Automatic three-dimensional scene rehearsal method, device and equipment based on text script

Abstract

The invention relates to the technical field of artificial intelligence and natural language processing, and provides an automatic three-dimensional scene preview method, device and equipment based on a text script, and the method comprises the steps: a system receives text script data, starts a script semantic analysis engine to carry out structured processing, and extracts a DSD object; according to the content in the DSD object, a three-dimensional scene basic environment is automatically constructed in a three-dimensional engine environment through semantic-visual mapping and programmed generation rules; and in the three-dimensional scene basic environment, automatically deploying roles, props and interaction elements according to the key information and a preset layout rule, and generating a three-dimensional rehearsal scene. According to the method, the technical problems of low efficiency, high cost and unsmooth communication of a traditional manufacturing process are solved, and efficient production from texts to three-dimensional scenes is realized.

Description

Technical Field

[0001] The embodiments of the present invention generally relate to the fields of artificial intelligence and natural language processing technology, and particularly to methods, apparatus and devices for automated 3D scene pre-visualization based on text scripts. Background Technology

[0002] Digital content industries such as film, animation, and games are core components of the modern cultural and creative industries, with complex and interconnected production processes. In the pre-production stage, converting abstract text scripts into intuitive 3D visualizations is crucial. This helps the creative team anticipate and plan shots, staging, pacing, and scene layout before investing significant resources, effectively avoiding costly rework in post-production. However, traditional script visualization processes suffer from significant technical bottlenecks. They heavily rely on manual operation, typically requiring creators to manually locate assets, build scenes, set lighting, place characters, and create animations in 3D software based on the script description—a time-consuming, labor-intensive, and costly process. When the script is revised, adjustments to the 3D scenes often require starting from scratch, leading to long iteration cycles and severely impacting creative efficiency. Furthermore, communication between different departments often lacks a unified visual reference, easily causing misunderstandings. While existing technologies have made progress in 3D asset management and rendering, they have not resolved the semantic gap between unstructured text scripts and structured 3D scenes, lacking a complete technical solution that can intelligently complete the transformation from text to 3D scenes. Summary of the Invention

[0003] To address the above issues, this invention deeply integrates natural language processing and content generation technologies to automatically parse and map semantic information in the script into 3D visual elements. By establishing automated rules encompassing scene construction, character layout, and lighting atmosphere, it rapidly generates interactive 3D pre-production scenes. Furthermore, by introducing a large language model to construct a feedback loop, it intelligently transforms human review comments into revised script drafts and drives real-time updates to the 3D scenes. This solves the technical problems of low efficiency, high costs, and poor communication in traditional production processes, achieving highly efficient production from text to 3D scenes.

[0004] According to embodiments of the present invention, a method, apparatus, and device for automated 3D scene pre-visualization based on text scripts are provided.

[0005] In a first aspect of the invention, a method for automated 3D scene pre-visualization based on a text script is provided. The method includes: Step S01: The system receives text script data, starts the script semantic parsing engine for structured processing, and extracts DSD objects; Step S02: Based on the content in the DSD object, automatically construct the basic environment of the 3D scene in the 3D engine environment through semantic-visual mapping and procedural generation rules; Step S03: In the basic 3D scene environment, based on key information and preset layout rules, automatically deploy characters, props and interactive elements to generate a 3D pre-visualization scene.

[0006] Furthermore, the specific steps of step S01 are as follows: Step S011: The system receives the text script file uploaded by the user and performs text preprocessing on the script content: it determines the format of the text script file. If it is a markup language-based format, it calls the corresponding XML parser to extract the preset tag content, maps and constructs a DSD object. Step S012: If it is plain text format, the script content in plain text format is standardized; a pattern matching method based on regular expressions is used to perform semantic block segmentation on the standardized plain text content. According to script writing conventions, scene title blocks, action description blocks, character name blocks and dialogue blocks are identified and separated. Sentence segmentation is performed on the segmented action description blocks to generate preprocessed text units in sentences. Step S013: Initiate the deep learning-based entity recognition process, perform entity recognition and annotation on the preprocessed text units, input the text units one by one into the NER model to traverse each word in the text, and output a predefined entity label; the system merges the predefined entity label sequence output by the model and extracts the complete entity from the text; the predefined entity label includes at least: scene title, interior and exterior scene identifiers, scene location, time identifier, character name, action description, key props, and atmosphere keywords; Step S014: Start the relation extraction process based on syntactic analysis, analyze the logical relations between the identified entities, and for sentences containing at least two identified entities, call the dependency parsing engine to generate the dependency relation tree of the sentence. The dependency relation tree reveals the grammatical subordination relationship between words in the sentence. Step S015: Define a set of relation extraction rule templates, match specific syntactic structure paths on the dependency relation tree to identify semantic relations between entities, and store the successfully matched entities and their relations in the form of structured subject-verb-object triples; the semantic relations include at least: the execution relationship between roles and actions, the domination relationship between actions and props, and the interaction relationship between roles. Step S016: Integrate the entities and logical relationships to generate and output a hierarchical DSD object.

[0007] Furthermore, the DSD object adopts a key-value pair data format, and its top-level structure is a list of scenes organized in the original order of the script. Each scene object in the list contains a unique scene ID and a structured parsing result of the scene title entity. The structured analysis results include interior and exterior scene identifiers, location, time, and a list of atmosphere keywords extracted from the scene description; Each scene description contains a content sequence consisting of multiple content block objects arranged chronologically. The content block objects are defined as character action blocks or character dialogue blocks based on their properties. The character action block contains the character entity performing the action, the original text description of the action, and a list of relation triples related to the action. The character dialogue block contains the character entity speaking, the text content of the dialogue, and character performance or emotion cues extracted from parentheses.

[0008] Furthermore, the specific steps of step S02 are as follows: Step S021: The system iterates through the scene list in the DSD object and automatically creates an independent 3D level named after its scene title in the 3D engine environment for each scene object in the list. Step S022: Call a pre-built 3D asset library. Each 3D model in the asset library is labeled with multi-dimensional metadata tags and associated with a high-dimensional feature vector generated by a pre-trained multimodal model, representing its visual and semantic features. Step S023: Start the semantic-visual mapping engine, extract scene location and atmosphere keywords from the current scene object, and use a multimodal model to convert the keyword combination into a query semantic vector; Step S024: By calculating the cosine similarity between the query semantic vector and the feature vectors of all models in the asset library, the assets are ranked according to their matching degree, and the main environmental model and a series of related prop models with a matching degree higher than the preset threshold are retrieved. Step S025: Automatically deploy the retrieved main environmental model to the newly created 3D level as the basic scene, and automatically arrange the retrieved prop models according to the preset procedural content generation rules and a rule library containing three types of rules: physical and spatial constraints, semantic association and functional layout, and context modifiers, to complete the initial construction of the scene.

