Task-Driven Video Production Group Dialogue Management Method
By obtaining task context and element identifier sets from the video production platform, identifying shared elements and forming element transfer relationships, the problem of inaccurate task association information in video production is solved, thereby improving the accuracy and efficiency of the video production process.
Patent Information
- Application Number
- CN202611618624.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing video production platforms cannot locate the actual source and transmission direction of shared elements in task association information, resulting in inconsistencies between the task context corresponding to the video production instruction and the actual processing chain. This may lead to omissions of source tasks that provide target elements or inclusion of tasks that have not actually received the element.
By acquiring the task context and feature identifier set, shared features are identified and categorized into value fragments, reference fragments, or time-series fragments. Fragments that are copied and executed after the task are executed are used as fragment copies. Fragment copies are detected and recovery requests are sent to restore the records of the source task, thus forming feature transfer relationships and ensuring the accurate transfer of instruction target features in the task group.
It reduces task omissions and irrelevant context calls, improves the accuracy and efficiency of the video production process, and avoids the reprocessing of irrelevant video content.
Smart Images

Figure CN123123676A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of video processing technology, and in particular to a task-driven video production group dialogue management method. Background Technology
[0002] The video production platform receives user input regarding scriptwriting, character design, storyboard design, shot generation, voice-over, subtitles, music, and editing requirements, and then generates or modifies the video content using corresponding tasks. A character's costume, appearance resources, and screen time are typically recorded across character design, storyboard, and shot generation tasks. When a user continuously modifies a video through the same entry point, the output of the previous production task can be read by the next, and the results of the next task can be passed on to editing or special effects tasks.
[0003] Video production dialogue management requires linking user input, the current video project, the selected video clip, and related task context within the same processing session. If the task context is too broad, unrelated scenes, characters with the same name, and production records from other time periods will be included in the current processing round; if the task context is too narrow, tasks providing the target elements or those actually receiving the changed results will be missed. Furthermore, the actual content passed between tasks can change with material replacement, task reruns, and cache updates. Therefore, task type and execution order can only provide a preliminary association and cannot fully represent the actual transmission path of shared elements.
[0004] In existing technologies, video production dialogues can employ retrieval-Augmented Generation (RAG) technology. This technology retrieves context based on the text similarity between user input and project records, and can also invoke upstream or downstream tasks according to pre-configured task connections. While some platforms retain result retrieval records that identify the receiving task and execution batch, they typically do not record specific shared elements, segment types, or source versions. Based on this information, only potentially related tasks can be identified, making it difficult to confirm the actual source and transmission direction of a particular shared element in this execution.
[0005] Therefore, when multiple video production tasks process the same video element consecutively, text matching, static linking, or reading records containing only receiving tasks and execution batches are insufficient to accurately define the task context required for this round of dialogue. This may result in omission of the source task that provides the target element, or inclusion of tasks that did not actually receive the element. Summary of the Invention
[0006] This application provides a task-driven video production group dialogue management method to solve the problem that task association information cannot locate the actual source and transmission direction of shared elements, resulting in inconsistency between the task context corresponding to the video production instruction and the actual processing chain.
[0007] To address the aforementioned technical problems, this application provides a task-driven video production group dialogue management method, comprising: Obtain the task context and feature identifier set. Each feature identifier consists of an object attribute main identifier and a location range field. Identify features with the same object attribute main identifier and overlapping location ranges in at least two tasks as shared features. Record the context fragments of shared features and classify them into value fragments, reference fragments, or time series fragments according to the record format. Based on the results, read the records, identify the task that reads the execution result of the other task among the two tasks that record the same shared element as the subsequent task, and copy the fragment that records the shared element in the subsequent task as a fragment copy; detect the fragment copy according to the fragment type to obtain the picture layer, the actual read address or the connecting frame; If any probe result does not reflect the corresponding content, the location range of the probed fragment is used as the request range. A recovery request is sent to the task with the same object attribute main identifier and fragment type, and whose local record range intersects with the request range. The task that returns the record and recovers the fragment copy is used as the candidate source. The source probe content is placed into the fragment copy of the candidate source, so that the subsequent execution task is executed again according to the original reading route. The source task is confirmed based on the context access record, input version, and local results, forming the feature transfer relationship. Extract the target element identifier of the instruction; read-type instructions use the output task of the selected video segment as the seed and add the source task in reverse along the element transmission relationship; change-type instructions use the task of saving the target element as the seed; when the target element saving task has an outward relationship, it is added to the receiving task in the forward direction; when the target element is only used locally by the saving task, the task is retained and the change is executed directly until no new tasks are added and a task group is formed. Execute instructions within the task group context, update video content according to the segment type of the changed elements, and output the video production result.
