Material generation method and electronic device

Abstract

The embodiment of the application discloses a material generation method and an electronic device. The method comprises the following steps: after receiving a material generation request submitted by a user and determining associated user input condition information, the following steps are executed by a generative language model: user intention understanding is performed according to the input condition information, and a structured expression material generation target is generated; an execution plan for the material generation target is generated according to the input condition information and the structured expression material generation target; a material generation task is executed according to a plurality of nodes included in the execution plan to complete the material generation target; wherein during the execution of each step process, the required policy organization information is loaded on demand, and the corresponding task target is completed by understanding the loaded policy organization information. Through the embodiment of the application, path autonomous planning can be realized, the dependence on artificial definition workflow is reduced, the cost is reduced, and the efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of content generation technology, and in particular to methods and electronic devices for generating materials. Background Technology

[0002] With the continuous expansion of the business scale of product information service systems (also known as "e-commerce platforms"), the demand for content production has become increasingly complex and diversified. It has evolved from simply generating a single product main image to a complex production system encompassing multiple types and tasks, including product titles, selling point copy, product main images, scene images, marketing posters, SKU (Stock Keeping Unit) images, detail images, and product videos. From an input perspective, the input conditions for content production are extremely complex, potentially including product information, product images (main image, white background image, model image), reference images, vague user needs, platform specifications, style requirements, category differences, historical context, etc., and these conditions are often incomplete, inconsistent, and constantly changing. From a task perspective, there are interdependent relationships between content production tasks; for example, generating a title and selling points first guides image generation. From a constraint perspective, constraints such as the accuracy of product information, platform style compatibility, and marketing compliance overlap, leading to an exponential increase in process combinations.

[0003] In response to the above situation, the existing technical solutions mainly rely on engineers to predefine workflows and combine them with single-point AI (Artificial Intelligence) capabilities for material production. However, different workflows need to be defined for different material production needs. Whenever a new type of material production need appears, engineers need to redesign and maintain a new workflow, which results in high development costs and long iteration cycles. Summary of the Invention

[0004] This application provides a material generation method and electronic device that enables autonomous path planning, reduces reliance on manually defined workflows, lowers costs, and improves efficiency.

[0005] This application provides the following solution: A method for generating materials, comprising: After receiving the user's request to generate materials and determining the associated user input information, the following steps are performed using a generative language model: Based on the input condition information, the user intent is understood, and a structured expression of the material generation target is generated; Based on the input condition information and the structured expression of the material generation target, an execution plan is generated for the material generation target; The material generation task is executed according to multiple nodes included in the execution plan to achieve the material generation goal; wherein, at least some of the tasks on the nodes are completed by calling pre-encapsulated tools; In the process of executing each step, the required strategy organization information is progressively loaded on demand according to the task objective, and the corresponding task objective is completed by understanding the loaded strategy organization information.

[0006] This also includes: When understanding user intent based on the input condition information, if it is determined that there is a lack of context information in the input condition information, the information is obtained by calling a pre-encapsulated query tool to complete the missing context information, so as to generate a structured expression material generation target in the state of complete context information.

[0007] This also includes: When performing user intent understanding based on the input condition information, if the input condition information includes image-related information, then the image understanding model is invoked to perform intent understanding on the input condition information.

[0008] This also includes: During the execution of the material generation task, the generation results at the corresponding node are verified by loading strategy organization information related to quality verification. If the results do not meet the standards, the execution plan is dynamically adjusted to trigger the regeneration of the corresponding materials.

[0009] This also includes: Save context state information during multiple rounds of task execution within the same session; During the process of understanding user intent and / or generating execution plans, the context state information is queried through the session ID so as to combine the context state information for understanding user intent and / or generating execution plans.

[0010] This also includes: After the session returns the result of generating materials on one or more nodes and receives feedback from the user on the result, if the feedback is negative, the execution plan is dynamically adjusted according to the feedback to regenerate materials on the corresponding node.

[0011] This also includes: After completing the material generation task, obtain the quality score information of the material generation results, and add the execution plans that meet the quality score conditions as reusable strategy organization information to the strategy organization information set.

[0012] The user input condition information includes: generating a material package for the target product to be published in the target system, wherein the information of the target product in the user input condition information is expressed through a product details page link and / or product ID; The step of understanding user intent based on the input condition information and generating a structured expression of the material generation target includes: Based on the input condition information, user intent is understood, and by loading the strategy organization information related to the definition of material packages in the target system, a structured material generation target is generated. The structured material generation target includes: title text, selling point copy, main image, and scene image. The nodes in the execution plan include: obtaining product information, generating title text, generating selling point copy, generating prompt text for image generation, generating main image, and generating scene image; The execution of the material generation task according to the multiple nodes included in the execution plan includes: Based on the material generation tasks at multiple nodes, the necessary strategy organization information is progressively loaded to complete the corresponding material generation tasks.

