress. The model training uses an 8:2 ratio to divide the training and test sets. The mean squared error formula is as follows: in, The number of samples represents the total number of data samples used to calculate the loss function. Indicates the first The predicted computing power requirement for each sample is the result predicted by the model. Indicates the first The actual computing power requirements of each sample are real, known, and accurate data. Using the above formula as the loss function, the Adam optimizer was trained for 100 rounds with an initial learning rate of 0.001, which was reduced to 0.5 every 20 rounds to ensure that the prediction error on the test set was controlled within 8%. Dynamic computing power scheduling module: Based on the computing power demand value output by the two-stage prediction model, combined with the real-time computing power resource status of the platform, it executes a dynamic scheduling strategy by stage and node. First, it constructs a two-level computing power resource pool of edge nodes and cloud cluster. Lightweight GPUs such as T4 and A30 are deployed on the edge nodes, while high-performance GPUs such as A100 and H100 are deployed on the cloud cluster, supporting multi-card distributed computing. The resource monitoring module collects the load, remaining video memory, and task queue length of each node once per second to generate a real-time resource status matrix. Resources are allocated according to the computing power requirements of different production stages: Lightweight GPUs on edge nodes are used in the pre-processing stage of materials, and computing power is allocated to meet these requirements. ,in, Total number of pixels in the image; These are the operands for preprocessing. The maximum allowed time for pre-processing is 5 seconds; the 3D generation stage schedules high-performance GPUs in the cloud cluster based on the corrected computing power requirements, using a formula. Calculate the required number of cards, where The total computing power required to generate the corrected graphics; The A100 single-card computing power is set at 100 TFLOPS, while NVLink technology ensures sufficient bandwidth for inter-card data transfer. ,in This refers to the amount of data exchanged between cards. The maximum allowed communication time is 0.5 seconds; during the optimization phase, computing power is dynamically adjusted based on the generated quality score. If the score is lower than the preset threshold, 2-4 GPUs are added; if the score meets the standard, 1-2 GPUs are maintained or reduced; during the format conversion phase, idle GPUs on edge nodes or remaining computing power in the cloud cluster are utilized to allocate computing power to meet the requirements. ,in This refers to the amount of model data. This represents the coding complexity coefficient. The maximum allowed time for format conversion is 3 seconds; The scheduling decision is updated every 5 seconds based on the latest resource status matrix and generation progress. When the load on a node exceeds 0.8 and there are idle nodes, some tasks on that node are migrated to the idle nodes. For user-defined high-urgency tasks, priority is applied using the following formula: Increase scheduling priority, among which Task urgency level (leveled from 1 to 5, to be preset according to actual needs); To optimize node load, prioritize scheduling high-performance nodes; The computing power cache reuse module is designed with a cache reuse mechanism for high-frequency similar tasks to reduce redundant computations. When a new task is submitted, its cosine similarity to the input feature vectors of historical tasks is calculated. in, For the new task feature vector; For historical task feature vectors; It is the dot product of two vectors; These are the magnitudes of the vectors. This formula quantifies the similarity between the feature vectors of two tasks, outputting a similarity value between -1 and 1. The closer the value is to 1, the higher the similarity. When the similarity is not lower than 0.85, the tasks are considered similar, triggering a caching mechanism to directly read the cached intermediate features, skipping repetitive preprocessing and some 3D generation calculations. The intermediate features are cached on a high-speed NVMe hard drive, named and stored according to the task type and input feature hash value, including semantic segmentation maps, depth maps, point cloud data, implicit field parameters, etc., with a 72-hour cache validity period, automatically cleared after the validity period. Historical tasks with a generation quality score below 80 are not cached. When the platform algorithm model is updated, the intermediate feature cache generated based on the old model is deleted, while controlling the cache to occupy no more than 30% of the total storage capacity. When the threshold is reached, some cache is deleted according to the least recently used principle. Energy Consumption and Quality Balance Module: This module constructs an energy consumption and quality balance model based on user scenario requirements, providing three operating modes for users to choose from: Precision Priority Mode is suitable