[0008] Optionally, fragment copies are detected by fragment type and feature transitive relationships are confirmed, including: Save the pre-detection records of each segment to be detected in the task, as well as the original state summary of other context segments that should not be changed in this detection except for the segment to be detected and its associated output. Copy each segment to be detected to obtain a context copy. Each context copy is independently subjected to the corresponding probe; if the result of any probe does not reflect the corresponding probe content, the task is blocked and then executed to read the corresponding pre-probe record and send the corresponding recovery request. The task that returns a local record and restores the context copy is identified as a candidate source for the corresponding fragment type, forming a candidate relationship to be verified; candidate relationships for different fragment types are retained independently, while candidate relationships corresponding to the same candidate source are merged. For each candidate relation to be verified, the corresponding fragment of the candidate source is copied to obtain a source copy. The local record version of the candidate relation is written into the base version field of the source copy. After the source probe content is inserted, a source copy version is generated with the base version and the current verification batch. The original read route of the task is executed again after saving. When the base version is consistent with the local record version, the actual input version is the source copy version, the context access record points to the source copy, the local result reflects the source probe content, and the original state summary has not changed, the candidate relation is confirmed and an element transit relation is formed. Otherwise, the candidate relation is deleted and the context is restored.
[0009] Optionally, sending a recovery request to the candidate tasks and establishing a candidate relationship includes: The location range of the detected segment is used as the request range; the task registration records are queried according to the project identifier, the object attribute master identifier of the shared element, and the segment type; the tasks whose local record range intersects with the request range are identified as candidate tasks, and a recovery request is sent. Each candidate task returns the current local record that matches the project ID, object attribute primary ID, and fragment type, as well as the source range, local record version, and task ID of the current local record; Each local record is written to the corresponding fragment copy. Tasks that can be written according to fragment type are identified as candidate sources, and the source range, the receiving range represented by the request range, and the local record version are written into the candidate relationship. If no local record that can be written is available, restore the corresponding fragment copy without forming a candidate relationship.
[0010] Optionally, the task to be executed later reads the original route to verify candidate relationships, including: Copy the corresponding fragment of the shared element in the candidate source to obtain the source copy, write the local record version in the candidate relationship into the base version field of the source copy; put the source probe content that is different from the previous probe content into the source copy, and generate the source copy version with the base version and the current verification batch; Keep the read configuration, task scheduling, and invocation order of the subsequent task unchanged, execute the subsequent task and record the context access address and the actual input version; the context access address represents the logical location or record address of the subsequent task accessing the task context record, which is different from the actual read address of the referenced resource; First, compare the base version with the local record version saved in the candidate relationship. Then, compare the context access address, the actual input version, and the local result with the logical location of the source copy, the source copy version, and the source probe content, respectively. If the base version is consistent with the local record version, the context access address points to the source copy, the actual input version is the source copy version, and the local results reflect the source probe content, the candidate relationship is confirmed, the intersection of the source range and the receiving range is taken as the relationship range, and an element transfer relationship is formed; otherwise, the candidate relationship is deleted, and the candidate source and the context of the subsequent task are restored.
[0011] Optionally, at least one seed task is determined based on the instruction type, and task groups are generated in a targeted manner, including: The task to be executed after directly generating the corresponding output is determined from the output record of the selected video segment, and serves as the seed task for the read-type instruction; Identify at least one task that saves target elements from the task context; use the target element saving task that is the source task in the confirmed relationship as the change seed for outward traversal, and retain the target element saving task that has not formed an outward relationship in the task group to perform local changes; Read the object attribute master identifier and target range from the target feature identifier; filter the feature transfer relationships from the relationship records that have the same object attribute master identifier and whose relationship range intersects with the target range; For read-type instructions, the source task is added in reverse from the seed task; for change-type instructions, the receiving task is added in forward from the change seed, and the process continues with the newly added tasks until no new tasks are added.
[0012] Optionally, the task context and feature identifier set are obtained and shared feature fragments are segmented, including: Production records are aggregated from video production projects according to task identifiers to obtain task context and element identifier sets. Each element identifier consists of an object attribute main identifier and a location range. Elements with the same object attribute main identifier and overlapping location ranges in at least two tasks are identified as shared elements. The visual attribute values, resource read addresses, or context fragments of shot transition time positions of the shared elements are recorded as value fragments, reference fragments, or time sequence fragments, respectively. The completion order is read from the task execution records, and the task is executed after determining the result read records indicating the received task and execution batch.
[0013] Optionally, the target feature identifier of the extraction instruction includes: Receive user input, combine the user input with the current video project and the selected video segment to obtain instructions; parse the operation type, target object, target attribute, video location and target content corresponding to the change operation from the instructions, combine the target object and target attribute into the object attribute main identifier, and use the video location as the location range field to obtain the target element identifier.
[0014] Optionally, the target element identifier corresponding to the instruction is obtained, including: By limiting the time range according to the selected video clip and the location range field, filter the task contexts with the same target object within the time range, and obtain the feature identifier with the same object attribute main identifier from the filtered task contexts to obtain the target feature identifier.
[0015] Optionally, the instructions are executed in the context of a task group, including: The execution order of each task in the task group is determined by the element transfer relationship. The instructions and the context of each task are input into the corresponding task in sequence, so that the task to be executed later reads the execution result of the source task and obtains the task execution result. For read-type instructions, the shared elements read are recorded. For change-type instructions, the changed shared elements, fragment types and target content are recorded.
[0016] Optionally, the video content is updated according to the segment type of the changed elements, and the video production result is output, including: Read the changed shared elements, segment types, and target content from the task execution results. Convert the target content into picture parameters, resource addresses, or time parameters, and update the resource content or transition frames corresponding to the picture layer and resource address respectively. Combine the updated content with the unchanged video content and write it into the corresponding context segment to obtain the video production result.
