Batch denoising method for edge AIGC service
By designing a batch denoising model and stacking algorithm in the edge network, optimizing the denoising steps and task scheduling, the problem of balancing AIGC service quality and latency in the edge network is solved, and high-quality and low-latency AIGC services are achieved.
Patent Information
- Application Number
- CN202511491929.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-01-23
AI Technical Summary
Existing technologies have not yet effectively utilized batch processing techniques in edge networks to provide high-quality and low-latency AIGC services, making it difficult to achieve a balance between generation latency and service quality.
A batch denoising method for edge AIGC services is designed. By establishing a batch denoising model and constructing a generation optimization problem, the STACKING algorithm is used for optimization. Combined with clustering, packing and batch processing techniques, the denoising steps and task scheduling are optimized to maximize generation quality and reduce latency.
It maximizes the average AIGC service quality under the constraint of generated content latency, providing high-quality and low-latency AIGC services to multiple users in the edge network, accurately representing the relationship between the GenAI denoising steps and service quality, and reducing computational complexity.
Smart Images
Figure CN121387477A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence, and more particularly to a batch denoising method for edge AIGC services. Background Technology
[0002] Artificial intelligence-generated content (AIGC) services refer to the application of generative artificial intelligence (GenAI) models to provide personalized and diverse content generation services, such as image and video generation. With the proliferation of mobile devices, the demand for high-quality and low-latency AIGC services in edge networks is rapidly increasing. For example, integrating GenAI into edge networks enables new in-vehicle services, such as augmented reality road simulation. Edge AIGC services require the use of GenAI models deployed on edge servers to generate content based on user needs.
[0003] Mainstream image and video generation primarily utilizes the diffusion-based GenAI model, which progressively denoises Gaussian noise through multiple denoising steps to generate content. Increasing the number of denoising steps generally improves the quality of generated content but also increases generation latency. To balance AIGC service quality and responsiveness, establishing a quantitative relationship between denoising steps and service quality is crucial. Several studies have already explored the relationship between the denoising steps in diffusion-based GenAI and AIGC service quality.
[0004] Furthermore, each denoising step in the diffusion-based GenAI model consumes significant computational resources, resulting in substantial computational latency for iterative denoising. Some research utilizes batch processing to support AI services in wireless edge networks. Batch processing, a fundamental technique in deep learning, groups multiple services into a single batch for parallel processing. By reusing AI models loaded into the graphics processing unit (GPU) to execute multiple tasks, it reduces the frequency of memory accesses and amortizes memory access time, thereby improving computational resource utilization and reducing computational latency per service. However, no research has yet utilized batch processing techniques to provide AIGC services in edge networks.
[0005] Therefore, those skilled in the art are dedicated to developing a batch denoising method for edge AIGC services. This method maximizes the average AIGC service quality under content generation latency constraints, providing high-quality and low-latency AIGC services to multiple users in edge networks. Summary of the Invention
[0006] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is to provide AIGC services in edge networks using batch processing technology.
[0007] To achieve the above objectives, this invention provides a batch denoising method for edge AIGC services, comprising the following steps: Step 1: Establish a batch denoising model and construct a generation optimization problem; Step 2: Solve using the optimized batch denoising STACKING algorithm.
[0008] Furthermore, in step 1, the optimization objective is to maximize the average generated quality under the content generation delay constraint.
[0009] Furthermore, in step 1, the relationship between the number of denoising steps and content quality is modeled in the implicit denoising diffusion model.
[0010] Furthermore, the quality of the content is measured using the Fraser initial distance.
[0011] Furthermore, each denoising step in the AIGC service is treated as a denoising task, and denoising tasks are scheduled in a batch under the constraint of the dependencies between denoising steps.
[0012] Furthermore, the STACKING algorithm does not depend on the specific characteristics of the content quality function.
[0013] Furthermore, the STACKING algorithm process includes clustering, packaging, and batch processing.
[0014] Furthermore, the STACKING algorithm utilizes the parallel computing capabilities of batch denoising to reduce the denoising delay at each step.
[0015] Furthermore, the generation optimization problem is:
[0016]
[0017] ,
[0018]
[0019]
[0020] Among them, symbols Indicates any, symbol Indicates belonging to, For service The number of denoising steps in the generated image. Indicates the first The start time of each batch, Indicates service The Was the denoising task assigned to the first...? One batch, Indicates the quantity of services or equipment. Represents a collection of service or device indexes. Indicates service The smaller the value, the higher the quality of the generated content. Indicates the batch quantity. Represents a batch index set. Indicates service Generate an index set for image denoising steps. Represents the denoising delay function. Indicates the first The size of each batch Indicates removal Batch index set, Indicates removal The set of indexes for denoising steps. Indicates service Content generation delay Indicates service The deadline.
