A multi-agent collaboration-based three-dimensional Gaussian code generation system based on multi-dimensional feedback

Abstract

The application provides a three-dimensional Gaussian code generation system based on multi-agent cooperation of multi-dimensional feedback, integrates visual quality verification, performance analysis and static code analysis into a dynamic optimization cycle, and uses multi-agent cooperation for 3D Gaussian splat code optimization; meanwhile, cross-task optimization migration and automatic error detection and repair are carried out by using a knowledge base; compared with existing manual optimization or single-agent optimization solutions, the application uses multi-dimensional feedback integration (visual quality + performance analysis + static analysis) to enter the optimization process, can be dynamically queried to the optimization experience knowledge base, corresponds to the reconstructed scene, and thus provides higher performance portability and adaptability, and can efficiently and stably generate and optimize Gaussian splat codes for complex 3D scenes, such as large 3D reconstruction, real-time rendering application and dynamic scene.

Description

Technical Field

[0001] This invention belongs to the fields of large language models, artificial intelligence, and neural rendering technology, and particularly relates to a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback. Background Technology

[0002] Existing methods for generating 3D Gaussian code typically rely on specific acquisition devices to capture multi-view sequences of images of a scene or object, combine this with corresponding pose information, train a 3D Gaussian neural representation of the scene or object, and then render it. Alternatively, they employ a training-free multi-agent workflow specifically designed for CUDA kernel generation and optimization. This workflow is inspired by iterative optimization processes and can incorporate hardware feedback for code generation. Another approach is to achieve Gaussian code generation by optimizing the GEMM (Generalized Matrix Multiplication) kernel.

[0003] Currently, existing technologies have not yet achieved an automated solution for generating 3D Gaussian code. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a 3D Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback, which automates the generation of 3D Gaussian code, reduces labor and time costs, and improves productivity.

[0005] A 3D Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback includes a code generation / tuning agent, a 3D reconstruction subsystem, a visual quality verification agent, a performance analysis agent, and a static code analysis agent;

[0006] The code generation / tuning agent outputs 3D Gaussian sputtering training code and parameter configuration based on the input dataset, optimization target, and training configuration. The parameter configuration includes the training process, inference mechanism, and rendering settings. The 3D reconstruction subsystem is used to execute the generated 3D Gaussian sputtering training code, perform the training, inference and rendering process of 3D Gaussian sputtering, and generate the trained PLY model, 3D scene point cloud data and the rendered image of the test set. The visual quality verification agent is used to acquire rendering quality-related data, including peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learned perceptual image patch similarity (LPIPS) of the rendered image; and then filters out erroneous data from the rendering quality-related data. The performance analysis agent is used to monitor the performance indicators of the 3D reconstruction subsystem, including frame rate, single-frame image generation time, GPU utilization, and memory bandwidth usage; then, based on the performance indicators, it performs bottleneck location analysis, resource utilization analysis, and records memory access logs. The static code analysis agent is used to analyze the code structure characteristics of the 3D Gaussian sputtering training code, including code correctness checks, function setting and call hierarchy checks, and parameter setting rationality checks. Then, based on the code structure characteristics, it evaluates the code organization and memory access efficiency, and outputs optimization suggestions, refactoring schemes, and memory access trajectories.

[0007] Furthermore, a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback also includes a shared knowledge base; The shared knowledge base is used to collect expert optimization experience, optimization rule templates, and analysis results from visual quality verification agents, performance analysis agents, and static code analysis agents. Then, based on the collected data, it outputs optimization strategies and fusion feedback scores for the 3D Gaussian sputtering training code, and stores the 3D Gaussian sputtering training code and its corresponding optimization strategies as historical cases.

[0008] Furthermore, a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback also includes a feedback fusion and priority ranking module; The feedback fusion and priority ranking module is used to comprehensively evaluate the 3D Gaussian sputtering training code based on the analysis results of the visual quality verification agent, performance analysis agent, and static code analysis agent, thereby outputting a comprehensive optimization decision for the 3D Gaussian sputtering training code, updating the shared knowledge base based on the generated comprehensive optimization decision, and simultaneously updating the priority of optimization experience in the shared knowledge base.