[0017] The technical solution provided by this invention has the following beneficial effects: This invention defines shared elements by using the object attribute main identifier and location range. It confirms the element transmission relationship through fragment copy detection, recovery request and original read route verification, so that the relationship corresponds to the source record and effective video range of the actual access. Read commands obtain the source task in reverse along the relationship, and change commands obtain the receiving task in forward along the relationship. When the target element is only used locally, the saved task is directly changed, thereby reducing task omission, irrelevant context calls and reprocessing of irrelevant video content. 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 description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are 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 A flowchart illustrating a task-driven video production group dialogue management method provided in this application embodiment.
[0020] Figure 2 A timing diagram illustrating the candidate relationship formation and verification process provided in the embodiments of this application.
[0021] Figure 3 This is a flowchart illustrating the seed task determination and orientation of task groups provided in an embodiment of this application. Detailed Implementation
[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0023] It should be noted that this application applies to the processing of multiple video production tasks that a user continuously invokes through the same interactive entry point. Grouped dialogue means selecting the task and its corresponding task context from the video production project for the current round of processing in response to a single instruction; the task context consists of task inputs, task outputs, resource call records, time position records, and execution records grouped according to the same task identifier.
[0024] For example, existing diffusion model video generation workflows and node-based workflows can transfer character resources, shot conditions, or stage results between character setting, storyboard generation, shot generation, and editing. These task organization methods do not rely on a specific video generation model or fixed task names. In such workflows, the same task may directly use its own records or read the execution results of other tasks; the task name and execution order cannot determine the actual source of specific shared elements, therefore it is necessary to establish element transfer relationships based on the data actually read by the task and select the task for the current round accordingly.
[0025] Based on this, such as Figure 1 As shown, this embodiment provides a task-driven video production group dialogue management method, including the following steps: S1. Obtain the task context and feature identifier set; the feature identifier consists of an object attribute primary identifier and a location range field; identify features with the same object attribute primary identifier and intersecting location ranges in at least two tasks as shared features, and classify shared feature fragments into value fragments, reference fragments, or time-series fragments according to record format. S2. Determine the task to be executed after reading the result of another task based on the result read record; copy the corresponding fragment of the task to be executed and probe according to fragment type to obtain the image layer, actual read address, or connecting frame. S3. If any probe result does not reflect the corresponding content, use the location range of the probed fragment as the request range, and send a recovery request to the task with the same object attribute primary identifier and fragment type, and whose local record range intersects with the request range; take the task that returns the record and restores the fragment copy as a candidate source, insert the source probe content into the candidate source fragment copy, and execute according to the original read route; confirm the source task based on the context access record, input version, and local results to form a feature transfer relationship. S4: Extract the target element identifier of the instruction; for read-type instructions, add them to the source task in reverse from the selected segment output task; for change-type instructions, add them to the receiving task in forward from the target element save task; if there is no outward relationship, retain the target element save task, and traverse until no new tasks are added to form a task group. S5: Execute the instruction in the context of the task group, update the video content according to the segment type of the changed element, and output the video production result.
[0026] In S1, production records in the video production project are first aggregated according to task identifiers to obtain the context of each task. The task context consists of context fragments recording task input, task output, resource calls, time positions, and execution processes. The video production platform can save these records using existing JSON data formats. Subsequently, object identifiers, attribute identifiers, and video positions are extracted from the context fragments to obtain the element identifier set for each task. The object identifier and attribute identifier form the main object attribute identifier, and the position range field records the corresponding scene, shot, or frame interval. The main object attribute identifier and the position range field together form a single element identifier. The position range can be a single interval or composed of multiple scene or shot sub-intervals; if two position ranges intersect, it indicates that they share at least one common scene, shot, or frame interval. For example, the female lead and her clothing color form the main object attribute identifier for clothing color. Scenes three through five constitute the position range corresponding to this main object attribute identifier, and together they identify the female lead's clothing color element in scenes three through five.
[0027] After extracting feature identifiers, the feature identifier sets of each task are compared. Only when at least two tasks record the same primary object attribute identifier and their corresponding location ranges intersect, are the same primary object attribute identifiers identified as shared feature identifiers, and the corresponding features identified as shared features. Both conditions must be met simultaneously; if either condition is not met, the feature is not treated as the same shared feature. A task context can contain multiple feature fragments, each indexed by a feature identifier composed of the primary object attribute identifier and its own location range field. When feature fragments in different tasks meet the aforementioned conditions, the corresponding task is called the task that records shared features, and the corresponding fragment is called a shared feature fragment. Thus, a task context contains feature fragments, feature identifiers indicate the object attributes and locations corresponding to feature fragments, and shared feature identifiers are used to associate corresponding fragments in different tasks; no separate data object, the feature task, is established.
[0028] Next, shared element fragments are categorized according to the existing field formats in the production log. A context fragment is classified into only one type based on its actual recording format: value fragments record visual attribute values such as color, texture, and shot size that can be presented by a local image layer; reference fragments record the read addresses of image, audio, model, or video resources; and temporal fragments record the start and end points of shots or shot transitions that can be reflected in connecting frames. Therefore, the "or" in value fragment, reference fragment, or temporal fragment indicates a choice of classification for a single fragment; the same shared element can be recorded by multiple context fragments, thus corresponding to one, two, or three fragment types overall. Value fields can be directly read during the image generation process, address fields point to resources in the project's media library, and time fields correspond to the shot transition frame positions on the video timeline.