[0009] Furthermore, the physical and spatial constraint rules described in step S025 are used to define the physical relationships between props, including ensuring that the bottom of the prop model automatically fits the top of the supporting surface model, avoiding mutual penetration between models through bounding box detection, and preserving unobstructed passages in the scene based on pathfinding grid data that conform to the character's movement logic. The semantic association and functional layout rules described in step S025 are used to perform clustered layout based on the functional attributes of props. These rules define the parent-child and companion relationships between different categories of props. The context modifier rules described in step S025 are used to dynamically adjust the layout details and status of props based on the atmosphere keywords extracted from the digital scene description object.

[0010] Furthermore, the specific steps of step S03 are as follows: Step S031: Based on the content block object of the current scene in the DSD object, retrieve and load the corresponding 3D character model from the 3D asset library into the 3D level; Step S032: Activate the cinematic layout engine and analyze the core interaction types of the current scene based on the content sequence in the DSD object, including at least: dialogue interaction, action conflict, or single-player scene; Step S033: Based on the analyzed interaction type, the cinematic layout engine selects and applies a corresponding layout strategy from a preset camera language rule library to calculate and determine the initial position and orientation of each character model in the 3D level; the camera language rule library includes at least: a 180-degree axis rule for dialogue interaction, which automatically establishes a virtual dialogue axis connecting the dialogue characters and places the character models on both sides of the axis to form a confrontation or communication layout; and a rule of thirds composition for single-person or environmental interaction scenes, which places the character on the key point of interest or the golden ratio line of the scene; Step S034: Parse the action description text within the character action block in the DSD object, and retrieve the most matching animation clip from a preset animation library containing standard animation clips by keyword matching or semantic similarity calculation. Step S035: Assign the retrieved animation clip to the corresponding character model and set it as the character's initial idle animation, or set it as an interactive animation that can be triggered by a specific event; Step S036: Based on the relation triplet information stored in the DSD object, establish automated interaction logic between the character model and the prop model, thereby realizing automated interaction pre-play between the character and the prop, i.e., 3D pre-play scene.

[0011] Furthermore, it also includes step S04: the system provides interactive review and context-bound feedback functions, and generates a revised script by calling the large language model through API based on the feedback information; Specifically, in the 3D pre-visualization scene, a first-person or third-person perspective camera that can be freely controlled is provided to the user, allowing the user to review the automatically constructed scene, character layout and lighting atmosphere from different angles in different ways; The system provides a user-activated feedback mode that switches from scene roaming control to feedback pointer control, allowing users to precisely locate the area or object they intend to point to on the 3D rendered screen. When a user clicks on the scene in feedback mode, the system performs a ray projection calculation from the screen's 2D coordinates to the 3D world coordinates to determine the specific 3D spatial location of the user's click and to identify whether there is an interactive 3D model object at that location. The system provides a user input interface, allowing users to input text-based annotations or modification suggestions for the current location point or identified objects. After the user submits the annotation, the system automatically captures and generates a structured context-bound feedback package. The feedback package contains at least the following data: the text annotation content entered by the user, the 3D world coordinates calculated by ray projection, the unique identifier ID of the object if the user inputs it, and the camera position parameters of the current user review camera, including position, rotation, and field of view. Based on the 3D spatial position or object ID recorded in the feedback package, the system traces back and locates the original script paragraph or entity in the DSD object that generated the region or object. The system constructs an input prompt for a large language model, which integrates the following information: the located original script paragraph as the basic context, the user's input text annotations as modification instructions, and the text description of the 3D scene context as supplementary context. The prompt will be sent to a common large language model via the API interface, and the model will be requested to generate a revised script text that conforms to the script format based on all the information provided, and the text will be received as the revised script. Step S05: Based on the updated key information in the revised script and the 3D pre-visualization scene, achieve closed-loop iterative optimization from text to visual scene; The specific steps are as follows: Step S051: The revised script returned by the large language model is presented to the user in the user interface. The interface displays the original script paragraphs before revision and the suggested text after revision side by side for the user to compare, confirm or manually edit. Step S052: After the user confirms and accepts the revised script, the system performs a difference analysis on the text before and after the revision, accurately compares the text changes between the original script paragraphs and the final confirmed revised draft, and identifies all modified, added or deleted text content. Step S053: The system inputs the difference content into the script semantic parsing engine of step S01, and performs semantic parsing on the changed parts to generate new entities and relationships, or update or delete existing entities and relationships. Step S054: Based on the results of the re-parsing, update the corresponding data structure in the DSD object to ensure that the DSD object is fully synchronized with the latest script content; Step S055: The system starts the dependency tracking mechanism, analyzes the changes between the updated DSD object and the original DSD object, and determines all 3D scene elements affected by this update based on the preset dependency graph. These elements include, but are not limited to, scene models, prop layouts, lighting parameters, or character animations. Step S056: For all identified affected 3D scene elements, the system triggers a partial or complete scene reconstruction process; if the change only involves prop positions or lighting parameters, a partial update is performed; if the change involves core elements such as scene locations, a complete reconstruction of the entire 3D level is performed. Step S057: The reconstruction process will call the corresponding steps in steps S02 and S03, and automatically regenerate or adjust the affected 3D scene parts using the data in the updated DSD object. Step S058: After the reconstruction is completed, the updated 3D pre-visualization scene is presented to the user in real time, thus completing a complete closed-loop iteration from 3D scene feedback to text script modification, and then to automatic 3D scene update. The user continues to review the new scene and repeat this process until a satisfactory pre-visualization effect is obtained.

[0012] In a second aspect of the invention, an apparatus for automated 3D scene pre-visualization based on a text script is provided. The apparatus includes: Text script processing module: Used for the system to receive text script data, start the script semantic parsing engine for structured processing, and extract DSD objects; 3D Scene Basic Environment Module: Used to automatically construct the basic 3D scene environment in the 3D engine environment based on the content in the DSD object through semantic-visual mapping and procedural generation rules; 3D Preview Scene Module: Used to automatically deploy characters, props, and interactive elements in a 3D scene environment based on key information and preset layout rules, generating a 3D preview scene.

[0013] In a third aspect of the invention, an electronic device is provided. The electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the program to implement the method according to a first aspect of the invention.

[0014] In a fourth aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method according to a first aspect of the invention.

[0015] This invention, through the deep integration of natural language processing and content generation technologies, automatically parses and maps semantic information in the script into three-dimensional visual elements; by establishing automated rules including scene construction, character layout, and lighting atmosphere, it quickly generates interactive three-dimensional pre-visualization scenes; and by introducing a large language model to build a feedback loop, it intelligently transforms human review opinions into script revisions and drives the real-time updating of three-dimensional scenes, solving the technical problems of low efficiency, high cost, and poor communication in traditional production processes, and achieving efficient production from text to three-dimensional scenes.