[0009] Furthermore, the feedback fusion and priority ranking module uses the following method to comprehensively evaluate the 3D Gaussian sputtering training code based on the analysis results of the visual quality verification agent, performance analysis agent, and static code analysis agent: The correctness of the 3D Gaussian sputtering training code is checked. If an error occurs or the result is abnormal, the code generation / tuning agent is immediately rolled back to regenerate the 3D Gaussian sputtering training code. If the correctness check passes, different optimization strategies are adopted based on the actual needs of the project to evaluate image quality and inference rendering performance. The optimization strategies include speed-first and quality-first strategies. Then, new modification prompts are generated based on multi-dimensional feedback to resolve optimization conflicts and output a comprehensive optimization decision, knowledge base update, and optimization priority queue. The speed-first strategy is as follows: the inference rendering performance score is given higher priority, and the size of the generated point cloud data is controlled by adjusting the densification threshold, pruning importance score, point cloud data numerical precision, and color sphere harmonic function order. The quality-first strategy is to increase image quality assessment scores, including PSNR, SSIM, and LPIPS.

[0010] Furthermore, a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback also includes an optimization strategy generation module; The comprehensive optimization decision is based on the 3D Gaussian sputtering training code output by the feedback fusion and priority sorting module. At the same time, a new optimization scheme is formulated with reference to the shared knowledge base, and the corrected 3D Gaussian sputtering training code, optimization strategy and memory / schedule / parameter optimization scheme are output.

[0011] Furthermore, a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback also includes an algorithm parameter adjustment module, a scheduling optimization application module, and a memory optimization application module; The algorithm parameter adjustment module is used to execute the parameter optimization scheme output by the optimization strategy generation module, including adjusting the number and precision of Gaussians, inputting the current parameters, performance feedback, and original code, and outputting the optimized hyperparameter configuration and learning rate adjustment scheme. The scheduling optimization application module is used to execute the scheduling optimization scheme output by the optimization strategy generation module, including optimizing thread scheduling, reducing warp divergence, inputting the original code and scheduling optimization strategy, and outputting parallel scheduling optimization code and execution efficiency prediction. The memory optimization application module is used to execute the memory optimization scheme output by the optimization strategy generation module, including improving memory access mode, reducing bandwidth bottlenecks, inputting original code and memory optimization strategy, and outputting memory-optimized code and video memory usage estimate.

[0012] Furthermore, a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback also includes a knowledge base update and learning module; The knowledge base update and learning module is used to store the optimization experience in the shared knowledge base. The optimization experience includes the mapping relationship between successful strategies, optimization cases, scene features, and optimization effects. It also outputs an updated knowledge base and optimization rule templates to provide historical reference for subsequent optimizations.

[0013] Furthermore, a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback also includes an optimized target verification module and an optimized code output module; The optimization target verification module is used to determine whether the corrected 3D Gaussian sputtering training code output by the optimization strategy generation module has reached the preset quality and performance thresholds, and outputs the verification results and performance improvement report. The optimized code output module is used to generate a final deployable code package, performance comparison report, deployment document and portable optimized code when the judgment result of the optimization target verification module is yes; when the judgment result of the optimization target verification module is no, it returns to the code generation / tuning agent to regenerate the 3D Gaussian sputtering training code.

[0014] Beneficial effects: 1. This invention provides a 3D Gaussian code generation system based on multi-agent collaboration using multi-dimensional feedback. It integrates visual quality verification, performance analysis, and static code analysis into a dynamic optimization loop, thereby employing multi-agent collaboration to optimize 3D Gaussian splash code. Simultaneously, it utilizes a knowledge base for cross-task optimization migration and automatic error detection and repair. Compared to existing manual optimization or single-agent tuning solutions, this invention uses multi-dimensional feedback integration (visual quality + performance analysis + static analysis) into the optimization process. The dynamically queryable optimization experience knowledge base corresponds to the reconstructed scene, thus providing higher performance portability and adaptability. It can efficiently and robustly generate and optimize Gaussian splash code for complex 3D scenes, such as large-scale 3D reconstruction, real-time rendering applications, and dynamic scenes.

[0015] 2. This invention provides a 3D Gaussian code generation system based on multi-agent collaboration using multi-dimensional feedback. By utilizing hardware-aware scheduling and memory optimization strategies, the generated 3D Gaussian splash code is automatically adapted to different GPU architectures (such as NVIDIA, AMD, or emerging accelerators). This ensures that performance improvements implemented on one platform can be consistently migrated to other platforms without manual readjustment. By integrating visual quality verification and repair agents into the feedback loop, common sources of 3D Gaussian splash code execution failures, such as visual artifacts, memory access errors, or warp divergence issues, are eliminated. This significantly reduces the burden on developers who need to manually debug and recompile the code.