[0029] After obtaining the shared elements and fragment types, S1 proceeds to the detection and processing of the actual transmission relationship between tasks.
[0030] In S2, based on the result, the task to be executed after reading the result of the other task is determined from two tasks that record the same shared feature and whose position ranges intersect. The corresponding fragment of the shared feature recorded in the task to be executed is then copied. Subsequently, the same format probe value is inserted into the value fragment and only the image layer is generated. The reference fragment is pointed to a blank copy and only the reference is resolved to obtain the actual read address. The timing fragment is moved to the post-execution probe position and only the connecting frame is generated.
[0031] In S2, the completion order and result reading records are read from two tasks that record the same shared feature identifier. The task that actually reads the batch results of the other task is called the subsequent task. At this stage, the other task is only a possible data source and cannot be determined as the source task solely based on the order of execution. The result reading record is only used to narrow down the detection scope and is not used as proof of the specific feature source. After determining the subsequent task, the corresponding segment of the shared feature recorded in the subsequent task is copied to obtain a segment copy. The segment copy is stored separately from the formal task context and is used to carry the detection content. The detection results are not directly written into the formal video project.
[0032] For value fragments, values different from the original visual attribute values are selected from the allowed content of the same field as identically formatted probe values. The data type, field format, and interface reading method of the identically formatted probe values are consistent with the original visual attribute values, and their content is distinguishable from the original values. These probe values originate from the allowed content of the current field, without setting additional numerical thresholds. After placing the identically formatted probe values into the value fragment, only the image layer containing the shared features is generated. The image layer is an object layer or partial mask layer that contains the shared features and can be generated independently. By reading the generation record accompanying the image layer and directly comparing the visual attribute values in the generation record with the identically formatted probe values, it can be determined whether the image layer presents the probe values.
[0033] In conjunction with value fragment processing, for reference fragments, the field structure of the original referenced resource is copied, and the content field is set to empty, resulting in a blank copy. The blank copy retains the resource type, field structure, and accessible addresses, but does not store valid image, audio, model, or video content. The reference fragment is then pointed to the blank copy, and only the reference is resolved, recording the resource addresses accessed during resolution to obtain the actual read address. The actual read address is the resource location actually accessed by the subsequent task during this resolution, and can use an existing Uniform Resource Identifier (URI). By directly comparing the actual read address with the read address of the blank copy, it can be determined whether the reference fragment controlled the current resource read.
[0034] After reference detection is completed, for a time segment, a position that differs from the original time position and can generate a corresponding connecting frame is selected from the unoccupied frames on the timeline as the subsequent detection position, prioritizing unoccupied frames after the original time position. If there are no subsequent free frames at the end of the project, an unoccupied frame before the original position is selected. The time segment is moved to the subsequent detection position, and only connecting frames related to the shot's time change are generated. For example, if the original time position is frame 240, and frames 241 to 259 are occupied by other shots while frame 260 is free, frame 260 can be selected. By reading the time record carried by the connecting frame, it can be determined whether the subsequent detection position of the connecting frame has changed.
[0035] After obtaining the image layer, actual read address, or connecting frame, S2 checks and provides feedback item by item according to the actual existing fragment types. If the detection result reflects the corresponding detection content, it means that the current task directly uses the corresponding fragment in its own context, restores the fragment copy, and ends the source detection of that fragment. Since no inter-task transfer occurs, no element transfer relationship is formed, and subsequent change-type instructions can still directly determine the task as the local change seed based on the storage location of the target element in the task context. If the detection result does not reflect the corresponding detection content, only the corresponding fragment type enters the restoration and candidate source verification. The image layer, actual read address, or connecting frame corresponds to the actual existing fragment type; if any existing type does not meet the condition, the corresponding restoration process will be triggered.
[0036] For detection results in S2 that do not reflect the corresponding content, a recovery request is sent to the task that meets the conditions of object attribute main identifier, location range and fragment type; a candidate relationship is formed based on the returned local record, and the candidate relationship is verified while maintaining the original reading route to form the element transfer relationship from the source task to the subsequent execution task.
[0037] For each type of detection result obtained by S2, S3 determines whether restoration is needed. A value restoration request is issued when the image layer does not present detection values of the same format; an address reconnection request is issued when the actual read address does not point to a blank copy; and a time reset request is issued when the connecting frame does not subsequently execute a detection position change. Each type of request carries a project identifier, a shared feature identifier, a requested location range, and a fragment type. The shared feature identifier is the main identifier of the object attribute, and the requested location range is taken from the location range field of the detected fragment in the subsequent task.