[0016] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0017] The beneficial effects of this invention are: 1. This invention deeply integrates natural language processing technology with content generation technology, solving the semantic gap problem between unstructured creative text and structured three-dimensional scenes in the traditional process, and realizing the intelligent translation of script intent; 2. This invention can directly generate a complete 3D pre-visualization scene from a text script, including scene environment, prop layout, lighting atmosphere and preliminary character arrangement, which greatly shortens the manual production time, greatly improves the efficiency of content production and significantly reduces production costs. 3. This invention constructs a closed-loop iterative system from visual feedback to text revision by introducing a large language model, allowing creators to intuitively provide modification suggestions in a 3D environment and quickly obtain intelligent script revisions and synchronously updated 3D scenes, thus achieving rapid creation and efficient collaboration. 4. This invention can be widely applied to fields such as film and television pre-production, game level design, and virtual shooting. By providing real-time interactive visual references, it effectively avoids communication barriers caused by misunderstandings and significantly reduces the risk of rework in post-production. Attached Figure Description

[0018] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart of an automated 3D scene pre-visualization method based on a text script, according to an embodiment of the present invention, is shown. Figure 2 A system architecture diagram according to an embodiment of the present invention is shown; Figure 3 A system operation flowchart according to an embodiment of the present invention is shown; Figure 4 A preliminary rendering of a scene based on an embodiment of the present invention is shown; Figure 5 A complete 3D pre-visualization scene rendering is shown according to an embodiment of the present invention; Figure 6 A block diagram of an apparatus for automated 3D scene pre-visualization based on a text script, according to an embodiment of the present invention, is shown. Figure 7 A schematic diagram of an automated 3D scene pre-playing device based on a text script, according to an embodiment of the present invention, is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] According to embodiments of the present invention, a method, apparatus, and device for automated 3D scene pre-visualization based on text scripts are proposed. By deeply integrating natural language processing technology and content generation technology, the semantic information in the script is automatically parsed and mapped into 3D visual elements. By establishing automated rules including scene construction, character layout, and lighting atmosphere, interactive 3D pre-visualization scenes are quickly generated. A large language model is introduced to construct a feedback loop, which intelligently transforms human review opinions into script revisions and drives the real-time updating of 3D scenes. This solves the technical problems of low efficiency, high cost, and poor communication in traditional production processes, and realizes efficient production from text to 3D scenes.

[0021] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0022] Figure 1 This is a schematic flowchart of an automated 3D scene pre-visualization method based on a text script, according to an embodiment of the present invention. The method includes: Step S01: The system receives text script data, starts the script semantic parsing engine for structured processing, and extracts DSD objects; Step S02: Based on the content in the DSD object, automatically construct the basic environment of the 3D scene in the 3D engine environment through semantic-visual mapping and procedural generation rules; Step S03: In the basic 3D scene environment, based on key information and preset layout rules, automatically deploy characters, props and interactive elements to generate a 3D pre-visualization scene.

[0023] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0024] To provide a clearer explanation of the above-described method for automated 3D scene pre-visualization based on text scripts, a specific embodiment will be used for illustration below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0025] The following example will further illustrate the method of automated 3D scene pre-visualization based on text scripts: Step S01: The system receives text script data, starts the script semantic parsing engine for structured processing, and extracts DSD objects.

[0026] The specific steps are as follows: Step S011: The system receives the text script file uploaded by the user and performs text preprocessing on the script content: it determines the format of the text script file. If it is a markup language-based format, it calls the corresponding XML parser to extract the preset tag content such as scene, character, and dialogue, and directly maps and constructs the DSD object required for subsequent steps based on the preset tag content.

[0027] Step S012: If the script is in plain text format, the script content is standardized, including removing redundant blank lines and headers / footers, unifying full-width / half-width characters, and standardizing line breaks to a uniform standard. A regular expression-based pattern matching method is used to semantically segment the standardized plain text content. Based on scriptwriting conventions, scene title blocks, action description blocks, character name blocks, and dialogue blocks are identified and separated. Sentence segmentation is performed within the segmented action description blocks to generate preprocessed text units in sentence form.

[0028] Step S013: Start the deep learning-based entity recognition process, perform entity recognition and annotation on the preprocessed text units, load a Named Entity Recognition (NER) model based on the Transformer architecture and fine-tuned with a large-scale script corpus. The model can understand the specific language habits and entity patterns in the field of film and television scripts. Input the text units one by one into the NER model. The model traverses each token in the text and outputs a predefined entity label. The system merges the predefined entity label sequence output by the model and extracts the complete entity from the text.

[0029] Predefined entity tags should include at least: scene title, interior / exterior scene identifiers, scene location, time identifier, character name, action description, key props, and atmosphere keywords.

[0030] Step S014: Start the relation extraction process based on syntactic analysis, analyze the logical relationships between the identified entities, and for sentences containing at least two identified entities, call the dependency parsing engine to generate the dependency tree of the sentence. The dependency tree reveals the grammatical subordination relationships between words in the sentence.

[0031] Step S015: Define a set of relation extraction rule templates, match specific syntactic structure paths on the dependency relation tree to identify semantic relations between entities, and store the successfully matched entities and their relations in the form of structured subject-verb-object triples.

[0032] Semantic relationships include at least: the execution relationship between a character and an action, the domination relationship between an action and an item, and the interaction relationship between characters; for example, the domination relationship between a character and an item can be identified by matching the subject-verb-object syntactic path that connects the character entity and the item entity.

[0033] Step S016: Integrate the entities and logical relationships to generate and output a hierarchical digital scene description object, i.e., a DSD object.

[0034] DSD objects use a key-value pair data format. Their top-level structure is a list of scenes, organized in the original order of the script. Each scene object in the list contains a unique scene ID and a structured parsing result of the scene title entity.

[0035] The structured analysis results include interior and exterior scene identifiers, location, time, and a list of atmosphere keywords extracted from the scene description.

[0036] Each scene description object contains a content sequence, which consists of multiple content block objects arranged chronologically. The content block objects are defined as character action blocks or character dialogue blocks based on their properties. The character action block contains the character entity performing the action, the original text description of the action, and a list of relation triples related to the action. The character dialogue block contains the character entity speaking, the text content of the dialogue, and character performance or emotion cues extracted from parentheses.

[0037] In this embodiment, text script data is received, the specific content of which is: EXT. Riverside Cabin - Daytime It was a sunny day. Old Wang and the fisherman were by the river.

[0038] The fisherman sat fishing, while Old Wang scattered bait.

[0039] Fisherman

[0040] (leisurely)

[0041] The weather is really nice today.

[0042] The script is determined to be in plain text format. After standardizing the content of the plain text script, the following content is obtained: EXT. Riverside Cabin - Daytime It was a sunny day. Old Wang and the fisherman were by the river.