[0013] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of any of the preceding methods.

[0014] An electronic device, comprising: One or more processors; and A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of any of the preceding methods.

[0015] A computer program product includes a computer program / computer executable instructions that, when executed by a processor in an electronic device, implement the steps of any of the preceding methods.

[0016] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides an automated e-commerce content production system based on Agentic AI. Upon receiving a content generation request, this system automates the generation of structured objectives and autonomously plans the execution plan using AI. The capability execution layer then executes tasks at each node along the specific path to generate the desired content. This method, where a generative language model dynamically generates task execution plans based on user input via natural language, replaces traditional manual predefined workflows. Adding any new input conditions or content requirements eliminates the need for engineers to redesign the workflow, reducing development costs and shortening iteration cycles. Furthermore, the progressive strategy loading mechanism at the strategy organization layer allows for the on-demand loading of relevant strategy organization information based on the current task type, rather than loading all information at once. This mechanism simplifies the context to the minimum set required for the current task, effectively reducing the model's contextual burden, improving decision-making quality, and ensuring the business compliance of the produced content.

[0017] Of course, any product implementing this application does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the system architecture provided in the embodiments of this application; Figure 2 This is a flowchart of the method provided in the embodiments of this application; Figure 3 This is a schematic diagram of the material generation process provided in the embodiments of this application; Figure 4 This is a schematic diagram of progressive loading provided in an embodiment of this application; Figure 5 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] To facilitate understanding of the technical solutions provided in the embodiments of this application and to compare them with the prior art, the following is a brief introduction to the specific implementation methods of the prior art based on engineer-built workflows and their drawbacks.

[0022] Suppose a merchant lists a product on platform A and also needs to distribute it on platform B. However, products listed on platform B must meet platform B's material requirements, including main image size, style, title, and selling points. Therefore, a new material package adapted to platform B's requirements needs to be generated. If the user submits a material generation request with very little information—for example, only an image of the product or a title—a designer needs to build a workflow to meet this specific material generation requirement. This workflow typically corresponds to an execution script that follows a fixed process. For example, to generate an image, editing instructions might be generated first, then the image might be generated, and so on.

[0023] However, if the goal is to generate a resource package for platform B, but the user's input conditions change, for example, by inputting a product link and a sentence, the above-mentioned link is no longer suitable, and a different link needs to be designed to meet this requirement.

[0024] Furthermore, if the products also need to be distributed to the C platform, the material requirements of the C platform are different from those of the B platform. Therefore, for various different input scenarios, it is necessary to build multiple different links for the material generation requirements of the C platform.

[0025] In short, different workflows need to be built for various input information and different material generation goals, resulting in a large number of workflows to be built. Furthermore, after a workflow is built, users (usually merchants) may feel that the materials generated by that workflow do not match their store's style and make specific requests regarding store style. At this point, the already built workflow needs to be modified, thus making the later maintenance costs of the workflow very high.

[0026] In response to the above situation, this application provides an automated e-commerce material production system based on Agentic AI. This system can use a generative language model as the core decision engine, adopt a multi-layer architecture design, and achieve autonomous understanding, autonomous planning and autonomous execution of the entire e-commerce material production process through core mechanisms such as AI autonomous planning and progressive strategy loading (in the preferred mode, it can also include continuous self-optimization).

[0027] See Figure 1 Specific multi-layered structures may include: The Goal Understanding Layer is responsible for addressing the question of "what the user actually wants." The system receives user input regarding material production requirements. This layer supports multimodal input formats, including: a one-sentence natural language description, reference images, product links, a set of reference images, and continuously evolving modification suggestions throughout multiple interactions, etc. At this Goal Understanding Layer, the system can use generative language models as its core to semantically analyze user intent. If missing contextual information exists, it can be automatically filled in, such as querying product information or images based on product IDs. If the input requirement includes non-textual information such as images, it can also call upon appropriate image understanding tools for interpretation, and so on. In short, through this Goal Understanding Layer, vague requirements can be transformed into structured goal descriptions that the system can execute. For example, suppose a user inputs a product link and a sentence: "Generate a set of material packages for this product, suitable for platform A." After parsing by this Goal Understanding Layer, the generated structured goal can be expressed as: "Title + Selling Points + Main Image × 3 + Scene Image × 2."