[0038] It should be understood that recovery requests do not pre-specify source tasks. Candidate queries simultaneously satisfy the following conditions: identical project identifier, identical shared feature identifier, identical fragment type, and intersection of the location range of the candidate task's local record with the requested location range; the inclusion relationship falls under one type of intersection. Tasks meeting the above conditions are called candidate tasks during the source verification phase. Each candidate task returns the current local record, source location range, local record version, and task identifier. The executed task then writes each local record to the corresponding fragment copy. Candidate tasks that can complete the write operation according to fragment type form candidate relationships to be verified, and the source location range and local record version are saved along with the candidate relationship. Candidate tasks are obtained from independent task registration records; therefore, candidate discovery does not depend on the element transit relationships to be formed.
[0039] For shared elements with only one fragment type, only the corresponding feedback and recovery requests are processed. For example, the character setting task saves the color of the female protagonist's clothing, and the camera generation task also records the main identifier of the same clothing color object attribute, with the two location ranges intersecting. The clothing color fragment from the camera generation task is copied, and blue is selected from the allowed values in the color field as the probe value of the same format. Only the frame layer containing the female protagonist's clothing is generated.
[0040] For example, if the original clothing color is red, and the color field in the image layer generation record is also red, the shot generation task uses the shot area covered by the detected clothing color fragment as the requested location range and issues a value restoration request. When the clothing color record of the character setting task intersects with the requested location range, it enters the candidate range and returns the current clothing color record, the source location range, and the version. After writing the returned record to an isolated color fragment copy, a candidate relationship is formed. Subsequently, the value fragment element transfer relationship from the character setting task to the shot generation task is determined through the execution result of the original read route. The color field uses direct comparison between values, without setting color difference boundaries.
[0041] like Figure 2 As shown, the data transfer between the execution task, candidate task, candidate relationship record, and source copy after S3 represents the candidate relationship formation and verification process. For shared elements whose values, references, and time-series information are recorded separately by different context fragments, S3 independently probes and records relationships according to fragment type; the three types of information are located in three fragments respectively, which does not mean that a fragment belongs to all three types at the same time. First, the pre-probe record of each fragment to be probed is saved, and the actual existing fragments are copied to obtain context copies. At the same time, other context fragments that should not be changed in this probe, except for the currently probed fragment and its explicitly associated output, are read, and the current probe object is excluded according to the location range and fragment identifier; the original state digest is generated according to the identifier, content version, resource address, and time position of the remaining fragments. The original state digest is used to confirm whether the probe has changed irrelevant content. For example, the existing SHA-256 hash algorithm can be used to generate the digest according to a fixed field order.
[0042] Next, probes corresponding to the fragment type are performed in each context copy. If a frame layer does not present a probe value, only the pre-probe value record is removed from the input of the fragment copy containing that value, and a value restoration request is issued. If an actual read address does not point to a blank copy, only an address reconnection request is issued for the referenced fragment copy. If a connecting frame does not change with the probe position, only a time reset request is issued for the timing fragment copy. If one, two, or three feedbacks do not meet the conditions, the recovery process for the corresponding fragment type is initiated. Fragment copies that did not trigger a request are restored according to the pre-probe record and do not participate in this source determination.
[0043] After recovery is complete, each request is associated with the primary identifier of the shared feature's object attributes, the requested location range, and the corresponding fragment type to obtain a request record. The request record stores the recovery request, candidate tasks, the returned current local record, the source location range, the local record version, and the processing time. For each request, the candidate task identifier that can be written to the corresponding context copy according to fragment type is read, and it is checked whether the corresponding context copy has been recovered to the returned local record. Subsequently, the current state summary is generated using the same batch of other context fragments and the same field sequence used when generating the original state summary.
[0044] When a context copy is restored to the returned local record and the current state summary matches the original state summary, the fragment type, candidate task identifier, subsequent execution task identifier, object attribute primary identifier, source location range, receiving location range, local record version, and pending verification status are written to the candidate relationship record; the receiving location range is the requested location range. After the candidate relationship is confirmed, the intersection of the source location range and the receiving location range is taken as the relationship location range and written into the element transit relationship. Different fragment types form candidate relationships separately; when the same candidate task corresponds to two or more fragment types, the fragment types can be merged and written into the same composite candidate relationship record. If a fragment does not obtain a writable return record or the copy is not restored, only the corresponding fragment copy is restored and no corresponding candidate relationship is formed.
[0045] For each candidate relation to be verified, the corresponding fragment recorded in the candidate task is copied to obtain a source copy, and the local record version saved by the candidate relation is written to the base version field of the source copy. Subsequently, source probe content that differs from both the original content and the previous probe content is inserted into the source copy, and a source copy version is generated using the base version and the batch identifier from this verification execution record. The base version is used to correspond to the formal local record returned by the candidate task, and the source copy version is used to identify the content of this isolation verification; inserting probe content does not rewrite the local record version in the formal context of the candidate task. The source copy remains bound to the original logical read position of the candidate task and does not change the read configuration of subsequent tasks.
[0046] Subsequently, the task is executed again according to the original task schedule, maintaining the task connection, configuration reading, and invocation order. During execution, the context access address and actual input version of the task are recorded, the local processing result corresponding to the fragment type is obtained, and a verification status summary is generated according to the field range of the original status summary. During verification, it is first confirmed that the base version of the source copy is consistent with the local record version saved in the candidate relationship. Then, it is determined whether the context access address points to the logical position of the source copy, whether the actual input version is equal to the source copy version, whether the local processing result reflects the source probe content, and whether the verification status summary is consistent with the original status summary. When all conditions are met, the candidate relationship is confirmed, the intersection of the source location range and the receiving location range is written as the relationship location range, and an element transfer relationship corresponding to the fragment type is formed; if any condition is not met, the candidate relationship is deleted. The source copy version is only used in this verification, and the relationship record continues to be saved as the local record version of the base version; when the official local record version changes, the candidate query and verification are re-executed.