[0016] 3. This invention provides a 3D Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback. Through knowledge base enrichment, it achieves continuous learning, enabling each optimization cycle to contribute new rules, templates, and performance insights to the shared knowledge base. Over time, the system becomes more intelligent and can handle unseen 3D scenes, thereby bringing cumulative optimization benefits and a self-improving framework. 4. This invention provides a 3D Gaussian code generation system based on multi-agent collaboration and multi-dimensional feedback, capable of handling the Gaussian splashing requirements of modern large-scale 3D systems, including applications based on large-scale 3D reconstruction, real-time rendering, and dynamic scenes. Its modular multi-agent structure supports efficient expansion without causing an exponential increase in development or tuning workload. Attached Figure Description

[0017] Figure 1 The present invention provides a flowchart of a three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0019] First, we will introduce the two core mechanisms involved in this invention: multi-agent collaborative optimization and multi-dimensional feedback mechanism.

[0020] I. Multi-agent cooperative optimization This invention innovatively integrates multiple agents with different roles for 3D Gaussian splatter code generation. Starting with multi-view image data input, multiple agent instances are initialized. After the code generation / tuning agent generates initial code, it is evaluated in parallel by three specialized agents: visual quality verification, performance analysis, and static code analysis. Feedback from each agent is aggregated into a shared knowledge base, and coordinated decision-making by a multi-dimensional feedback integration center drives an automatic error correction and optimization loop. Through iterative improvement, the system ultimately outputs high-quality 3D Gaussian splatter code that is portable across GPU platforms, has optimized memory access, and maximizes bandwidth utilization, forming a closed-loop optimization system.

[0021] II. Multidimensional Feedback (Visual Quality, Performance Analysis, Static Analysis) This invention introduces a unified feedback loop that combines visual quality checks, runtime performance analysis, and static code analysis. These complementary dimensions enable the system to identify functional errors and performance bottlenecks, thereby enabling fine-grained optimization strategies and allowing 3D Gaussian splash code to dynamically adapt to specific scenarios and hardware requirements.

[0022] Based on this, the present invention provides a 3D Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback, comprising a code generation / tuning agent, a 3D reconstruction subsystem, a visual quality verification agent, a performance analysis agent, and a static code analysis agent, the workflow of which is as follows: Figure 1 As shown, it includes the following steps: 1. Multi-view image dataset + camera pose information input: The system receives the original multi-view image dataset and camera pose parameters, performs data preprocessing and standardization, and provides input data for subsequent 3D reconstruction.

[0023] 2. Initialization - Multiple Agent Instances: The system initializes multiple dedicated agent instances, including code generation agent, visual quality verification agent, performance analysis agent, and static code analysis agent, forming a multi-agent collaborative framework.

[0024] 3. Execution code generation / tuning agent (3D Gaussian sputtering kernel execution module): The agent generates or tunes 3D Gaussian sputtering training code and parameter configuration based on the input dataset and optimization target. The parameter configuration includes the training process, inference mechanism and rendering settings.

[0025] The prompts include system prompts and user prompts: System prompts include code template examples (such as training and rendering code for classic 3D Gaussian reconstruction, and the code implementation for CUDA acceleration), and general rules for generating code, such as limiting the user platform (e.g., RTX 4090) and version information (e.g., CUDA 11.8), the number of training iterations, the order of the spherical harmonic function, etc. User prompts provide customized generation information for the user, such as requirements for the size of the generated PLY file, rendering speed requirements, etc.

[0026] 4. Perform 3D Gaussian 3D reconstruction based on generated code: Execute the generated 3D Gaussian sputtering training code to perform the training, inference and rendering process of 3D Gaussian sputtering, and generate the trained PLY model, 3D scene point cloud data and rendered images of the test dataset, providing a basis for subsequent feedback collection.

[0027] Parallel feedback collection phase: The system simultaneously initiates feedback collection from three dimensions, forming a parallel processing mechanism. 5. Execution - Visual Quality Verification Agent (Visual Quality Feedback Collection Module): Responsible for collecting rendering quality-related data, including peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learned perceptual image patch similarity (LPIPS) scores of the rendered images in the test set; and then filtering out erroneous data from the rendering quality-related data.