[0028] Task Planning Layer: In this embodiment, engineers no longer pre-write complete workflows for each input combination and each material target. Instead, AI automatically generates execution plans based on the specifically generated structured targets and can dynamically modify them. Of course, the system can still have toolkits, strategy kits, and reusable historical paths, allowing AI to dynamically plan execution paths within these constraints and capabilities. In other words, the system does not rely on rigid workflows predefined manually for specific material requirements. Instead, it dynamically generates or adjusts execution plans based on user goals, available tools, strategy organization information, and historical reusable paths. For example, for the structured targets in the aforementioned example, the specifically planned execution plan could include: querying product information → generating a title → generating selling point copy → generating a prompt (text for generating images) → generating 3 main images → generating 2 scene images → compliance verification, etc. Compliance verification can also be performed at each node or some nodes to more promptly identify and correct problems. Furthermore, regarding the generation of specific execution plans, for simpler tasks, the AI ​​model can generate plans based on its own knowledge. For more complex tasks, the AI ​​model can acquire relevant professional knowledge, detection rules, execution standards, etc., by loading strategy organization sets to assist in generating the execution plan. Compared to manually predefined workflows, the AI ​​autonomous planning mechanism eliminates the need for engineers to redevelop workflows when adding new material requirements, significantly reducing development costs and enabling the system to flexibly respond to complex scenarios such as changes in input conditions and mid-task adjustments.

[0029] The capability execution layer unifies AI capabilities from different sources and in different forms into a standardized toolset that is schedulable, orchestratable, and scalable. For example, for the material generation scenario in a product information service system, the specific toolset can cover seven categories: image understanding, image generation, information retrieval, copywriting generation, SKU combination, material evaluation, and human-computer interaction, as shown in Table 1. Table 1

[0030] Each tool can be registered in a structured manner, including at least the tool name, tool category, capability description, applicable conditions, input parameter format, required fields, output result format, and exception return format. This allows the task planning layer to select the appropriate tool from the registered tools based on the task objectives of the current node when generating the execution plan, rather than relying on a manually hard-coded call chain. For example, when the current node needs to understand user-uploaded images, it selects an image understanding tool; when it needs to generate main images, scene images, or poster backgrounds, it selects an image generation tool; when it needs to retrieve product information, product images, or SKU information based on the product ID, it selects an information query tool; when it needs to generate titles, selling points, or marketing copy, it selects a copywriting generation tool; when it needs to generate combined sales plans or SKU titles, it selects an SKU combination tool; when it needs to determine whether materials are compliant, aesthetically pleasing, or conform to platform specifications, it selects a material evaluation tool; when it needs user confirmation of intermediate results or supplementary feedback, it selects a human-computer interaction tool, and so on. Before actually invoking a tool, the capability execution layer can organize the invocation parameters according to the tool's input parameter format and check whether required fields, parameter types, number of images, size requirements, platform information, and category information meet the tool invocation conditions. After the tool executes, the system can write the tool's output results into the current session context, such as product information, candidate images, generated materials, copywriting results, quality review reports, and user feedback, for subsequent nodes to use. When a tool invocation fails, returns an empty result, or the tool's output results fail quality verification, the system can feed back the failure reason or quality review result to the task planning layer. The task planning layer can then reorganize the parameters, replace the tool, retry the current node, supplement context information, or adjust the subsequent execution plan accordingly.

[0031] The strategy organization layer primarily provides relevant contextual information, including industry knowledge, platform rules, and category experience, to the target understanding layer and task planning layer. In this embodiment, the specific strategy organization information can be loaded progressively. That is, industry knowledge, platform rules, and category experience can be dynamically loaded based on the current task type, rather than loading all strategy organization information at once, thus simplifying the context content to the minimum set required for the current task. Specifically, the strategy organization information in this embodiment can be managed using hierarchical strategy units. Each strategy unit may not be a single prompt word, but rather consists of "entry description + phased sub-strategies + callable tools / scripts + output format constraints," etc. The system does not load all strategy content at once during runtime, but rather expands it step-by-step: first, it loads the strategy entry information. The system first determines which strategies the current task might need based on the name, capability description, applicable conditions, and invocation timing of each strategy unit. Examples include: platform category specification strategy, copywriting generation strategy, image production strategy, quality review strategy, and layout generation strategy. Sub-strategies are loaded according to task stages. Only the strategy content required for the current stage is loaded at different task stages. For example: intent resolution rules are loaded during the intent resolution phase; task decomposition rules are loaded during the task breakdown phase; strategy scheduling rules are loaded during the execution scheduling phase; retry and degradation rules are loaded when exceptions occur; and delivery format rules are loaded during the delivery phase. Alternatively, in the aforementioned example, assuming a material package for platform A needs to be generated, only the definitions of the specific material package and the requirements for specific individual materials (e.g., image size, resolution) from platform A need to be loaded; strategies from other platforms do not need to be loaded. Furthermore, loading can be further subdivided by sub-task type. When entering a specific task, the system continues to load more detailed strategy content based on the sub-task type. For example, after entering an image production task, not all image rules need to be loaded; instead, they can be subdivided according to image type: main image production strategy is loaded when generating the main image; SKU consistency strategy is loaded when generating the SKU image; poster layout strategy is loaded when generating the marketing poster; detail page module strategy is loaded when generating the detail image; tool selection strategy is loaded only when tool parameters need to be selected; and material package specification strategy is loaded only when generating the complete material package. Cross-cutting strategies, such as platform category specification strategies and quality review strategies, are invoked on demand. These strategies are not limited to a single material production task but are called as needed at stages such as copywriting, image creation, layout, and quality review. For example, before image production, the platform specification strategy is called to obtain size, style, and category visual language; before delivery, the quality review strategy is called to check compliance, consistency between text and images, and visual quality. The loaded strategy content enters the current task context and is used to constrain model generation, tool selection, parameter organization, and quality judgment. Unmatched strategy content does not enter the context, thus avoiding context bloat.A strategy can trigger the loading of new strategies: During the execution of a strategy, other strategies can be triggered as needed by the task. For example, if the image production strategy detects that the current image is a "main image with selling points", it will trigger the copywriting generation strategy to generate short copy, then trigger the layout generation strategy to complete the text overlay, and finally trigger the quality review strategy to check the results.