[0043] The fisherman sat fishing, while Old Wang scattered bait.

[0044] Fisherman

[0045] (leisurely)

[0046] The weather is really nice today.

[0047] The standardized content is semantically segmented to generate preprocessed text units in sentence form: [

[0048] {"type": "Scene Heading", "content": "EXT. Riverside Cabin - Daytime"}, {"type": "Action", "content": "A sunny day."}, {"type": "Action", "content": "Old Wang and the fisherman are by the river."} {"type": "Action", "content": "A fisherman sits and fishes,"} {"type": "Action", "content": "Old Wang is scattering bait."} {"type": "Character", "content": "Fishing Guy"}, {"type": "Parenthetical", "content": "(leisurely)"}, {"type": "Dialogue", "content": "The weather is really nice today."} ].

[0049] Specifically, if the text script data is available, the specific content is as follows: <content> <paragraph type="Scene Heading"> <text> EXT. Riverside Cabin - Daytime< / text> < / paragraph> <paragraph type="Action"> <text> It was a sunny day. Old Wang and the fisherman were by the river.< / text> < / paragraph> <paragraph type="Action"> <text> The fisherman sat fishing, while Old Wang scattered bait.< / text> < / paragraph> <paragraph type="Character"> <text> Fisherman< / text> < / paragraph> <paragraph type="Parenthetical"> <text> (leisurely)< / text> < / paragraph> <paragraph type="Dialogue"> <text> The weather is really nice today.< / text> < / paragraph> < / content> It is a format based on a markup language. When calling the corresponding XML parser, the system will directly parse these XML tags and their contents and map them directly to construct the final DSD object. Therefore, this processing path does not need to go through subsequent intermediate steps such as standardization and purification, semantic block segmentation, NER entity recognition, and relationship extraction, thus achieving efficient processing of structured data.

[0050] Entity recognition and annotation are performed on the text unit, and the following results are obtained: Scene title: [EXT. River Cottage - Daytime] (Scene title) (Parsed out: [EXT.], [River], [Cottage], [Daytime]) Action description: A [sunny] (atmosphere keyword) day.

[0051] Action description: [Lao Wang] (character name) and [Fisherman] (character name) are by the river.

[0052] Action description: [Fisherman] (character name) is sitting fishing, Action description: [Lao Wang] (character name) is spreading [bait] (key prop).

[0053] Character name: [Fisherman] (character name)

[0054] Mood hint: (悠闲地) [Leisurely]

[0055] Dialogue content: The weather is really nice today.

[0056] The system analyzes "The fisherman is sitting fishing" and "Lao Wang is spreading bait": 1. Dependency syntactic analysis: Generate core syntactic structures such as "Fisherman - sitting fishing" and "Lao Wang - spreading - bait"; 2. Rule template matching: Successfully match two key relationships; 3. Store triples: { "subject": "Fisherman", "predicate": "sitting fishing", "object": null} { "subject": "Lao Wang", "predicate": "spreading", "object": "bait"}.

[0057] Finally, the system integrates all recognized entities and extracted logical relationships to generate a DSD object. Regardless of whether the input is in plain text or markup language format, the structure of the final output DSD object is as follows: { "scenes": { "scene_id": "scene_001", "title_details": { "raw_text": "EXT. River Cottage - Day", "int_ext": "EXT", "location": "River Cottage", "time_of_day": "Day", "atmosphere_keywords": ["Sunny", "Relaxed"] }, "content_sequence": { "type": "action_block", "content": "On a sunny day. Lao Wang and the angler are by the river. The angler is sitting fishing and Lao Wang is sprinkling bait.", "relation_triples": { "subject": "The angler", "predicate": "sitting fishing", "object": null}, { "subject": "Lao Wang", "predicate": "sprinkling", "object": "bait"} }, { "type": "dialogue_block", "character": "The angler", "parenthetical": "(Relaxed)", "dialogue": "What a nice day today." } } } Through the above steps, the present invention has successfully converted the original script data in different formats into unified structured instructions that can be accurately understood and executed by the downstream 3D generation module, fully demonstrating the technical solution of the present invention in the script parsing stage.

[0058] ​​​Step S02: Based on the content in the DSD object, automatically construct the basic environment of the 3D scene in the 3D engine environment through semantic-visual mapping and procedural generation rules.

[0059] like Figure 2 As shown, the specific steps are as follows: Step S021: The system iterates through the scene list in the DSD object and automatically creates an independent 3D level named after its scene title in the 3D engine environment for each scene object in the list.

[0060] Step S022: Call a pre-built 3D asset library. Each 3D model in the asset library is labeled with multi-dimensional metadata tags and associated with a high-dimensional feature vector generated by a pre-trained multimodal model, representing its visual and semantic features.

[0061] Step S023: Start the semantic-visual mapping engine, extract scene location and atmosphere keywords from the current scene object, and use a multimodal model to convert the keyword combination into a query semantic vector.

[0062] Step S024: By calculating the cosine similarity between the query semantic vector and the feature vectors of all models in the asset library, the assets are ranked according to their matching degree, and the main environmental model and a series of related prop models with a matching degree higher than the preset threshold are retrieved.

[0063] Step S025: Automatically deploy the retrieved main environmental model to the newly created 3D level as the basic scene, and automatically arrange the retrieved prop models according to the preset procedural content generation rules and a rule library containing three types of rules: physical and spatial constraints, semantic association and functional layout, and context modifiers, to complete the initial construction of the scene.

[0064] Physics and spatial constraint rules are used to define the physical relationships between props, including ensuring that the bottom of the prop model automatically fits the top of the supporting surface model, preventing mutual penetration between models through bounding box detection, and preserving unobstructed passages in the scene based on pathfinding mesh data that conform to the character's movement logic.

[0065] Semantic association and functional layout rules are used to cluster props according to their functional attributes. These rules define the parent-child and companion relationships between props of different categories. For example, a model of the category "chair" is placed next to a model of the category "table" at a preset interaction point, and a model of the category "table lamp" is placed on the surface of a model of the category "bedside table" to form a functionally reasonable visual combination.

[0066] Context modifier rules are used to dynamically adjust the layout details and state of props based on the atmosphere keywords extracted from the digital scene description object. For example, when the keyword "messy" is detected, the rule will trigger a slight random rotation and displacement of props of the specified category and increase the generation density of small clutter models. When the keyword "tidy" is detected, a layout method aligned with the scene's predetermined grid will be triggered to reflect different scene states.

[0067] Based on the time stamps parsed from the scene objects, basic lighting components are generated and configured. The lighting components include at least directional light sources, sky atmosphere, and sky lighting, and their initial rotation angles and intensities are set. At the same time, based on the atmosphere keywords parsed from the scene objects, the color temperature, intensity, and post-processing effect parameters of the lighting components are automatically adjusted. For example, for the keyword "tension," the system will apply a low-illuminance, high-contrast cool-toned lighting scheme and add volumetric fog effects to create a visual atmosphere that matches the script description.