[0047] For example, the presentation of the female protagonist's image on screen can simultaneously involve the visual attribute values saved by the character setting task, the character resource addresses saved by the material management task, and the shot transition positions saved by the storyboard task. When all three types of probes trigger recovery, the three requests can respectively return the current local records and versions of the three tasks mentioned above. Based on this, the system forms three candidate relationships for different segment types and performs verification one by one while maintaining the original reading route. If two of them are actually provided by the same task, the corresponding segment types can be merged after verification. During the candidate discovery phase, return records where the version has changed are retained. The current source after material replacement, task rerun, or cache update can still be verified. When a relationship record is marked as invalid, the corresponding probe is re-executed, and the verified relationship records are reused in other processing rounds.
[0048] In S4, the target element identifier of the instruction is extracted using the element transfer relationship formed in S3, and the seed task is determined according to the instruction type. Read-type instructions are added to the source task in reverse, and change-type instructions are added to the receiving task in forward. The target element saving task that has not formed an outward relationship is retained as a local change seed until no new tasks are added to form a task group.
[0049] Upon entering S4, user input is received based on the relationship record obtained in S3. The user input is then combined with the current video project and the selected video segment to generate an instruction. The task group represents the set of tasks and task contexts that need to be invoked in this round of instructions; the current video project limits the project to which the production record belongs, and the selected video segment and the video position in the instruction together limit the location range for this round of processing. The operation type, target object, target attribute, and video position are parsed from the instruction; if the operation type changes, the target content is further parsed. The target object and target attribute are combined into the object attribute main identifier, and the video position is recorded as a location range field. These two fields together form the target element identifier; the target element identifier uses the same field structure as the element identifier in S1.
[0050] After completing the instruction parsing, the seed task is first determined. For read-type instructions, the task identifier that directly generates the output of the selected video segment is read from the output record, and the corresponding output task is used as the seed; for example, when querying the generated footage of the third scene, the shot generation task that directly generates that footage is used as the seed. For change-type instructions, task contexts with the same object attribute primary identifier are first filtered within the location range, and at least one task that saves the target element is determined as the seed. When the save task is the source task in the confirmed element transfer relationship, it is added to the receiving task in a forward direction starting from that task; when the save task has no outward confirmed relationship, but the detection result of S2 has already reflected the local segment content, the task is still directly added to the task group as a local change seed, but it is no longer traversed outward. Thus, even if the target element is only used locally by a certain task, a change seed can still be obtained. When different segment types are saved by different tasks, each save task is used as a seed, and other tasks with only similar identifiers are not pre-added to the task group.
[0051] like Figure 3 As shown, the task group formation process includes determining the seed task, determining the traversal direction according to the instruction type, and ending the traversal when no new tasks are added. After determining the seed task, the object attribute main identifier and target location range are read from the target feature identifier; only feature transfer relationships with the same object attribute main identifier and whose relationship location range intersects with the target location range are determined as relationships to be traversed, and both conditions must be met simultaneously. Read-type instructions add the source task in reverse from the output task seed along the relationship to be traversed; change-type instructions add the receiving task in forward from the target feature save task with outward transfer relationship. After each task is added, the connected relationships are filtered and tasks are deduplicated with the newly added task until no new tasks are generated in this round of filtering. When the target location range contains multiple non-intersecting target sub-intervals, traversal is performed for each target sub-interval separately.
[0052] For any target sub-interval, the intersection of the position range of the first relationship to be traversed and the target sub-interval is determined as the current valid position range; the current valid position range represents the video position commonly corresponding to all relationships in the current traversal chain. When continuing to filter relationships with newly added tasks, connected tasks are added only when the object attribute main identifier of the next relationship is the same and the relationship position range intersects with the current valid position range, and the current valid position range is updated to the intersection of the two; if the intersection is empty, no task is added along the next relationship. Each target sub-interval uses an independent traversal queue and visited task record. After traversal, the task sets of each sub-interval are merged into a task group; when the same task appears in multiple sub-interval task sets, the task identifier is deduplicated and the corresponding record of each target sub-interval and the corresponding context segment is retained.
[0053] For example, the target location range includes two separate scenes, the third and fifth. The editing task saves the context fragments of the third and fifth scenes respectively. When the location ranges of the third and fifth scenes do not intersect, traversal starting from the target sub-interval of the third scene, even if an editing task is added, will not continue along the fifth scene's relationship to add downstream tasks; the task chain corresponding to the fifth scene is traversed separately by the target sub-interval of the fifth scene. When an editing task appears in both sub-interval task sets simultaneously, the merged task group only removes duplicate task identifiers, still retaining the corresponding records of the context fragments of the third and fifth scenes respectively, so that the context is read according to the corresponding target sub-interval during task execution.
[0054] Furthermore, after obtaining the task group in step S4, the following task execution and video content update process begins.