[0032] In other words, strategy organization information can include multiple hierarchical strategy units, each of which can include entry description information and several sub-strategy information. The entry description information describes the scope of capabilities, applicable conditions, and invocation timing of the strategy unit; the sub-strategy information describes the specific processing rules under different execution stages, different material types, or different abnormal scenarios. After receiving a material generation request, the system can progressively load the corresponding strategy units or sub-strategy information based on the current execution stage and specific requirements such as material type and platform category. Strategy information not related to the current task objective or current execution node is not loaded into the current context. A loaded strategy unit can trigger the loading of other strategy units based on its execution result or intermediate state, forming a multi-level strategy scheduling chain. For example, an image production strategy can trigger a copywriting generation strategy when it identifies the need to overlay selling point copy, a layout generation strategy when generating a layout image, and a quality review strategy after output is complete. In short, a strategy is not a flat rule base, but rather a hierarchical, routable, recursively invoked, and stage-based execution knowledge unit.

[0033] The memory and state layer is responsible for maintaining the complete context state during multi-round task execution, including: user dialogue history, queried product information, currently available image set, reusable intermediate results, user preferences, and current task progress. It can be managed uniformly at the session level using SessionId to ensure uninterrupted multi-step tasks. Furthermore, this memory and state layer can build a closed-loop mechanism for continuous system self-optimization: for each successfully completed end-to-end material production task, the complete decision path can be automatically recorded and offline quality scored. High-quality paths can be refined into reusable business strategies for priority reference in subsequent similar tasks, enabling continuous self-evolution of the system's orchestration capabilities.

[0034] It should be noted that in practical implementation, the various layers mentioned above can be coupled together. For example, in the target understanding layer, strategy organization information can be loaded when necessary, and context information can be reviewed through the memory and state layer; in the task planning layer, relevant strategy organization information can also be loaded, and so on. Furthermore, all the above layers can be implemented by the same AI model. Regarding the capability execution layer, for tasks at each node, if the AI ​​model is capable of completing them, it can do so itself; if its capabilities are insufficient, it can call relevant tools. Specifically, the strategy organization information and tools existing in the system can be provided to the AI ​​model in advance, along with the names and functional descriptions of the tools, so that the model can load or call them when needed.

[0035] The specific implementation schemes provided in the embodiments of this application will be described in detail below.

[0036] First, this application provides a method for generating materials, see [link to relevant documentation]. Figure 2 The method may specifically include: S201: After receiving the user's material generation request and determining the associated user input condition information, the following steps S201 to S204 are processed through the generative language model.

[0037] In this embodiment of the application, when a user needs to generate materials, the specific input conditions can be expressed in various ways, and the specific types of input information can also be various. For example, as mentioned above, it can include text expressed in natural language, as well as images, links, dialogue content, and so on.

[0038] S202: Based on the input condition information, understand the user's intent and generate a structured expression of the material generation target.

[0039] After receiving the input condition information, the specific material generation process can begin. Specifically, firstly, a generative language model (a type of AI model primarily providing text understanding and text generation capabilities) can be used as the core for understanding user intent. During this process, if the input condition information contains information in other modalities besides text, other modal AI models can be invoked for auxiliary understanding. Furthermore, if the input condition information reveals missing contextual information, relevant tools can be used to complete the contextual information. For example, suppose the input condition information expresses the need to generate a material package for a certain platform based on a product image, but only provides a product link or ID, relevant query tools can be used to query information such as the product image. After completing the contextual information completion, structured target information is generated. In other words, when understanding user intent based on the input condition information, if it is determined that there is missing contextual information, a pre-packaged query tool is invoked to retrieve the information and complete the missing contextual information, so that a structured expression of the material generation target can be generated with the completed contextual information.