[0068] In this embodiment, the system traverses the scenes list in the DSD object and finds the first scene object (scene_001). Based on its raw_text field "EXT.River Cabin-Daytime", a new 3D level file is automatically created in the 3D engine's project content browser and named Previs_Level_EXT_River_Cabin_Day.

[0069] The system engine extracts key information from the DSD object's title_details: location: "riverside cabin" and atmosphere_keywords: ["sunny"]. Then, the keyword combination "sunny daytime riverside cabin" is input into a pre-trained multimodal model to generate a query vector representing this semantic meaning. The system then ranks the queries based on their cosine similarity to the feature vectors of all models in the 3D asset database. The main environmental model with the highest matching degree was retrieved, such as a riverside environment kit called SM_Riverside_Environment_Kit, which includes models of rivers, grasslands, trees, etc. The SM_Wooden_Cabin model was also retrieved.

[0070] Related prop models with high matching degree were also retrieved, such as: SM_Fishing_Chair (fishing chair), SM_Fishing_Rod (fishing rod), etc.

[0071] The system deploys the retrieved SM_Riverside_Environment_Kit model into the level, generating the basic river and riverside terrain. Then, the SM_Wooden_Cabin model is placed on the riverbank. Afterwards, props are placed using the PCG rule base. Physical and spatial constraints: Based on DSD analysis, the system identifies two core characters in the scene (Old Wang and the fisherman). Therefore, it procedurally places two SM_Fishing_Chairs on the flat ground along the riverbank. The system automatically performs ground-hugging operations to ensure they do not float or embed themselves in the ground. Bounding box detection ensures that the two chairs maintain reasonable social and physical distances from each other, as well as from other environmental models such as huts and trees, preventing them from intersecting.

[0072] Semantic Association and Functional Layout Rules: The system further analyzes the specific behaviors of the characters based on the DSD object: the core activity of "Fishing Guy" is "sitting and fishing," while the core activity of "Old Wang" is "sprinkling bait" (a standing action). Therefore, the PCG engine will arrange a functional fishing area around these two characters and activities. The system logically marks one SM_Fishing_Chair as "occupied" and associates it with the "Fishing Guy" character. The other SM_Fishing_Chair is marked as "idle." Although it is prepared for "Old Wang," since "Old Wang's" current behavior is standing, it is not marked as occupied. At the same time, the system loads the SM_Fishing_Rod model that directly interacts with "Fishing Guy" into the scene, but sets its initial state to "hidden" and disables its physical collision through procedural rules. This prop is now logically marked as being held by "Fishing Guy," but is completely invisible visually, awaiting final attachment and activation with the "Fishing Guy" character model in subsequent step S03.

[0073] (3) Context modifier rule: Based on the keywords "sunny" and "leisurely" extracted from the DSD object, the system triggers the "natural" modifier rule. This rule applies small, programmatic random scaling and rotation to natural elements such as trees and grass in the scene to avoid repetition and enhance the natural and relaxed atmosphere of the scene.

[0074] The system configures the lighting and atmosphere. Based on the `time_of_day: "daytime"` and `atmosphere_keywords: ["sunny", "leisurely"]` in the DSD, the system creates a strong, warm-colored directional light source to simulate sunlight on a sunny day, setting its angle to a height consistent with normal daylighting logic. Simultaneously, the system creates a sky atmosphere and skylight components, associating them with the directional light source to automatically generate a bright blue sky background and ample global ambient light. Then, based on the atmosphere keywords "sunny" and "leisurely," the system automatically adjusts the scene's post-processing effects: increasing exposure compensation for a brighter, clearer overall image; reducing contrast for smoother light and shadow transitions, creating a relaxed visual experience; and slightly increasing saturation to make the colors of the natural environment more vivid. Finally, the system adds and enables a volumetric cloud component, generating sparse, white clouds to further enhance the "sunny daytime" visual atmosphere.

[0075] After all the above steps, the system automatically generates a sunny, inviting fishing scene at a riverside cottage. This environment includes terrain, buildings, fixed props, and a complete lighting system consistent with the "sunny daytime" description. Items requiring character interaction are loaded and ready, preparing for subsequent character deployment and interaction binding. The final effect is as follows... Figure 4 As shown.

[0076] Through this step, the system successfully transformed the DSD data into a 3D visualization environment that users can directly review, laying a solid foundation for subsequent role deployment.

[0077] Step S03: In the basic 3D scene environment, based on key information and preset layout rules, automatically deploy characters, props and interactive elements to generate a 3D pre-visualization scene.

[0078] The specific steps are as follows: Step S031: Based on the content block object of the current scene in the DSD object, retrieve and load the corresponding 3D character model from the 3D asset library into the 3D level.

[0079] Step S032: Start the cinematic layout engine and analyze the core interaction types of the current scene based on the content sequence in the DSD object. The interaction types include at least: dialogue interaction, action conflict or single-player scene.

[0080] Step S033: Based on the analyzed interaction type, the cinematic layout engine selects and applies the corresponding layout strategy from a preset camera language rule library to calculate and determine the initial position and orientation of each character model in the 3D level.

[0081] The camera language rule base includes at least: a 180-degree axis rule for dialogue interaction, which automatically establishes a virtual dialogue axis connecting the dialogue characters and places the character models on both sides of the axis to form a confrontation or communication layout; and a rule of thirds composition for single-person or environmental interaction scenes, which places the characters on the key points of interest or the golden ratio line of the scene.

[0082] Step S034: Parse the action description text within the character action block in the DSD object, and retrieve the most matching animation clip from a preset animation library containing standard animation clips by keyword matching or semantic similarity calculation.

[0083] Step S035: Assign the retrieved animation clip to the corresponding character model and set it as the character's initial idle animation, or set it as an interactive animation that can be triggered by a specific event.

[0084] Step S036: Based on the relation triplet information stored in the DSD object, establish automated interaction logic between the character model and the prop model, thereby realizing automated interaction pre-playing between the character and the prop, i.e., 3D pre-playing scene. For example, for a character A picking up prop B, the system will trigger an event at a predetermined keyframe on the interaction animation timeline of character A, attaching the prop B model from its original position to the hand bone node of the character A model.

[0085] In this embodiment, the system analyzes the content sequence of the DSD object and identifies the list of characters appearing in the scene as: ["Old Wang", "Fishing Guy"]. Subsequently, the system retrieves and loads the corresponding 3D character models from the 3D asset library into the Previs_Level_EXT_River_Cabin_Day level, loading SKM_OldMan_A (as "Old Wang") and SKM_OldMan_B (as "Fishing Guy").