for scenarios such as industrial design and film and television production, with a quality threshold set at 90 points, computing power allocation configured at 1.1 times the predicted value, a quality weight of 0.7, and an energy consumption weight of 0.3; Efficiency Priority Mode is suitable for scenarios such as rapid prototyping and sketch preview, with a quality threshold set at 80 points, computing power allocation configured at 0.8 times the predicted value, an energy consumption weight of 0.7, and a quality weight of 0.3; Balanced Mode is suitable for general scenarios, with a quality threshold set at 85 points, computing power allocation configured according to the predicted value, and both quality and energy consumption weights at 0.5. Employing a reinforcement learning training balance regulation model, with system state... As input, where Computing power already used; , , These are respectively the generation of quality score, energy consumption, and real-time load; adjusted by computing power. As output, This represents the incremental computing power consumption during the preprocessing stage, which is the additional amount of computing resources consumed by the data to complete preprocessing operations such as cleaning and format conversion before entering the formal graphics generation process. This represents the incremental computational power consumption during the graphics generation stage. This stage generates 2D to 3D graphics based on the input data and is the core of the entire computation process. The computational power consumed is mainly used for operations such as graphics rendering and model calculation. This refers to the incremental computational power consumption during the optimization phase. It is the increased computational power required to perform post-processing optimization operations (such as smoothing and detail enhancement) on the generated 3D graphics in order to improve graphics quality and generation efficiency. This represents the incremental computational power consumption during the conversion phase. It refers to the additional computational power consumption caused by changes in data format and dimension during the process of converting 2D graphics data into a 3D graphics data structure. The model parameters are updated through temporal difference learning, using the following formula: ,in The learning rate; Discount factor; The updated state value; The value function of the state before the update represents the state value at time step [step]. System state Value; Indicates at time step The system status, such as the current resource allocation status and task queue status of the AI computing platform; In time step In the context of AI-powered 2D to 3D graphics generation, the immediate reward obtained after performing an action may represent the energy savings and increased efficiency gained after completing a certain number of graphics generation tasks. It is a time step System state Value; Real-time monitoring and adjustment module: Establishes a closed loop for real-time monitoring and feedback adjustment throughout the entire process. It collects multi-dimensional indicators such as computing power utilization, generation quality score, execution time, energy consumption, and memory usage at each generation stage, constructing a monitoring indicator matrix. The generation quality score is calculated by weighting three dimensions: geometric accuracy, texture consistency, and topological rationality. The formula is as follows: ,in To score geometric accuracy, the generated 3D graphic is compared with the standard 3D model corresponding to the original 2D graphic in terms of geometric features. Spatial distance measurement algorithms (such as Hausdorff distance) are used to quantify the differences between vertices, edges, and faces between the models. The smaller the difference, the higher the score. To score texture consistency, feature matching and similarity calculation are performed on the texture of 2D image and the surface texture of 3D model. Perceptual hashing algorithm or deep learning texture feature extraction network (such as VGG network) is used to extract texture feature vectors and then calculate cosine similarity. The higher the similarity, the higher the texture consistency score. To score the topological rationality, the topological properties of the model, such as vertex connectivity and face integrity, are checked based on the topological structure rules of the 3D graphics. The number and type of errors are counted by a topological error detection algorithm. The fewer the errors, the higher the topological rationality score. Set an adjustment trigger threshold; when the computing power utilization rate is below 0.4 for 3 consecutive seconds, press [the appropriate button]. Reduce computing power The reduction ratio is 0.1-0.2; when the computing power utilization rate is higher than 0.85 and lasts for 3 seconds, it is reduced by... Additional computing power, An additional addition ratio of 0.2-0.3; Every day at midnight, the previous day's scheduling data is analyzed offline. A random forest algorithm is used to build a scheduling effect evaluation model, identify the optimal