[0021] Further, step 2 includes: Step 2.1, Clustering: Based on the expected number of denoising tasks for the service and the ideal maximum number of denoising tasks that the service can currently complete, the service is divided into two clusters; Step 2.2, Packaging: Group the services in the cluster, and batch processing prioritizes services with strict latency requirements; Step 2.3, Batch processing: Batch noise reduction.
[0022] Existing batch processing techniques are not suitable for AIGC services and cannot provide high-quality and low-latency AIGC services to multiple users in edge networks. This invention proposes a batch denoising method to provide high-quality and low-latency AIGC services to multiple users in edge networks. Specifically, an AIGC generation optimization problem is constructed based on the proposed batch denoising model, and then a specific feature stacking algorithm independent of the content quality function is designed to solve this problem. This invention treats each denoising step of the AIGC service as a denoising task, and schedules denoising tasks in the batch under the constraint of the dependencies between denoising steps; in addition, the design principle of the stacking algorithm is derived from the parallel computing capability of batch processing, and the initial denoising step based on the diffusion-based GenAI model has a greater impact on the quality of the generated content than subsequent steps.
[0023] Existing relationships between GenAI denoising steps and AIGC service quality are either overly simplistic or only applicable to specific models, making them unusable directly. This invention models the relationship between the number of denoising steps and the quality of generated content in the Denoising Diffusion Implicit Model (DDIM), where content quality is measured by FID. Extensive experiments were conducted to obtain a large amount of real-world data, and based on this, a function representing this relationship was derived through fitting.
[0024] Compared with the prior art, the present invention has the following obvious substantive features and significant advantages: 1. This invention maximizes the average AIGC service quality under the constraint of content generation latency, thereby providing high-quality and low-latency AIGC services to multiple users in the edge network.
[0025] 2. This invention precisely illustrates the relationship between the GenAI denoising step and the AIGC service quality.
[0026] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description
[0027] Figure 1 This is a preferred embodiment of the invention: inference delay vs. batch size. Figure 2 This is a preferred embodiment of the present invention: FID score vs. number of denoising steps; Figure 3 This is a preferred embodiment of the relationship between average FID score and number of services. Figure 4 This is a preferred embodiment of the relationship between the average FID score and the minimum latency requirement. Figure 5 This is a preferred embodiment of the STACKING algorithm implemented in the present invention to show the execution status of edge AIGC services. Detailed Implementation
[0028] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.
[0029] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.
[0030] This invention proposes a batch denoising method for multi-user AIGC services in edge networks, with image generation as a typical application scenario. To maximize average generation quality under content generation latency constraints, a batch denoising model is designed and a generation optimization problem is constructed. Based on requirements, the relationship between the number of denoising steps and content quality in the Denoising Diffusion Implicit Model (DDIM) is accurately modeled, where content quality is measured by FID. Inspired by the powerful parallel computing capabilities of batch denoising to reduce denoising latency per step, and considering that the initial denoising step has a greater impact on generation quality than subsequent steps, a STACKING algorithm is designed to optimize batch denoising. This algorithm follows a clustering-packaging-batch processing flow. The STACKING algorithm is independent of the specific characteristics of the content quality function and effectively reduces computational complexity.
[0031] 1. System Model and Problem Formulation
[0032] Consider a single-cell system, where one edge server is... Each edge device provides AIGC services, and the device index is... Applications on the device share pre-trained GenAI models, such as DDIM, deployed on the server to enable AIGC services with diverse generation latency requirements. The batch denoising model will be explained in detail below. It's important to note that due to the one-to-one correspondence between "device" and "service," these two terms will be used interchangeably. The number of denoising steps in generating the image is denoted as . .
[0033] 1.1 Batch Denoising Model
[0034] Batch denoising is employed to achieve rapid content generation. For service k, the number of denoising steps is expressed as... Its index set is Each denoising step is treated as an independent denoising task. During batch denoising, the edge server will... The noise reduction task of each service is assembled into There are batches, and the batch index set is denoted as... . Each batch is processed sequentially and meets the requirements. ,in Indicates the first The start time of each batch.
[0035] make As a binary variable, it represents the service. The Was the denoising task assigned to the first...? One batch:
[0036] in the case of, ;otherwise, To achieve the desired content quality, each denoising task in the service must be performed once.