[0086] The system activates a cinematic layout engine to locate the characters. Engine analysis of the DSD (Discretionary Scripting) reveals that the scene contains both core actions of multiple characters (such as "sitting by the river" and "baiting") and dialogue. Based on this composite information, the system classifies the core interaction type of the current scene as "accompanying dialogue within a parallel activity." Based on this judgment, the engine does not adopt a traditional face-to-face dialogue layout but instead calls a more context-appropriate layout strategy. The system reads that one of the fishing chairs has been logically marked as occupied by "the fisherman," so it precisely places the character model of SKM_OldMan_B (the fisherman) on this chair. Simultaneously, the system identifies "Old Wang's" core action as "scattering bait," a standing action. Therefore, the engine places the character model of SKM_OldMan_A (Old Wang) next to another empty chair, in a standing position near the riverbank. This layout provides reasonable space for "Old Wang's" standing action and also subtly suggests his identity as a participant in the scene through his proximity to the empty chair.

[0087] After completing the character layout, the system enters the animation matching and application phase. For the seated SKM_OldMan_B, the system matches and sets A_Fishing_Idle as its initial idle animation based on its description of "sitting and fishing". For the standing SKM_OldMan_A, the system matches a throwing interaction animation based on its core action of "scattering bait" and sets it to be triggerable by an event.

[0088] Finally, the system executes the automated interaction logic construction. Based on the logic markers in step S02, the system activates the previously "hidden" SM_Fishing_Rod model instance and attaches it to the right-hand bone node of the SKM_OldMan_B character model, perfectly integrating it with the character's idle animation pose. Simultaneously, the system establishes "splashing" interaction logic for SKM_OldMan_A; when the pre-play reaches a specific moment, its model will automatically play the corresponding interactive animation, completing the key actions described in the script.

[0089] After all the above steps, the system automatically generates a complete 3D pre-visualization scene that includes the scene environment, dynamic and static characters, props, and core interactions. The final effect is as follows: Figure 5 As shown.

[0090] Step S04: The system provides interactive review and context-bound feedback functions, and generates a revised script by calling the large language model through API based on the feedback information.

[0091] Specifically: In the 3D pre-visualization scene, users are provided with a freely controllable first-person or third-person perspective camera, allowing them to review the automatically constructed scene, character layout, and lighting atmosphere from different angles through roaming, panning, rotating, and zooming.

[0092] The system provides a user-activated feedback mode that switches from scene roaming control to feedback pointer control, allowing users to precisely locate the area or object they intend to point to on the 3D rendered screen. When a user clicks on the scene in feedback mode, the system performs a ray projection calculation from the screen's 2D coordinates to 3D world coordinates to determine the specific 3D spatial location of the user's click and identify whether there is an interactive 3D model object at that location.

[0093] The system displays a user input interface, allowing users to input text-based annotations or modification opinions for the current location point or identified object.

[0094] After a user submits an annotation, the system automatically captures and generates a structured context-bound feedback package. This package contains at least the following data: the user-inputted text annotation, the 3D world coordinates calculated from ray projection, a unique identifier (ID) for the object if it exists, and the camera position parameters for the current user review camera, including position, rotation, and field of view. Based on the 3D spatial location or object ID recorded in the feedback package, the system traces back and locates the original script paragraph or entity in the DSD object that generated the region or object.

[0095] The system constructs an input prompt, Prompt, for large language models. The prompt integrates the following information: the located original script paragraph as the basic context, the user's input text annotations as modification instructions, and the text description of the 3D scene context as supplementary context.

[0096] The prompt will be sent to a common large language model via the API interface, requesting the model to generate a revised script text that conforms to the script format based on all the information provided, and then receiving the text as the revised script.

[0097] In this embodiment, the user can control a third-person perspective camera and freely roam the level. For example, during the review process, the user felt that "Old Wang's" action of directly "sprinkling bait" seemed abrupt and lacked a preparatory action, and hoped to add an action of "picking up the fish food box" before sprinkling bait.

[0098] At this point, the user can activate the system's feedback mode, move the pointer to the character model "Old Wang" (SKM_OldMan_A), and click. The system immediately performs ray casting calculations and successfully identifies the clicked object as SKM_OldMan_A. Subsequently, a user input interface pops up, where the user can enter their specific suggestions for improvement: "The movements are not fluid enough; Old Wang should grab the fish food box first, and then scatter the bait."

[0099] After submitting the comment, the system automatically captures and generates a structured context-bound feedback package, the content of which is shown below: { "user_text": "The actions weren't fluid enough; Lao Wang should have taken the fish food box first, and then sprinkled the bait." "clicked_object_id": "SKM_OldMan_A", "world_coordinates": { "x": 150.0, "y": 320.5, "z": 90.0}, "camera_params": { "location": { "x": 50.0, "y": 450.0, "z": 120.0}, "rotation": { "pitch": -15.0, "yaw": -90.0, "roll": 0.0}, "fov": 90.0 } } Based on the clicked_object_id "SKM_OldMan_A" in the feedback packet, the system traced back to the role corresponding to the model, which was "Old Wang". Next, in the content sequence of the DSD object, the system located the action_block containing "Old Wang's" core actions, whose original script content was: "A sunny day. Old Wang and the fisherman were by the river. The fisherman sat fishing, and Old Wang was scattering bait." The system integrates all captured information into a high-quality, structured input prompt, ready to be sent to the large language model. Task instructions: You are a professional screenwriter. Please optimize a specified original script segment based on the user's editing instructions and context. Please maintain a consistent script format and style, and only output the revised text.

[0100] Original script excerpt: It was a sunny day. Old Wang and the fisherman were by the river. The fisherman was sitting and fishing, while Old Wang was scattering bait.

[0101] User modification instructions: The movements weren't fluid enough; Lao Wang should have taken the fish food box first, and then sprinkled the bait.

[0102] 3D scene context: In the 3D pre-visualization scene, the user clicked on the character model named "Old Wang" and provided the aforementioned modification command. The scene is an outdoor riverside scene on a sunny day, featuring two characters: "Old Wang" and "Fisherman".

[0103] The system sends the structured prompts mentioned above to the large language model service via the API interface. After understanding all the context information, the model generates and returns the following revised version: It was a sunny day. Old Wang and the fisherman were by the river. The fisherman was sitting and fishing. Old Wang picked up a fish food box from the ground and then sprinkled bait into the water.

[0104] Step S05: Based on the updated key information in the revised script and the 3D pre-visualization scene, achieve closed-loop iterative optimization from text to visualized scene, such as... Figure 3 As shown; Step S051: The revised script returned by the large language model is presented to the user in the user interface. The interface displays the original script paragraphs before revision and the suggested text after revision side by side, so that the user can compare, confirm or manually edit them.