scheduling strategy pattern, and update it to the online scheduling system to achieve continuous iterative optimization of the scheduling strategy. Example: A high-precision 2D to 3D generation task in an industrial design scenario; First, a two-level computing resource pool is constructed, consisting of edge nodes and a cloud cluster. Ten servers equipped with T4 GPUs are deployed at the edge nodes, and eight servers equipped with eight A100 GPUs are deployed in the cloud cluster, supporting NVLink technology to achieve high-speed communication between the cards; K8s is used as the computing power scheduling system, and NVMe high-speed storage devices are deployed as cache media, with a total storage capacity of 100TB; We collected 100,000 data entries from 2D to 3D industrial design tasks over the past 12 months on the platform, which were divided into training and testing sets in an 8:2 ratio. The input material features were standardized, with image resolution ranging from 256×256 to 8192×8192. Edge density was calculated using the Canny operator, and texture complexity was represented by the energy value of the gray-level co-occurrence matrix. An LSTM network incorporating an attention mechanism is used as the initial prediction sub-module, while a GRU network is used as the dynamic correction prediction sub-module during training. During training, the mean square error of the loss function gradually converges to 0.006, and the prediction error on the test set is 7.2%, which meets the accuracy requirements. The trained two-stage prediction model is then deployed to the platform's computing power demand prediction module, which outputs the computing power demand values for each stage based on the real-time input features of the industrial design 2D drawings. The user uploads a 1024×1024 resolution 2D drawing of an industrial part, selects the precision-priority mode, and sets the task urgency level to 4. The computing power demand prediction module outputs initial computing power requirements: 10 TFLOPS for material preprocessing, 800 TFLOPS for 3D generation, 300 TFLOPS for optimization, and 8 TFLOPS for format conversion. During the process, the dynamic correction prediction submodule collects real-time data and adjusts the computing power requirement for the 3D generation stage to 850 TFLOPS. The dynamic computing power scheduling module allocates resources based on the resource status matrix: one T4 GPU (15 TFLOPS computing power) is scheduled on the edge node for material preprocessing to meet the computing power requirement; the 3D generation stage allocates resources according to the formula... The required number of GPUs was calculated, and nine A100 GPUs from two servers in the cloud cluster were scheduled. NVLink technology was used to ensure inter-GPU communication bandwidth. During the optimization phase, since the initial quality score was 88 points, which was lower than the 90-point threshold for the precision-first mode, two more A100 GPUs were added, increasing the computing power to 500 TFLOPS. During the format conversion phase, idle T4 GPUs from edge nodes were scheduled to complete the format conversion and export. During scheduling, the resource status matrix is updated every 5 seconds. Due to the high urgency of the task, the scheduling priority is adjusted accordingly. Priority was given to ensuring the supply of computing power, and there was no shortage of computing power. Because this type of industrial part is a high-frequency processing task on the platform, the similarity between the input feature vectors of the new task and the historical task is 0.88, triggering the caching mechanism. The cached semantic segmentation map and depth map are read directly, skipping some pre-processing calculations of the materials. This step reduces the time taken from 3 seconds to 1 second. The energy consumption and quality balancing module runs in precision-first mode. The reinforcement learning balancing adjustment model dynamically adjusts the computing power according to the real-time system status. The final 3D model quality score is 92 points, which meets the precision requirements of industrial design. The energy consumption of a single task is 120kWh, which is 35% lower than the traditional static allocation mode. Throughout the entire task execution cycle, the system's computing power utilization remained within a healthy range of 85%-92%, without triggering any abnormal warnings. Only during the optimization phase, because the 3D depth score (PSNR=28.5dB) of the initially generated graphics did not meet the preset standard (≥30dB), the system automatically activated the computing power supplementation mechanism: dynamically allocating 20% of the GPU resources from the spare computing power resource pool, and redistributing the computing load in conjunction with the balancing adjustment model; after adjustment, and real-time calibration by the two-stage prediction model, the generation quality quickly reached the standard (PSNR=32.1dB) after 3 iterations. The final total task time was 160 seconds, achieving a 28% improvement in generation efficiency compared to the traditional static computing power allocation mode (222 seconds).