[0037] No. The size of each batch, i.e. the number of denoising tasks assigned, is calculated as follows:
[0038] from Figure 1 It was observed that the first The denoising delay of each batch can be modeled as:
[0039] Among them, parameters and These are constants that are related to specific application scenarios and GPU hardware. express Norm. Figure 1 The curve showing the denoising latency as a function of batch size is presented. The measured data was obtained using a DDIM model pre-trained on the CIFAR-10 dataset on an NVIDIA GeForce RTX 3050 graphics card, with the fitted values for the corresponding parameters being... , .
[0040] Serve The total time required to complete the noise reduction task also represents the service's performance. The content generation delay can be expressed as:
[0041] Due to the execution order of batch processing and the dependencies between denoising tasks in a single service, the following two constraints must be satisfied. First, up to the... The reasoning for the first batch is completed, the first batch... Only one batch can be processed:
[0042] in, Secondly, regarding services until the The first noise reduction task is complete. Only one denoising task can be processed:
[0043] in, .
[0044] 1.2 Problem Formulation
[0045] Considering service The content generation delay cannot exceed the preset deadline. :
[0046] Under the deadline constraint, the AIGC service quality maximization problem is formulated as follows:
[0047]
[0048] The goal is to maximize the quality of AIGC services by optimizing the denoising steps, batch start time, and task-batch allocation. Furthermore, (1), (2), (6), and (7) impose constraints on batch denoising; (8) ensures that the deadline constraint is met. A smaller function Q indicates higher generation quality, and minimizing this function maximizes generation quality.
[0049] Problem (P0) is difficult to solve for the following reasons: ·Serve Content generation quality It is an implicit function of the number of denoising steps, and usually exhibits a complex form. Figure 2 In this study, based on images generated using DDIM pre-trained on CIFAR-10, the FID score is used to evaluate the image generation quality, and its relationship with the number of denoising steps is shown. With the increase in denoising steps, the FID score initially drops sharply, then gradually levels off, indicating an improvement in image quality. A power-law function is used to accurately fit the measurement data.
[0050] Content generation latency depends on task batch allocation, where The presence of norms and product terms introduces significant complexity.
[0051] 2. Algorithm Design
[0052] from Figure 1 It can be observed that the parameters Greater than the parameter This indicates that the overhead of loading the model from memory by the GPU is significantly higher than the computational latency caused by increasing the batch size. Therefore, the first approach to batch denoising optimization is to maximize the number of denoising tasks in each batch. This improves computational efficiency. However, Increased latency can prevent some services with strict latency constraints from completing sufficient denoising tasks, thus limiting the quality of generated content. For example... Figure 2As shown, the initial denoising step has a far greater impact on image quality than subsequent steps. Therefore, the second approach to batch denoising optimization is to balance the number of denoising steps across different services as much as possible.
[0053] Inspired by the two ideas mentioned above, a stacking algorithm was designed, utilizing auxiliary variables. To solve for (P0), where Defined as the expected number of denoising tasks per service. Specifically, the STACKING algorithm includes the following three steps.
[0054] 1) Clustering: In the first stage... Record each service before each batch begins. Number of denoising tasks completed so far and update its remaining delay to
[0055] Each service at present Maximum number of noise reduction tasks that can still be completed Calculated as
[0056] Therefore, each service The ideal maximum number of denoising tasks that can be completed at present is expressed as:
[0057] By comparison and All services are divided into two clusters: one is
[0058] And the other is its complement. For ease of description, All services are in accordance with Sort in ascending order.
[0059] 2) Packaging: After clustering, a packaging step is performed to group services from the two clusters for subsequent batch processing. This is because batch processing always prioritizes services with smaller cluster sizes. The service is designed with a packaging strategy based on the following two scenarios.
[0060] Case 1: When At that time, prioritize services with strict delay requirements by setting... Pack them together. However, if If the value is small, this setting will delay processing. This could reduce the number of denoising tasks that can be completed by the deadline, as the service in the middle may be affected. This is because appropriately increasing the batch size while packaging... For some services with strict latency requirements, set the batch size to:
[0061] in This item ensures that for all , Establishment means packaging Some services will not be reduced. The number of denoising tasks completed by each service. In addition, and They represent middle maximum value and The minimum value.
[0062] Case 2: When At that time, there are no services that need to be prioritized for batch processing and The desired size is as large as possible, with an upper limit of 1. However, too large May significantly reduce The number of denoising tasks for services with strict latency requirements. Therefore, setting...
[0063] To satisfy all services Maximize under the premise . This ensured ,Right now The number of denoising tasks that each service can ultimately complete will not be less than the threshold. ,in Represents a set middle The minimum value.