[0105] Step S052: After the user confirms and accepts the revised script, the system performs a difference analysis on the text before and after the revision, accurately compares the text changes between the original script paragraphs and the final confirmed revised draft, and identifies all modified, added or deleted text content.

[0106] Step S053: The system inputs the difference content into the script semantic parsing engine of step S01, and performs semantic parsing on the changed parts to generate new entities and relationships, or update or delete existing entities and relationships.

[0107] Step S054: Based on the results of the re-parsing, update the corresponding data structure in the DSD object to ensure that the DSD object is fully synchronized with the latest script content.

[0108] Step S055: The system initiates the dependency tracking mechanism, analyzes the changes between the updated DSD object and the original DSD object, and determines all 3D scene elements affected by this update based on the preset dependency graph. These elements include, but are not limited to, scene models, prop layouts, lighting parameters, or character animations.

[0109] Step S056: For all 3D scene elements identified as affected, the system triggers a partial or complete scene reconstruction process; if the change only involves prop positions or lighting parameters, a partial update is performed; if the change involves core elements such as scene locations, a complete reconstruction of the entire 3D level is performed.

[0110] Step S057: The reconstruction process will call the corresponding steps in steps S02 and S03, and automatically regenerate or adjust the affected 3D scene parts using the data in the updated DSD object.

[0111] Step S058: After the reconstruction is completed, the updated 3D pre-visualization scene is presented to the user in real time, thus completing a complete closed-loop iteration from 3D scene feedback to text script modification, and then to automatic 3D scene update. The user can continue to review the new scene and repeat this process until a satisfactory pre-visualization effect is obtained.

[0112] Based on the same inventive concept, this invention also proposes an apparatus for automated 3D scene pre-visualization based on a text script. The implementation of this apparatus can be found in the implementation of the method described above; repeated details will not be repeated. Figure 6 As shown, the device 100 includes: Text script processing module 101: Used for the system to receive text script data, start the script semantic parsing engine for structured processing, and extract DSD objects; 3D Scene Basic Environment Module 102: Used to automatically construct the 3D scene basic environment in the 3D engine environment based on the content in the DSD object through semantic-visual mapping and procedural generation rules; 3D Preview Scene Module 103: Used to automatically deploy characters, props and interactive elements in a 3D scene base environment based on key information and preset layout rules to generate a 3D preview scene; Interactive review module 104: This module provides interactive review and context-bound feedback functions, and generates a revised script by calling the large language model via API based on the feedback information. Scene Iteration Optimization Module 105: Used to achieve closed-loop iterative optimization from text to visualized scene based on the updated key information of the revised script and the 3D pre-visualization scene.

[0113] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0114] like Figure 7As shown, the device includes a central processing unit (CPU), which can perform various appropriate actions and processes based on computer program instructions stored in read-only memory (ROM) or loaded from storage units into random access memory (RAM). The RAM can also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0115] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0116] The processing unit executes the various methods and processes described above, such as method steps S01 to S05. For example, in some embodiments, method steps S01 to S05 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of method steps S01 to S05 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute method steps S01 to S05 by any other suitable means (e.g., by means of firmware).

[0117] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0118] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0119] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0120] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0121] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for automated 3D scene pre-visualization based on text scripts, characterized in that, The method includes: Step S01: The system receives text script data, starts the script semantic parsing engine for structured processing, and extracts DSD objects; Step S02: Based on the content in the DSD object, automatically construct the basic environment of the 3D scene in the 3D engine environment through semantic-visual mapping and procedural generation rules; Step S03: In the basic 3D scene environment, based on key information and preset layout rules, automatically deploy characters, props and interactive elements to generate a 3D pre-visualization scene.

2. The method for automated 3D scene pre-visualization based on text scripts according to claim 1, characterized in that, The specific steps of step S01 are as follows: Step S011: The system receives the text script file uploaded by the user and performs text preprocessing on the script content: it determines the format of the text script file. If it is a markup language-based format, it calls the corresponding XML parser to extract the preset tag content, maps and constructs a DSD object. Step S012: If it is in plain text format, then standardize the script content in plain text format; A pattern matching method based on regular expressions is used to perform semantic block segmentation on standardized plain text content. According to script writing conventions, scene title blocks, action description blocks, character name blocks and dialogue blocks are identified and separated. Sentence segmentation is performed on the segmented action description blocks to generate preprocessed text units in sentences. Step S013: Initiate the deep learning-based entity recognition process, perform entity recognition and annotation on the preprocessed text units, input the text units one by one into the NER model to traverse each word in the text, and output a predefined entity label; the system merges the predefined entity label sequence output by the model and extracts the complete entity from the text; the predefined entity label includes at least: scene title, interior and exterior scene identifiers, scene location, time identifier, character name, action description, key props, and atmosphere keywords; Step S014: Start the relation extraction process based on syntactic analysis, analyze the logical relations between the identified entities, and for sentences containing at least two identified entities, call the dependency parsing engine to generate the dependency relation tree of the sentence. The dependency relation tree reveals the grammatical subordination relationship between words in the sentence. Step S015: Define a set of relation extraction rule templates, match specific syntactic structure paths on the dependency relation tree to identify semantic relations between entities, and store the successfully matched entities and their relations in the form of structured subject-verb-object triples; the semantic relations include at least: the execution relationship between roles and actions, the domination relationship between actions and props, and the interaction relationship between roles. Step S016: Integrate the entities and logical relationships to generate and output a hierarchical DSD object.

3. The method for automated 3D scene pre-visualization based on text scripts according to claim 2, characterized in that, The DSD object uses a key-value pair data format. Its top-level structure is a list of scenes, organized in the original order of the script. Each scene object in the list contains a unique scene ID and a structured parsing result of the scene title entity. The structured analysis results include interior and exterior scene identifiers, location, time, and a list of atmosphere keywords extracted from the scene description; Each scene description contains a content sequence consisting of multiple content block objects arranged chronologically. The content block objects are defined as character action blocks or character dialogue blocks based on their properties. The character action block contains the character entity performing the action, the original text description of the action, and a list of relation triples related to the action. The character dialogue block contains the character entity speaking, the text content of the dialogue, and character performance or emotion cues extracted from parentheses.

4. The method for automated 3D scene pre-visualization based on text scripts according to claim 1, characterized in that, The specific steps of step S02 are as follows: Step S021: The system iterates through the scene list in the DSD object and automatically creates an independent 3D level named after its scene title in the 3D engine environment for each scene object in the list. Step S022: Call a pre-built 3D asset library. Each 3D model in the asset library is labeled with multi-dimensional metadata tags and associated with a high-dimensional feature vector generated by a pre-trained multimodal model, representing its visual and semantic features. Step S023: Start the semantic-visual mapping engine, extract scene location and atmosphere keywords from the current scene object, and use a multimodal model to convert the keyword combination into a query semantic vector; Step S024: By calculating the cosine similarity between the query semantic vector and the feature vectors of all models in the asset library, the assets are ranked according to their matching degree, and the main environmental model and a series of related prop models with a matching degree higher than the preset threshold are retrieved. Step S025: Automatically deploy the retrieved main environmental model to the newly created 3D level as the basic scene, and automatically arrange the retrieved prop models according to the preset procedural content generation rules and a rule library containing three types of rules: physical and spatial constraints, semantic association and functional layout, and context modifiers, to complete the initial construction of the scene.