[0055] In practice, after the task group is determined, the task execution order is arranged according to the element transfer relationship in step S5, and the instructions are executed in the context of the task group; for the shared elements that have changed, the video content is updated according to the segment type and the video production result is output.
[0056] In detail, after the task group is determined, in step S5, the execution order of each task is first determined along the verified element transfer relationship, so that the execution result of the source task is processed before the corresponding receiving task. For example, existing topological sorting methods can be used to arrange the tasks in the relationship record; tasks at the same level that do not have a direct element transfer relationship are executed according to their respective contexts. Instructions and the context of each task are sequentially input into the corresponding task, causing the receiving task to read the execution result of the source task and obtain the task execution result. The execution result record for the task corresponding to a read instruction is the shared element read; the execution result record for the task corresponding to a change instruction is the changed shared element, fragment type, and target content.
[0057] Subsequently, the modified shared elements, segment types, and target content are read from the task execution results of the modification instructions. When the shared element is located in a value segment, the target content is converted into picture parameters and the corresponding picture layer is updated; when it is located in a reference segment, the target content is converted into a resource address, the read address is updated, and the resource content corresponding to the resource address is generated; when it is located in a time segment, the target content is converted into shot time parameters, the time position is updated, and a transition frame is generated. When a shared element corresponds to multiple segment types, the corresponding updates are performed separately. After completing the classification updates, the updated picture layer, the resource content corresponding to the resource address, or the transition frame are combined with the unchanged video content, and the picture parameters, the updated read address, or the time parameters are written into the task group to record the corresponding context segment of the shared element, thus obtaining the video production result.
[0058] For example, the user first inputs which character resource is currently being used by the female lead in the third scene. The user input, combined with the current project and the video clip from the third scene, generates a read command. This command is then parsed to obtain the female lead, the character resource, and the location range of the third scene, forming the primary identifier of the character resource object attribute. The system determines the shot generation task as the seed from the output record of the third scene video clip, and then reverses the process along the element transfer relationship of the character resources, adding the verified character setting task or material management task. During execution, the resource record of the source task is first retrieved, and then the shot generation task returns the currently used character resource information, recording the shared elements read in the task execution result.
[0059] To further illustrate, consider the user's input: change the female lead's red dress in the third scene to a blue short skirt while maintaining consistency in subsequent shots. This input forms a change instruction, which is parsed to obtain the female lead, clothing color and style, the location range of the third scene and subsequent shots, and the target content of the blue short skirt. The system uses the character setting task, which stores the main identifier of the clothing color object attribute, as the seed. If the character setting task has a verified outward propagation relationship, it is forward-added to the third scene shot generation task and the subsequent shot generation task within the location range. If the user only modifies the locally generated character preview of the character setting task, and the detection results already reflect the local clothing fragment, then the character setting task directly enters the task group as a local change seed, without needing to first establish an element propagation relationship pointing to other tasks.
[0060] When executing a change command, first convert the blue skirt to the color and clothing parameters used in the character's visual layer, and update the value fragment of the character setting task. If the target content contains new character assets, write the corresponding resource address into the reference fragment and generate the resource content corresponding to the resource address. If the shot transition time is affected by changes in clothing movement, write the shot timing information in the target content into the timing fragment and generate the transition frame. Then, execute the task sequentially, reading the execution results of the previous task to obtain the task execution results that record changes to shared elements, fragment types, and target content. Finally, combine the updated content with unchanged content such as the scene, voice-over, and background music, and write the updated parameters back to the corresponding context fragment in the task group, outputting a video production result where the clothing is consistent in the third scene and subsequent shots.
[0061] The task-driven video production group dialogue management method provided in this application has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.
Claims
1. A task-driven video production group dialogue management method, characterized in that, include: Obtain the task context and feature identifier set; the feature identifier consists of the object attribute primary identifier and the location range field; identify features with the same object attribute primary identifier and overlapping location ranges in at least two tasks as shared features, and classify the shared feature fragments into value fragments, reference fragments or time series fragments according to the record format; Based on the results, read the records to determine the task to be executed after reading the results of another task; After copying, execute the corresponding segment of the task and probe according to the segment type to obtain the image layer, actual read address or connecting frame; If any probe result does not reflect the corresponding content, the location range of the probed fragment is taken as the request range, and a recovery request is sent to the task with the same object attribute main identifier and fragment type and whose local record range intersects with the request range; the task that returns the record and recovers the fragment copy is taken as the candidate source, the source probe content is put into the candidate source fragment copy and executed according to the original read route; the source task is confirmed according to the context access record, input version and local results, and the feature transfer relationship is formed; Extract the target element identifier of the instruction; read-type instructions are added to the source task in reverse from the selected fragment output task, and change-type instructions are added to the receiving task in forward from the target element saving task; when there is no outward relationship, the target element saving task is retained, and a task group is formed when no new tasks are added. Execute instructions within the task group context, update video content according to the segment type of the changed elements, and output the video production result.