[0020] Based on the above description, this invention completely solves the problems of resource waste and supply-demand imbalance in the traditional fixed computing power allocation mode by constructing a dynamic computing power scheduling system with multi-module linkage and cooperation. Relying on a two-stage prediction model, it accurately captures the differences in computing power demand at each generation stage. Combined with the flexible scheduling of two-level resource pools at the edge nodes and cloud clusters, it realizes on-demand allocation and dynamic adaptation of computing power resources, avoiding the overabundance of computing power in low computing power demand stages and the underabundance of computing power in high computing power demand stages. At the same time, the computing power cache reuse mechanism reuses intermediate features for high-frequency similar tasks, greatly reducing the ineffective consumption caused by repeated calculations and further improving resource turnover efficiency. The real-time monitoring and dynamic correction closed loop of the whole process ensures that the computing power configuration always matches the task execution state, which qualitatively improves the utilization rate of platform computing power resources and significantly optimizes the overall efficiency of the generation process. Furthermore, a reinforcement learning-driven energy-quality balance model enables precise control of generated quality and energy consumption costs, meeting the differentiated needs of various scenarios. For high-precision scenarios such as industrial design and film production, high standards of generated quality are ensured by optimizing computing power configuration. For efficiency-priority scenarios such as rapid prototyping and sketch preview, energy consumption is reduced while meeting basic quality requirements, balancing practicality and economy. The generated quality scoring system ensures output quality from multiple dimensions, including geometric accuracy, texture consistency, and topological rationality. Combined with a real-time monitoring and adjustment mechanism, quality defects can be corrected in a timely manner, ensuring the reliability of the generated results. The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.
Claims
1. A 2D-to-3D graphics generation AIGC data platform based on AI computing power, characterized in that: It includes a computing power demand prediction module, a dynamic computing power scheduling module, a computing power cache reuse module, an energy consumption quality balancing module, and a real-time monitoring and adjustment module; The computing power demand prediction module is used to construct a two-stage prediction model and output the computing power demand values for each generation stage of 2D to 3D conversion. The dynamic computing power scheduling module achieves accurate allocation of computing power for each generation stage based on the computing power demand values and the real-time computing power resource status of the platform. The computing power cache reuse module caches and reuses intermediate features for high-frequency similar tasks to reduce redundant calculations. The energy consumption and quality balancing module adjusts the computing power configuration according to user scenario requirements to balance generation quality and energy consumption. The real-time monitoring and adjustment module dynamically corrects the computing power allocation scheme by collecting multi-dimensional operating indicators. The above modules work together to achieve dynamic adaptation of computing power resources and generation stages, thereby improving resource utilization.
2. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 1, characterized in that: The dual-stage prediction model includes an initial prediction submodule and a dynamic correction prediction submodule. The initial prediction submodule uses an LSTM network with an attention mechanism, taking the standardized input material feature vector as input and outputting the initial computing power requirement for each generation stage. The dynamic correction prediction submodule uses a GRU network to collect the computing power utilization, memory usage, and generation progress of the current generation stage in real time, and dynamically corrects the initial computing power requirement using the following formula: The initial computing power requirement is dynamically adjusted, whereby... The correction factor is between 0.1 and 0.
3. The output is a correction factor between -0.2 and 0.2; This represents the current computing power utilization rate. This represents the current video memory usage. This represents the current generation progress.
3. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 2, characterized in that: The input material feature vector includes image resolution normalization value, edge density, texture complexity, and one-hot encoding of the input type; wherein the image resolution normalization value is calculated using the following formula: in, ; .
4. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 3, characterized in that: The This represents the largest unit of resource that can be allocated between the two-tiered computing resource pools. This value can meet the peak computing power requirements for 3D graphics generation in ultra-high-definition and complex scenes, ensuring smooth operation in large projects or high-concurrency tasks. With the smallest resource allocation granularity, it aims to ensure the computing power supply for basic graphics processing tasks while avoiding excessive resource fragmentation. Combined with a two-stage prediction model to accurately output computing power requirements, it enables the cache reuse mechanism and the balancing adjustment model to operate more efficiently, improving resource utilization while taking into account generation efficiency and energy consumption costs.
5. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 2, characterized in that: The edge density is calculated using the following formula: in, Number of pixels at the edge; This represents the total number of pixels.
6. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 2, characterized in that: The texture complexity is calculated using the following formula: in, These are elements of the gray-level co-occurrence matrix.
7. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 1, characterized in that: The dynamic computing power scheduling module constructs a two-tier computing power resource pool consisting of edge nodes and a cloud cluster. Lightweight GPUs are deployed at the edge nodes for low-computing-power tasks such as material pre-processing and format conversion. High-performance GPUs are deployed in the cloud cluster, supporting multi-GPU distributed computing for high-peak-power tasks such as 3D geometry generation. The dynamic computing power scheduling module collects real-time data on the load, remaining video memory, and task queue length of each node through a resource status matrix, and allocates corresponding computing power resources according to the computing power requirements of each generation stage. The number of GPUs required for the 3D generation stage is calculated using the following formula: in The total computing power required to generate the corrected graphics; This represents the computing power of a single GPU; the A100 single-card computing power is set to 100 TFLOPS.
8. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 1, characterized in that: The computing power cache reuse module performs similarity calculation and cache judgment through the following steps: First, extract feature vectors. Specifically, for new tasks, extract the key features of their input data and construct multi-dimensional feature vectors; for historical tasks, retrieve the stored feature vectors from the cache records. These features cover image color distribution, texture structure, and object shape and contour information. Next, cosine similarity is calculated using the cosine similarity formula: in, For the new task feature vector; For historical task feature vectors; It is the dot product of two vectors; These are the magnitudes of the vectors; this formula quantifies the similarity between the feature vectors of two tasks, outputting a similarity value between -1 and 1, with the value closer to 1 indicating a higher similarity. The calculated similarity value is compared with a preset threshold of 0.
85. When the similarity is not lower than 0.85, it is determined that the new task and the historical task have a high similarity, triggering the caching mechanism, reading the cached intermediate features, and skipping the repeated calculation step. The intermediate features include the semantic segmentation map and depth map output from the material preprocessing, as well as the point cloud data and implicit field parameters from the 3D generation step. The computing power cache reuse module is set to a 72-hour cache validity period and uses the least recent use principle to clean up excess cache.
9. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 1, characterized in that: The energy consumption and quality balancing module trains a balancing adjustment model through reinforcement learning and provides three operating modes for users to choose from: precision-first mode, efficiency-first mode, and balanced mode. It takes system state as input and computing power adjustment as output, where system state includes, but is not limited to, the current computing power allocation and task load. It optimizes computing power configuration based on a reward function, the formula of which is as follows: in, For quality weights; Energy consumption weighting; To generate a quality score, ; For energy consumption, This represents the energy consumption value at full computing power. For computing power utilization; This function balances the generation quality, energy consumption, and computing power utilization, incentivizing the model to find the optimal configuration. The generated quality score is calculated by weighting three dimensions: geometric accuracy, texture consistency, and topological rationality. The formula is as follows: ,in Scoring for geometric accuracy; Scoring for texture consistency; Score the topological rationality; During the reinforcement learning iteration process, the model parameters are updated through temporal difference learning, as shown in the formula: ,in The learning rate; Discount factor; The updated state value; The value function of the state before the update represents the state value at time step [step]. System state Value; Indicates at time step The system status; In time step The immediate reward obtained after performing an action; It is a time step System state The value of.
10. The AIGC data platform for 2D-to-3D graphics generation based on AI computing power according to claim 1, characterized in that: The real-time monitoring and adjustment module collects the computing power utilization, generation quality score, execution time, energy consumption, and video memory usage of each generation stage once per second; when the computing power utilization is below 0.4 and remains below 0.4 for 3 seconds, it adjusts accordingly. Reduce computing power The reduction ratio is 0.1-0.2; when the computing power utilization rate is higher than 0.85 and lasts for 3 seconds, it is reduced by... Additional computing power, The additional ratio is 0.2-0.3.