5. The method for automated 3D scene pre-visualization based on text scripts according to claim 4, characterized in that, The physical and spatial constraint rules described in step S025 are used to define the physical relationships between props, including ensuring that the bottom of the prop model automatically fits the top of the supporting surface model, avoiding mutual penetration between models through bounding box detection, and preserving unobstructed passages in the scene based on pathfinding grid data that conform to the character's movement logic. The semantic association and functional layout rules described in step S025 are used to perform clustered layout based on the functional attributes of props. These rules define the parent-child and companion relationships between different categories of props. The context modifier rules described in step S025 are used to dynamically adjust the layout details and status of props based on the atmosphere keywords extracted from the digital scene description object.

6. The method for automated 3D scene pre-visualization based on text scripts according to claim 1, characterized in that, The specific steps of step S03 are as follows: Step S031: Based on the content block object of the current scene in the DSD object, retrieve and load the corresponding 3D character model from the 3D asset library into the 3D level; Step S032: Activate the cinematic layout engine and analyze the core interaction types of the current scene based on the content sequence in the DSD object, including at least: dialogue interaction, action conflict, or single-player scene; Step S033: Based on the analyzed interaction type, the cinematic layout engine selects and applies a corresponding layout strategy from a preset camera language rule library to calculate and determine the initial position and orientation of each character model in the 3D level; the camera language rule library includes at least: a 180-degree axis rule for dialogue interaction, which automatically establishes a virtual dialogue axis connecting the dialogue characters and places the character models on both sides of the axis to form a confrontation or communication layout; and a rule of thirds composition for single-person or environmental interaction scenes, which places the character on the key point of interest or the golden ratio line of the scene; Step S034: Parse the action description text within the character action block in the DSD object, and retrieve the most matching animation clip from a preset animation library containing standard animation clips by keyword matching or semantic similarity calculation. Step S035: Assign the retrieved animation clip to the corresponding character model and set it as the character's initial idle animation, or set it as an interactive animation that can be triggered by a specific event; Step S036: Based on the relation triplet information stored in the DSD object, establish automated interaction logic between the character model and the prop model, thereby realizing automated interaction pre-play between the character and the prop, i.e., 3D pre-play scene.

7. The method for automated 3D scene pre-visualization based on text scripts according to claim 1, characterized in that, It also includes step S04: The system provides interactive review and context-bound feedback functions, and generates a revised script by calling the large language model through API based on the feedback information; Specifically, in the 3D pre-visualization scene, a first-person or third-person perspective camera that can be freely controlled is provided to the user, allowing the user to review the automatically constructed scene, character layout and lighting atmosphere from different angles in different ways; The system provides a user-activated feedback mode that switches from scene roaming control to feedback pointer control, allowing users to precisely locate the area or object they intend to point to on the 3D rendered screen. When a user clicks on the scene in feedback mode, the system performs a ray projection calculation from the screen's 2D coordinates to the 3D world coordinates to determine the specific 3D spatial location of the user's click and to identify whether there is an interactive 3D model object at that location. The system provides a user input interface, allowing users to input text-based annotations or modification suggestions for the current location point or identified objects. After the user submits the annotation, the system automatically captures and generates a structured context-bound feedback package. The feedback package contains at least the following data: the text annotation content entered by the user, the 3D world coordinates calculated by ray projection, the unique identifier ID of the object if the user inputs it, and the camera position parameters of the current user review camera, including position, rotation, and field of view. Based on the 3D spatial position or object ID recorded in the feedback package, the system traces back and locates the original script paragraph or entity in the DSD object that generated the region or object. The system constructs an input prompt for a large language model, which integrates the following information: the located original script paragraph as the basic context, the user's input text annotations as modification instructions, and the text description of the 3D scene context as supplementary context. The prompt will be sent to a common large language model via the API interface, and the model will be requested to generate a revised script text that conforms to the script format based on all the information provided, and the text will be received as the revised script. Step S05: Based on the updated key information in the revised script and the 3D pre-visualization scene, achieve closed-loop iterative optimization from text to visual scene; The specific steps are as follows: Step S051: The revised script returned by the large language model is presented to the user in the user interface. The interface displays the original script paragraphs before revision and the suggested text after revision side by side for the user to compare, confirm or manually edit. Step S052: After the user confirms and accepts the revised script, the system performs a difference analysis on the text before and after the revision, accurately compares the text changes between the original script paragraphs and the final confirmed revised draft, and identifies all modified, added or deleted text content. Step S053: The system inputs the difference content into the script semantic parsing engine of step S01, and performs semantic parsing on the changed parts to generate new entities and relationships, or update or delete existing entities and relationships. Step S054: Based on the results of the re-parsing, update the corresponding data structure in the DSD object to ensure that the DSD object is fully synchronized with the latest script content; Step S055: The system starts the dependency tracking mechanism, analyzes the changes between the updated DSD object and the original DSD object, and determines all 3D scene elements affected by this update based on the preset dependency graph. These elements include, but are not limited to, scene models, prop layouts, lighting parameters, or character animations. Step S056: For all identified affected 3D scene elements, the system triggers a partial or complete scene reconstruction process; if the change only involves prop positions or lighting parameters, a partial update is performed; if the change involves core elements such as scene locations, a complete reconstruction of the entire 3D level is performed. Step S057: The reconstruction process will call the corresponding steps in steps S02 and S03, and automatically regenerate or adjust the affected 3D scene parts using the data in the updated DSD object. Step S058: After the reconstruction is completed, the updated 3D pre-visualization scene is presented to the user in real time, thus completing a complete closed-loop iteration from 3D scene feedback to text script modification, and then to automatic 3D scene update. The user continues to review the new scene and repeat this process until a satisfactory pre-visualization effect is obtained.

8. A device for automated 3D scene pre-visualization based on a text script, characterized in that, The device implements the method as described in any one of claims 1 to 7, comprising: Text script processing module: Used for the system to receive text script data, start the script semantic parsing engine for structured processing, and extract DSD objects; 3D Scene Basic Environment Module: Used to automatically construct the basic 3D scene environment in the 3D engine environment based on the content in the DSD object through semantic-visual mapping and procedural generation rules; 3D Preview Scene Module: Used to automatically deploy characters, props, and interactive elements in a 3D scene environment based on key information and preset layout rules, generating a 3D preview scene.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images