[0040] Specifically, structured target information means expressing the generation target that the user vaguely expresses in the input conditions in a more clear, accurate, and system-executable form, so as to improve the quality of subsequent material generation.

[0041] S203: Based on the input condition information and the structured expression of the material generation target, generate an execution plan for the material generation target.

[0042] After generating structured target information, a corresponding execution plan can be generated based on this information. This means determining which nodes the generation process involves and what specific tasks need to be performed at each node. In practice, before generating the execution plan, if necessary, specific strategy organization information can be loaded first for reference during plan generation. Specifically, when loading strategy organization information, a progressive loading approach can be adopted; that is, only the necessary strategy organization information is loaded based on the current task, while other information not needed at this stage is not loaded.

[0043] After the strategy organization information is loaded, the model can generate an execution plan based on the various knowledge types it contains. It's important to note that the specific strategy organization information may include a specific execution plan saved for the current structured target. In this case, the path saved in the strategy organization information can be directly used as the execution plan for the current task. If the strategy organization information does not directly include a specific execution plan saved for the current structured target, the generative language model can generate each node in the execution plan one by one.

[0044] S204: Execute material generation tasks according to the multiple nodes included in the execution plan to achieve the material generation target, wherein at least some of the tasks on the nodes are completed by calling pre-packaged tools. During the execution of each step, the required strategy organization information is progressively loaded on demand according to the task target, and the corresponding task target is completed by understanding the loaded strategy organization information.

[0045] After generating the execution plan, tasks at each node can be executed through the capability execution layer. Specifically, for each node's task, if the currently used generative language model has the corresponding capability, it can be executed directly by that model; otherwise, tools with the corresponding capability can be called. For example, a generative language model might have the ability to generate prompts for image generation models. In this case, the generative language model can directly generate the prompt. Of course, if the specific task is more complex, corresponding tools can be called to complete the task of generating raw image prompts. For tasks such as generating titles, generating selling point copy, generating main images, and generating scene images, since these tasks are often more specialized, they are usually achieved by calling corresponding tools. Of course, as generative language models continue to upgrade and develop, and as the system's internal self-optimization progresses, generative language models may acquire more capabilities. The specific implementation depends on the actual situation and is not limited here. Furthermore, as long as the capability execution layer connects to the corresponding tools, the system can expand to more material types through task planning and tool scheduling, such as product videos, video covers, multi-product combination material packages, and multi-platform style-adapted materials. When a new material type is added, the system generates an execution node corresponding to that material type through the task planning layer and calls the corresponding tool to complete the material generation, without the need to manually define a fixed workflow for that material type in advance.

[0046] Specifically, the content generation process involves multiple nodes in the execution plan. Each node may generate a portion of the content; for example, one node might generate title text, another might generate selling point copy, and so on. In practice, specific strategy organization information can be loaded as needed when executing tasks at each node. For instance, when generating a main image suitable for a particular platform, the required strategy organization information might include the platform's main image specifications (including size and resolution), the platform's main image style strategy (e.g., some platforms prefer clean white backgrounds as main images, while others prefer to include various selling point copy in the main image), and so on. Furthermore, quality verification nodes can be set at multiple nodes to perform quality checks on the content during generation, including whether the generated images comply with watermark compliance rules, aesthetic ratings, etc. (The rules and standards used in the verification process can also come from the strategy organization layer). If the quality of the content generated at a certain node is found to be unsatisfactory, the execution plan can be adjusted, and generation can be restarted.

[0047] The specific generated materials can correspond to the generation goal. For example, in the aforementioned example, if a product material package suitable for publication on a certain platform needs to be generated, the specific generated materials can include title text, selling point copy, main image, scene image, etc. Specifically, the title text could be: "Cloth-covered retro notebook with lined inner pages, wire-bound, non-diary notebook, available in dancing girl / squirrel / bunny / pony designs"; the selling point copy could include: "Cloth-covered feel, retro style, sturdy binding, vintage cloth cover, durable binding, high-value notebook, vintage cloth cover, binding won't fall apart, easy to write with lines," and so on.

[0048] It's worth noting that the system can create sessions between users and the AI. Users can initiate content generation requests and receive the generated content through dialogue with the AI ​​within these sessions. In practice, the system can also save context state information from multiple rounds of task execution within the same session. This allows for querying the current context state information within the session using the session ID during user intent understanding and / or execution plan generation.