[0028] 6. Execution-Performance Analysis Agent (Performance Analysis Feedback Collection Module): Monitors the performance metrics of the 3D reconstruction subsystem, including FPS (frame rate), single-frame image generation time, GPU utilization, and memory bandwidth usage; then, based on the performance metrics, it locates bottlenecks in the 3D reconstruction subsystem, analyzes resource utilization, and records memory access logs.

[0029] 7. Execution - Static Code Analysis Agent (Static Code Analysis Feedback Collection Module): Analyzes the code structure characteristics of the 3D Gaussian sputtering training code, including code correctness checks, function setting and call hierarchy checks, and parameter setting rationality checks. Then, based on the code structure characteristics, it evaluates code organization and memory access efficiency, and outputs optimization suggestions, refactoring schemes, and memory access trajectories.

[0030] 8. Visual Quality Verification and Error Detection (PSNR / SSIM Assessment and Visual Artifact Detection): Accurately quantifies image quality, identifies visual defects, and outputs error detection results, quality metrics, PSNR values, and SSIM scores.

[0031] 9. Runtime performance bottleneck analysis (GPU utilization and memory bandwidth analysis): Measure hardware resource utilization efficiency, locate bottlenecks, analyze resource utilization, and record memory access logs.

[0032] 10. Code Structure and Optimization Opportunity Identification (Code Structure and Memory Access Pattern Analysis): Evaluate code organization and memory access efficiency, and output optimization suggestions, refactoring schemes, and memory access trajectories.

[0033] 11. Shared Knowledge Base: It gathers expert optimization experience, optimization rule templates, and visual / performance / code analysis results. Then, based on the collected data, it outputs optimization strategies for 3D Gaussian sputtering training code, integrates feedback scores, and stores 3D Gaussian sputtering training code and its corresponding optimization strategies as historical cases.

[0034] 12. Multidimensional Feedback Integration and Decision Coordination Center (Feedback Fusion and Prioritization Module): This is the core decision point of the 3D Gaussian code generation system, which comprehensively evaluates the feedback data from three dimensions (image quality, inference rendering performance, and code quality).

[0035] The code correctness check has a veto power: if the runtime reports an error or the result is abnormal, it will directly revert to the code generation / tuning agent to regenerate the 3D Gaussian sputtering training code; If the correctness check passes, the image quality and inference rendering performance are evaluated based on the actual needs of the project, and two optimization strategies are adopted: Speed-first strategy: Inference rendering performance score is given higher priority, and the size of the generated point cloud data is controlled by adjusting densification threshold, pruning importance score, point cloud data numerical precision (FP16, FP8, etc.), and color sphere harmonic function order (1~4). Quality-first strategy: Increase image quality assessment scores (PSNR, SSIM, LPIPS, or their combined statistical results). Finally, based on multi-dimensional feedback, new modification prompts are generated to resolve optimization conflicts, and comprehensive optimization decisions, knowledge base updates, and optimization priority queues are output. 13. Automatic Error Correction and Optimization Loop (Optimization Strategy Generation Module): Based on the fusion feedback results and with reference to the shared knowledge base, it formulates specific optimization schemes and outputs corrected code, optimization strategies, and memory / schedule / parameter optimization schemes.

[0036] Application optimization phase: The optimization strategy is broken down into three implementation directions: 13. Algorithm Parameter Tuning Module (Memory Optimization Application Module): This module executes the parameter optimization schemes output by the optimization strategy generation module. It includes adjusting key parameters such as the number of Gaussians and accuracy (the main optimization items). The input parameters are the current parameters, performance feedback, and the original code. The output is the optimized hyperparameter configuration and learning rate adjustment scheme.

[0037] 12. Scheduling Optimization Application Module: This module executes the scheduling optimization scheme output by the optimization strategy generation module. This includes optimizing thread scheduling, reducing warp divergences (Nvidia uses Warp, AMD's equivalent concept is Wavefront, or equivalent concepts for other computing platforms. This optimization strategy is not limited to a specific computing platform). Input is the original code and the scheduling optimization strategy; output is the parallel scheduling optimization code and execution efficiency prediction.