[0064] 3) Batch processing: Collections Center front The latest noise reduction tasks for each service are packaged into the first... In each batch. If the services are packaged... Remaining delay Less than If the denoising task is complete, then the service is considered to have completed all denoising tasks. Accordingly, its denoising tasks are removed from the batch and... Subtract 1. Remaining tasks and the updated and They are then batch-processed together for noise reduction. Meanwhile, for the first... Each service in each batch Its denoising steps index Add 1, set And Updated to . No. After the second batch of noise reduction, services that violated the latency requirements were... Remove it from the batch to prevent it from being processed in subsequent batches.
[0065] The above clustering-packaging-batch processing process is repeated until... An empty value indicates that the denoising task for all AIGC services has been completed. Furthermore, The optimal value depends on the distribution of latency requirements for all services. To determine this value, a search is performed within a predefined range, selecting the value that minimizes the objective function (P0). The complete process is summarized in Algorithm 1.
[0066]
[0067] Figure 5 The execution of the STACKING algorithm for edge AIGC services is demonstrated. It can be observed that services with strict latency requirements are prioritized. Additional denoising tasks from services with less stringent latency requirements are included in batch processing, ensuring the quality of the images generated by the services, to maximize computational resource utilization. Furthermore, when only services with less stringent latency requirements remain, the server processes as many services as possible. Overall, services with similar latency requirements tend to have a similar number of denoising steps at the end. Services are completed close to the deadline, thus maximizing computational resource utilization.
[0068] Single-instance, greedy batch processing, and fixed-size batch processing schemes were set as benchmark schemes, and compared with the present invention through experiments: from Figure 3 It can be observed that as the number of services increases, the average FID score gradually rises, indicating that the average quality of AIGC services has decreased. The single-instance solution struggles to support multi-user AIGC services, highlighting the necessity of batch denoising. Under continuously increasing service load, neither greedy batch processing nor fixed-size batch processing solutions can effectively allocate computing resources, leading to a rapid increase in the average FID score and even service interruptions.
[0069] Figure 4The paper demonstrates how the average FID score changes with the minimum latency requirement, while the maximum latency requirement is fixed at 20 seconds. This invention consistently achieves the lowest average FID score, and the smaller the minimum latency requirement, the more significant the performance improvement. Compared to greedy batch processing and fixed-size batch processing schemes, this invention, with its flexible batch denoising strategy, performs better under strict latency constraints.
[0070] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A batch denoising method for edge AIGC services, characterized in that, Includes the following steps: Step 1: Establish a batch denoising model and construct a generation optimization problem; Step 2: Solve using the optimized batch denoising STACKING algorithm.
2. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, The optimization objective of step 1 is to maximize the average generated quality under the constraint of content generation delay.
3. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, Step 1 establishes the relationship between the number of denoising steps and content quality in the implicit denoising diffusion model.
4. The batch denoising method for edge AIGC services as described in claim 3, characterized in that, The quality of the content is measured using the Fraser initial distance.
5. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, Each denoising step in the AIGC service is treated as a denoising task, and denoising tasks are scheduled in a batch while adhering to the constraints of the dependencies between denoising steps.
6. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, The STACKING algorithm does not depend on the specific characteristics of the content quality function.
7. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, The STACKING algorithm process includes clustering, packaging, and batch processing.
8. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, The STACKING algorithm utilizes the parallel computing capabilities of batch denoising to reduce the denoising delay at each step.
9. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, The generation optimization problem is: ， Among them, symbols To indicate any, the symbol Indicates belonging to, For service The number of denoising steps in the generated image. Indicates the first The start time of each batch, Indicates service The Was the denoising task assigned to the first...? One batch, Indicates the quantity of services or equipment. Represents a collection of service or device indexes. Indicates service The smaller the value, the higher the quality of the generated content. Indicates the batch quantity. Represents a batch index set. Indicates service Generate an index set for image denoising steps. Represents the denoising delay function. Indicates the first The size of each batch Indicates removal Batch index set, Indicates removal The set of denoising step indexes, Indicates service Content generation delay Indicates service The deadline.
10. The batch denoising method for edge AIGC services as described in claim 1, characterized in that, Step 2 includes: Step 2.1, Clustering: Based on the expected number of denoising tasks for the service and the ideal maximum number of denoising tasks that the service can currently complete, the service is divided into two clusters; Step 2.2, Packaging: Group the services in the cluster, and batch processing prioritizes services with strict latency requirements; Step 2.3, Batch processing: Batch noise reduction.