[0049] In addition to dynamic path adjustments based on material verification results, the model can also adjust the path based on user feedback. Specifically, since users can interact with the AI ​​through system-created conversations, and relevant materials can be returned to the user in the conversation window, users can provide feedback on their satisfaction with the generated materials while receiving them. This feedback can be expressed using natural language. If the user's feedback is negative, meaning they are dissatisfied with the generated materials—for example, after generating the main image, although it may meet a platform's requirements, the user still needs it to match their store's unified style—the user can suggest modifications in the conversation interface. For instance, the user can type in the conversation window: "The generated main image doesn't quite match my store's style." The AI ​​model can then regenerate the task execution path based on the user feedback and the previously obtained strategy organization information, and regenerate the main image. If additional strategy organization information or other response information needs to be obtained, this can also be done through the strategy organization layer or capability execution layer. For example, a user might need the generated main image to match their store's style, but without specifying a unified style for all the main images in their store. Therefore, a query tool can be used to retrieve the main images of multiple products in the user's store. Then, by using image understanding tools, style information can be extracted from these main images, and the unified style can be summarized. For instance, if the store uses model images as main images, and previously generated images were flat lay images of clothing, then when adjusting the execution plan, the previous "Generate Main Image" node can be split into two nodes: "Generate Clothing Flat Lay Image" and "Generate Model Image as Product Main Image," and then the material can be regenerated.

[0050] After completing the material generation task, the system can also obtain quality score information on the generated materials. Execution plans that meet the quality score criteria are then added to the strategy organization information set as reusable strategy organization information. In other words, for each successfully completed end-to-end material production task, the system can automatically record the complete decision-making execution plan and perform offline quality scoring. High-quality paths can be extracted into reusable business strategies and prioritized for reference in subsequent similar tasks. (For the same or similar structured objectives, this execution plan can be directly reused or fine-tuned to improve efficiency and generation quality), enabling continuous self-evolution of the system's orchestration capabilities.

[0051] To facilitate understanding of the solutions provided in the embodiments of this application, the following will illustrate the methods described below. Figure 3 The example shown illustrates the specific implementation process. See also... Figure 3 First, the system can receive user input, including sentences in natural language, images, or product links. Then, a generative language model can perform intent parsing, including identifying the task type, filling in missing contextual information, and generating a structured objective. After generating the structured objective, AI can dynamically generate an execution plan, specifically, task decomposition, step sequencing, dependency determination, and so on. This process requires no manual pre-definition; it is generated autonomously by AI. The specific execution plan may vary depending on the specific input conditions. For example, if the input is a product ID, the execution path might include: querying product information → querying product images → generating title text → generating selling point copy → generating images (including main images, scene images, etc.). If the input includes images and instructions, the execution path might include: image content understanding → determining task type → selecting tools → generating results. If insufficient information is found during the generation process, the current step can be paused, and missing information can be filled in by calling tools, revising the execution plan, and then continuing execution. Finally, the materials can be generated, and users can confirm the materials. If they feel that the materials do not meet the requirements, they can point it out through dialogue. The AI ​​will then replan the task path and regenerate the materials. After multiple iterations, materials that meet the user's requirements can be generated.

[0052] In this approach, information can be loaded on demand, rather than in full, based on the specific tasks at each node in the execution plan. For example, ... Figure 4 As shown, assuming a node's task is to generate a marketing poster for platform A, it can load platform A's style specifications, marketing poster layout strategies, etc. If a node's task is to generate images of clothing models wearing the garment, it can load clothing category specifications, model wearing process, and image processing strategies, etc. If a node's task is to optimize product main images for platform B, it can load platform B's main image specifications, compliance verification rules, and aesthetic scoring rules, etc. If a node's task is to perform general image editing, it can skip loading e-commerce-related rules and directly execute user commands. This progressive loading method achieves simplified context, high-quality decision-making, guaranteed compliance, and flexible expansion: if new specifications, rules, standards, etc., are added directly to the strategy organization information set, and specific names and other information are provided to the AI ​​model for loading when needed.

[0053] In summary, this application provides an automated e-commerce content production system based on Agentic AI. This system, upon receiving a content generation request, can automatically generate structured objectives and autonomously plan the execution plan using AI. The capability execution layer then executes tasks at each node along the specific path to generate the desired content. This method, where a generative language model dynamically generates task execution plans based on user input via natural language, replaces traditional manually predefined workflows. Adding any new input conditions or content requirement types eliminates the need for engineers to redesign the workflow, significantly reducing development costs and shortening iteration cycles. Furthermore, the progressive strategy loading mechanism of the strategy organization layer allows for the on-demand loading of corresponding strategy organization information based on the current task type, rather than loading all information at once. This mechanism simplifies the context to the minimum set required for the current task, effectively reducing the model's contextual pressure, improving model decision quality, and ensuring the business compliance of the produced content.