2. The method according to claim 1, characterized in that, Detect fragment copies by fragment type and form feature transitive relationships, including: Save the pre-detection records of each segment to be detected in the task and the original state summary of other context segments that should not be changed in this detection, and copy each segment to be detected to obtain a context copy; Each context copy is independently subjected to the corresponding probe; if the result of any probe does not reflect the corresponding content, the task is blocked and then executed to read the corresponding pre-probe record and send a recovery request. The task that returns a local record and restores the context copy is identified as a candidate source for the corresponding fragment type, forming a candidate relationship to be verified; candidate relationships for different fragment types are retained independently, while candidate relationships corresponding to the same candidate source are merged. For each candidate relation, the corresponding fragment of the candidate source is copied to obtain a source copy, and the local record version of the candidate relation is used as the base version. After the source probe content is inserted, a source copy version is generated based on the base version and the current verification batch, and the task is executed according to the original read route. When the base version is consistent with the local record version of the candidate relation, the input version is the source copy version, the context access record points to the source copy, the local result reflects the source probe content and the original state summary is unchanged, the candidate relation is confirmed and an element transit relationship is formed; otherwise, the candidate relation is deleted and the context is restored.
3. The method according to claim 2, characterized in that, The task of sending a recovery request and returning a record and a copy of the recovered fragment as a candidate source includes: The location range of the detected segment is used as the request range; the task registration records are queried according to the project identifier, the object attribute master identifier of the shared element, and the segment type; the tasks whose local record range intersects with the request range are identified as candidate tasks, and a recovery request is sent. Receive the current local record that matches the project identifier, object attribute main identifier, and fragment type, as well as the source range, local record version, and task identifier of the current local record; Each local record is written to its corresponding fragment copy. Tasks that can complete the writing are identified as candidate sources, and the source range, the receiving range represented by the request range, and the local record version are written into the candidate relationship. If no local record that can be written is available, restore the corresponding fragment copy without forming a candidate relationship.
4. The method according to claim 3, characterized in that, The step of inserting source detection content into the candidate source fragment copy and executing according to the original read route includes: Copy the corresponding fragment of the shared element in the candidate source to obtain the source copy, and use the local record version in the candidate relationship as the base version; insert the source probe content that is different from the previous probe content, and generate the source copy version with the base version and the current verification batch; Keep the read configuration, task scheduling, and call order of the subsequent task unchanged, execute the subsequent task, and record the context access address and actual input version; When the base version is consistent with the local record version in the candidate relationship, the context access address points to the source copy, the actual input version is the source copy version, and the local result reflects the source probe content, the candidate relationship is confirmed, and the intersection of the source range and the receiving range is taken as the relationship range and an element transfer relationship is formed. Otherwise, delete the candidate relationship and restore the context of the candidate source and the subsequent task.
5. The method according to claim 4, characterized in that, Based on the instruction type, seed tasks are determined and task groups are formed by traversing the element transit relationships, including: The task that directly generates the corresponding output is determined from the output record of the selected video segment, serving as the seed task for the read-type instruction; Identify at least one task that saves target elements from the task context; use the target element saving task that is the source task in the confirmed relationship as the change seed for outward traversal, and retain the target element saving task that has not formed an outward relationship in the task group to perform local changes; Read the object attribute master identifier and target range from the target feature identifier; filter the feature transfer relationships from the relationship records that have the same object attribute master identifier and whose relationship range intersects with the target range; For read-type instructions, the source task is added in reverse from the seed task; for change-type instructions, the receiving task is added in forward from the change seed, and the process continues with the newly added tasks until no new tasks are added.
6. The method according to claim 1, characterized in that, The process of obtaining the task context and feature identifier set and dividing the shared feature fragments includes: Production records are aggregated from video production projects according to task identifiers to obtain task context and element identifier sets; elements with the same object attribute main identifier and overlapping location ranges are identified as shared elements; the visual attribute values, resource read addresses, or shot transition time positions of the shared elements are divided into value segments, reference segments, or time sequence segments; the completion order is read from the task execution records, and the task is executed after determining the result read records indicating the received task and execution batch.
7. The method according to claim 5, characterized in that, The target element identifier of the extraction instruction includes: Receive user input, combine the user input with the current video project and the selected video segment to obtain instructions; parse the operation type, target object, target attribute, video location and target content corresponding to the change operation from the instructions, combine the target object and target attribute into the object attribute main identifier, and use the video location as the location range to obtain the target element identifier.
8. The method according to claim 7, characterized in that, The target element identifier of the extraction instruction also includes: By limiting the time interval according to the selected video clip and location range, filter the task context with the same target object within the time interval, and obtain the feature identifier with the same object attribute main identifier from the filter results to obtain the target feature identifier.
9. The method according to claim 8, characterized in that, The execution of instructions in a task group context includes: The execution order of each task in the task group is determined by the element transfer relationship. The instructions and task context are sequentially input into the corresponding tasks so that the receiving task reads the execution result of the source task. For read-type instructions, the shared elements read are recorded. For change-type instructions, the changed shared elements, fragment types and target content are recorded to obtain the task execution result.
10. The method according to claim 9, characterized in that, The process of updating video content according to the segment type of the changed elements and outputting the video production result includes: Read the changed shared elements, segment types, and target content from the task execution results. Convert the target content into picture parameters, resource addresses, or time parameters, and update the resource content or transition frames corresponding to the picture layer and resource address, respectively. Combine the updated content with the unchanged video content and write it into the corresponding context segment to obtain the video production result.