[0038] 13. Memory Optimization Application Module: Executes the memory optimization scheme output by the optimization strategy generation module, including improving memory access patterns and reducing bandwidth bottlenecks (optional optimization). Inputs the original code and memory optimization strategy, and outputs the memory-optimized code and estimated video memory usage.

[0039] 14. Knowledge Base Update and Learning Module: Store the optimization experience (optimization cases, success / failure experiences) in the shared knowledge base, including the mapping relationship between successful strategies, scenario features and optimization effects, and output the updated knowledge base and optimization rule templates to provide historical reference for subsequent optimizations.

[0040] 15. Optimization Target Verification Module: Used to determine whether the corrected 3D Gaussian sputtering training code output by the optimization strategy generation module meets the preset quality and performance thresholds. Input the optimized code and performance target thresholds, and output the verification result (satisfied / unsatisfied) and performance improvement report.

[0041] 16. Optimized code output module (optimized 3D Gaussian sputtering code output): If the objective is achieved, that is, when the judgment result of the optimization objective verification module is yes, the final deployable code package, performance comparison report, deployment document and portable optimization code will be generated. If the goal is not achieved, that is, if the judgment result of the optimization goal verification module is negative, return to step 3 (execute code generation / tuning the agent) and continue iterative optimization to form a closed-loop optimization process.

[0042] It should be noted that this invention is also applicable to all extended systems with a multi-agent kernel, such as OpenClaw or OpenClaw-like products, products with extended upstream and downstream permissions, UI-encapsulated solutions, etc.

[0043] Therefore, this invention is based on multi-agent collaboration using a shared knowledge base. Multiple specialized agents operate within a coordinated framework. Each agent (a system entity that, in a given environment, acquires information through perception and autonomously makes decisions and takes actions based on goals or strategies to influence the state of the environment; the underlying decision-making is typically driven by a large model) focuses on different optimization aspects, such as memory management, scheduling strategies, or visual quality analysis. Their decisions are coordinated through a shared knowledge base containing optimization templates, rules, and historical experience, ensuring continuous improvement and knowledge transfer across tasks and hardware platforms.

[0044] Furthermore, this invention integrates an iterative repair mechanism that automatically corrects errors detected during visual quality verification or performance analysis using a specialized intelligent agent. This cycle minimizes human intervention, accelerates the 3D Gaussian splash code development cycle, and ensures that the optimized code remains functionally correct and performs efficiently in constantly changing 3D scenes and hardware environments.

[0045] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A multi-agent collaboration based on multi-dimensional feedback for three-dimensional Gaussian code generation system, characterized in that, This includes a code generation / tuning agent, a 3D reconstruction subsystem, a visual quality verification agent, a performance analysis agent, and a static code analysis agent; The code generation / tuning agent outputs 3D Gaussian sputtering training code and parameter configuration based on the input dataset, optimization target, and training configuration. The parameter configuration includes the training process, inference mechanism, and rendering settings. The 3D reconstruction subsystem is used to execute the generated 3D Gaussian sputtering training code, perform the training, inference and rendering process of 3D Gaussian sputtering, and generate the trained PLY model, 3D scene point cloud data and the rendered image of the test set. The visual quality verification agent is used to acquire rendering quality-related data, including peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learned perceptual image patch similarity (LPIPS) of the rendered image; and then filters out erroneous data from the rendering quality-related data. The performance analysis agent is used to monitor the performance indicators of the 3D reconstruction subsystem, including frame rate, single-frame image generation time, GPU utilization, and memory bandwidth usage; then, based on the performance indicators, it performs bottleneck location analysis, resource utilization analysis, and records memory access logs. The static code analysis agent is used to analyze the code structure characteristics of the 3D Gaussian sputtering training code, including code correctness checks, function setting and call hierarchy checks, and parameter setting rationality checks. Then, based on the code structure characteristics, it evaluates the code organization and memory access efficiency, and outputs optimization suggestions, refactoring schemes, and memory access trajectories.

2. The multi-agent collaboration based on multi-dimensional feedback for three-dimensional Gaussian code generation system of claim 1, wherein, It also includes a shared knowledge base; The shared knowledge base is used to collect expert optimization experience, optimization rule templates, and analysis results from visual quality verification agents, performance analysis agents, and static code analysis agents. Then, based on the collected data, it outputs optimization strategies and fusion feedback scores for the 3D Gaussian sputtering training code, and stores the 3D Gaussian sputtering training code and its corresponding optimization strategies as historical cases.