[0054] In addition, by using a generative language model as its core and combining it with other unified encapsulated tools, it can flexibly respond to multimodal conditional inputs such as "a sentence, an image, or a product link," and can also adjust the execution path in real time based on user feedback during multiple rounds of interaction, thus breaking free from the limitations of static predefined workflows.

[0055] Furthermore, through the continuous self-optimization mechanism of memory and state layers, the complete decision-making path (including tool call sequence, intermediate results, final output, etc.) can be automatically recorded for each successfully completed end-to-end material production task, and offline quality evaluation can be performed. High-quality paths are refined into reusable business strategies and given priority reference in subsequent similar tasks, realizing the continuous self-evolution of the system's orchestration capabilities and making the system "smarter" the more it is used. In other words, the solution provided in this application embodiment realizes the paradigm upgrade of e-commerce material production from "manual predefined workflow + AI execution" to "AI autonomous planning + continuous learning optimization," enabling the system to automatically produce all types of e-commerce materials, including product titles, selling point copy, product main images, scene images, marketing posters (multi-platform style adaptation), SKU images, product detail images, and product videos, with a relatively high end-to-end task completion rate, improving the efficiency and flexibility of e-commerce material production.

[0056] It should be noted that the embodiments of this application may involve the use of user data. In practical applications, user-specific personal data may be used in the scheme described herein within the scope permitted by applicable laws and regulations, provided that it complies with the applicable laws and regulations of the country (e.g., with the user's explicit consent, with the user being properly notified, etc.).

[0057] Corresponding to the foregoing method embodiments, this application also provides a material generation apparatus, which may include: The request receiving unit receives the material generation request submitted by the user, determines the associated user input conditions, and then performs the following steps through the generative language model: The structured target generation unit is used to understand user intent based on the input condition information and generate structured material generation targets. An execution plan generation unit is used to generate an execution plan for the material generation target based on the input condition information and the structured material generation target. An execution unit is configured to execute material generation tasks according to multiple nodes included in the execution plan to achieve the material generation target; wherein, at least some of the tasks on the nodes are completed by calling pre-packaged tools; In the process of executing each step, the required strategy organization information is progressively loaded on demand according to the task objective, and the corresponding task objective is completed by understanding the loaded strategy organization information.

[0058] In a specific implementation, the device may further include: The information completion unit is used to, when performing user intent understanding based on the input condition information, if it is determined that there is a lack of context information in the input condition information, then by calling a pre-encapsulated query tool to obtain information and complete the missing context information, so as to generate a structured expression material generation target in the state of complete context information.

[0059] Additionally, the device may also include: The image understanding model invocation unit is used to invoke the image understanding model to perform intent understanding on the input condition information when performing user intent understanding based on the input condition information. If the input condition information includes image-type information, then the image understanding model is invoked to perform intent understanding on the input condition information.

[0060] The quality verification unit is used to perform quality verification on the generated results at the corresponding node by loading strategy organization information related to quality verification during the execution of the material generation task. If the results do not meet the standards, the execution plan is dynamically adjusted to trigger the regeneration of the corresponding material.

[0061] Furthermore, the device may also include: The session information storage unit is used to store the context state information during multiple rounds of task execution in the same session; The query unit is used to query the context state information by session ID during the user intent understanding and / or execution plan generation process, so as to combine the context state information to perform user intent understanding and / or execution plan generation.

[0062] The feedback information receiving unit is used to dynamically adjust the execution plan based on the feedback information after receiving the material generation results of one or more nodes through the session and receiving feedback information from the user on the material generation results. If the feedback information is negative feedback, the unit will regenerate the material on the corresponding node.

[0063] Additionally, the device may also include: The quality score information acquisition unit is used to acquire quality score information about the material generation results after the material generation task is completed, and to add the execution plan that meets the quality score conditions as reusable strategy organization information to the strategy organization information set.

[0064] The user input condition information includes: generating a material package for the target product to be published in the target system, wherein the information of the target product in the user input condition information is expressed through a product details page link and / or product ID; At this point, the structured target generation unit can specifically be used for: Based on the input condition information, user intent is understood, and by loading the strategy organization information related to the definition of material packages in the target system, a structured material generation target is generated. The structured material generation target includes: title text, selling point copy, main image, and scene image. The nodes in the execution plan include: obtaining product information, generating title text, generating selling point copy, generating prompt text for image generation, generating main image, and generating scene image; The task execution unit can specifically be used for: Based on the material generation tasks at multiple nodes, the necessary strategy organization information is progressively loaded to complete the corresponding material generation tasks.

[0065] In addition, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in any of the foregoing method embodiments.

[0066] And an electronic device, comprising: One or more processors; and A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of the method described in any of the foregoing method embodiments.

[0067] A computer program product includes a computer program / computer executable instructions that, when executed by a processor in an electronic device, implement the steps of the method described in the foregoing method embodiments.

[0068] in, Figure 5 An exemplary architecture of an electronic device is shown, which may include a processor 510, a video display adapter 511, a disk drive 512, an input / output interface 513, a network interface 514, and a memory 520. The processor 510, video display adapter 511, disk drive 512, input / output interface 513, network interface 514, and memory 520 can communicate with each other via a communication bus 530.

[0069] The processor 510 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solution provided in this application.

[0070] The memory 520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 520 can store the operating system 521 for controlling the operation of the electronic device 500, and the basic input / output system (BIOS) for controlling the low-level operations of the electronic device 500. Additionally, it can store a web browser 523, a data storage management system 524, and a material generation and processing system 525, etc. The aforementioned material generation and processing system 525 can be the application program that specifically implements the aforementioned steps in this embodiment. In summary, when implementing the technical solution provided in this application through software or firmware, the relevant program code is stored in the memory 520 and executed by the processor 510.

[0071] Input / output interface 513 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0072] Network interface 514 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0073] Bus 530 includes a pathway for transmitting information between various components of the device, such as processor 510, video display adapter 511, disk drive 512, input / output interface 513, network interface 514, and memory 520.

[0074] It should be noted that although the above-described device only shows the processor 510, video display adapter 511, disk drive 512, input / output interface 513, network interface 514, memory 520, bus 530, etc., in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the solution of this application, and does not necessarily include all the components shown in the figures.

[0075] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0076] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0077] The above provides a detailed description of the material generation method and electronic device provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for generating materials, characterized in that, include: After receiving the user's request to generate materials and determining the associated user input information, the following steps are performed using a generative language model: Based on the input condition information, the user intent is understood, and a structured expression of the material generation target is generated; Based on the input condition information and the structured expression of the material generation target, an execution plan is generated for the material generation target; The material generation task is executed according to multiple nodes included in the execution plan to achieve the material generation goal; wherein, at least some of the tasks on the nodes are completed by calling pre-encapsulated tools; In the process of executing each step, the required strategy organization information is progressively loaded on demand according to the task objective, and the corresponding task objective is completed by understanding the loaded strategy organization information.

2. The method according to claim 1, characterized in that, Also includes: When understanding user intent based on the input condition information, if it is determined that there is a lack of context information in the input condition information, the information is obtained by calling a pre-encapsulated query tool to complete the missing context information, so as to generate a structured expression material generation target in the state of complete context information.

3. The method according to claim 1, characterized in that, Also includes: When performing user intent understanding based on the input condition information, if the input condition information includes image-related information, then the image understanding model is invoked to perform intent understanding on the input condition information.

4. The method according to claim 1, characterized in that, Also includes: During the execution of the material generation task, the generation results at the corresponding node are verified by loading strategy organization information related to quality verification. If the results do not meet the standards, the execution plan is dynamically adjusted to trigger the regeneration of the corresponding materials.

5. The method according to claim 1, characterized in that, Also includes: Save context state information during multiple rounds of task execution within the same session; During the process of understanding user intent and / or generating execution plans, the context state information is queried through the session ID so as to combine the context state information for understanding user intent and / or generating execution plans.

6. The method according to claim 5, characterized in that, Also includes: After the session returns the result of generating materials on one or more nodes and receives feedback from the user on the result, if the feedback is negative, the execution plan is dynamically adjusted according to the feedback to regenerate materials on the corresponding node.

7. The method according to claim 5, characterized in that, Also includes: After completing the material generation task, obtain the quality score information of the material generation results, and add the execution plans that meet the quality score conditions as reusable strategy organization information to the strategy organization information set.

8. The method according to any one of claims 1 to 7, characterized in that, The user input condition information includes: generating a material package for the target product to be published in the target system, wherein the information of the target product in the user input condition information is expressed through a product details page link and / or product ID; The step of understanding user intent based on the input condition information and generating a structured expression of the material generation target includes: Based on the input condition information, user intent is understood, and by loading the strategy organization information related to the definition of material packages in the target system, a structured material generation target is generated. The structured material generation target includes: title text, selling point copy, main image, and scene image. The nodes in the execution plan include: obtaining product information, generating title text, generating selling point copy, generating prompt text for image generation, generating main image, and generating scene image; The execution of the material generation task according to the multiple nodes included in the execution plan includes: Based on the material generation tasks at multiple nodes, the necessary strategy organization information is progressively loaded to complete the corresponding material generation tasks.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1 to 8.

10. An electronic device, characterized in that, include: One or more processors; as well as A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of the method according to any one of claims 1 to 8.

11. A computer program product comprising a computer program / computer-executable instructions, characterized in that, When the computer program / computer executable instructions are executed by a processor in an electronic device, they implement the steps of the method according to any one of claims 1 to 8.

Images