3. The multi-agent collaboration based on multi-dimensional feedback for three-dimensional Gaussian code generation system of claim 2, wherein, It also includes a feedback fusion and priority sorting module; The feedback fusion and priority ranking module is used to comprehensively evaluate the 3D Gaussian sputtering training code based on the analysis results of the visual quality verification agent, performance analysis agent, and static code analysis agent, thereby outputting a comprehensive optimization decision for the 3D Gaussian sputtering training code, updating the shared knowledge base based on the generated comprehensive optimization decision, and simultaneously updating the priority of optimization experience in the shared knowledge base.

4. The multi-agent collaboration based on multi-dimensional feedback for three- dimensional Gaussian code generation system of claim 3, wherein, The feedback fusion and priority ranking module is used to comprehensively evaluate the 3D Gaussian sputtering training code based on the analysis results of the visual quality verification agent, performance analysis agent, and static code analysis agent. The method is as follows: Check the correctness of the 3D Gaussian sputtering training code. If an error occurs or the result is abnormal, immediately return to the code generation / tuning agent to regenerate the 3D Gaussian sputtering training code. If the correctness check passes, different optimization strategies are adopted based on the actual needs of the project to evaluate image quality and inference rendering performance. The optimization strategies include speed-first strategy and quality-first strategy. Then, new modification prompts are generated based on multi-dimensional feedback to resolve optimization conflicts and output comprehensive optimization decisions, knowledge base updates and optimization priority queues. The speed-first strategy is as follows: the inference rendering performance score is given higher priority, and the size of the generated point cloud data is controlled by adjusting the densification threshold, pruning importance score, point cloud data numerical precision, and color sphere harmonic function order. The quality-first strategy is to increase image quality assessment scores, including PSNR, SSIM, and LPIPS.

5. The multi-agent collaboration based on multi-dimensional feedback for three-dimensional Gaussian code generation system of claim 3, wherein, It also includes an optimization strategy generation module; The comprehensive optimization decision is based on the 3D Gaussian sputtering training code output by the feedback fusion and priority sorting module. At the same time, a new optimization scheme is formulated with reference to the shared knowledge base, and the corrected 3D Gaussian sputtering training code, optimization strategy and memory / schedule / parameter optimization scheme are output.

6. The multi-agent collaboration based on multi-dimensional feedback for three-dimensional Gaussian code generation system of claim 5, wherein, It also includes an algorithm parameter adjustment module, a scheduling optimization application module, and a memory optimization application module; The algorithm parameter adjustment module is used to execute the parameter optimization scheme output by the optimization strategy generation module, including adjusting the number and precision of Gaussians, inputting the current parameters, performance feedback, and original code, and outputting the optimized hyperparameter configuration and learning rate adjustment scheme. The scheduling optimization application module is used to execute the scheduling optimization scheme output by the optimization strategy generation module, including optimizing thread scheduling, reducing warp divergence, inputting the original code and scheduling optimization strategy, and outputting parallel scheduling optimization code and execution efficiency prediction. The memory optimization application module is used to execute the memory optimization scheme output by the optimization strategy generation module, including improving memory access mode, reducing bandwidth bottlenecks, inputting original code and memory optimization strategy, and outputting memory-optimized code and video memory usage estimate.

7. The multi-agent collaboration based on multi-dimensional feedback for three- dimensional Gaussian code generation system of claim 6, wherein, It also includes a knowledge base update and learning module; The knowledge base update and learning module is used to store the optimization experience in the shared knowledge base. The optimization experience includes the mapping relationship between successful strategies, optimization cases, scene features, and optimization effects. It also outputs an updated knowledge base and optimization rule templates to provide historical reference for subsequent optimizations.

8. The three-dimensional Gaussian code generation system based on multi-agent collaboration with multi-dimensional feedback as described in claim 7, characterized in that, It also includes optimizations to the target verification module and the code output module; The optimization target verification module is used to determine whether the corrected 3D Gaussian sputtering training code output by the optimization strategy generation module has reached the preset quality and performance thresholds, and outputs the verification results and performance improvement report. The optimized code output module is used to generate a final deployable code package, a performance comparison report, a deployment document, and portable optimized code when the judgment result of the optimization target verification module is yes. If the result of the optimization target verification module is negative, then return to the code generation / tuning agent to regenerate the 3D Gaussian sputtering training code.

Images