Automobile marketing video generation method and system based on diffusion model

Abstract

The invention discloses an automobile marketing video generation method and system based on a diffusion model. The method comprises the following steps: firstly, performing deep feature extraction on an official photo of a target vehicle, and constructing a brand landmark feature parameter set; generating a parameterized motion template based on a marketing script, and generating scene static graphs in batches through an image diffusion model; performing brand consistency closed-loop verification on the static graph by adopting a feature matching algorithm, and automatically regenerating if the brand consistency does not reach the standard; encoding the static image and the motion template which pass the verification, inputting the encoded static image and motion template into a video diffusion model, constraining time sequence consistency through a cross-frame attention mechanism, and generating a video clip; and based on scene similarity, adaptively selecting a transition mode to perform intelligent splicing, and adding brand elements to synthesize a final video. According to the invention, brand security, parameterized fine control and intelligent reuse of generated assets of the automobile marketing video are realized, and the generation efficiency and quality are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence video generation, in particular to a car marketing video generation method and system based on a diffusion model. BACKGROUND

[0002] With the rapid development of generative artificial intelligence technology, image and video generation technology based on diffusion model has become an important tool in the field of content creation. In the car marketing industry, traditional marketing video production mainly relies on real scene shooting or three-dimensional animation rendering, which has the pain points of long production cycle (usually 3-7 days), high cost (single video cost tens of thousands to hundreds of thousands of yuan), slow creative iteration, etc. Although the existing technology has made preliminary attempts to generate static images using Stable Diffusion and then convert them into videos through Stable Video Diffusion (SVD), it still faces the following core technical bottlenecks:

[0003] Brand feature consistency out of control: During the generation process, the brand logo features of the vehicle (such as the front grille, the outline of the headlights, and the brand logo) are prone to distortion or drift, leading to brand infringement risks or decreased visual recognition. The existing technology simply uses official photos as reference input, lacking a systematic feature extraction and closed-loop verification mechanism.

[0004] Motion control granularity is coarse: The parameters such as camera movement, vehicle trajectory, and environmental dynamics rely on vague descriptions of natural language prompts (Prompts), making it difficult to achieve precise frame-level quantitative control. The generated video segments have problems such as unprofessional camera movement and motion jitter.

[0005] Generation efficiency and cost paradox: Although single generation is fast, each video needs to be re-inferred and calculated, which cannot save historical generation results. For different versions of marketing videos of the same car model, repeated calculations lead to resource waste, and cannot achieve marginal cost reduction in mass production.

[0006] Engineering deployment difficulties: Diffusion model inference requires high memory usage and long time consumption, and traditional monolithic architecture cannot support batch task parallelism. There is a lack of resource scheduling optimization solutions for the high-frequency and diverse needs of car marketing.

[0007] To address the above problems, the present application proposes an engineering system with brand feature closed-loop verification, parameterized motion template control, and material library incremental learning reuse, which realizes low-cost, high-efficiency, and brand-safe mass production of car marketing videos. SUMMARY

[0008] The present application aims to provide a car marketing video generation method and system based on a diffusion model, thereby solving the aforementioned problems in the prior art.

[0009] To achieve the above object, the technical scheme adopted by the present application is as follows:

[0010] In a first aspect, the present application discloses a vehicle marketing video generation method based on a diffusion model, comprising the following steps:

[0011] Obtain the standardized official photos of the target vehicle, analyze the vehicle brand logo feature parameter set through a deep feature extraction network, and the feature parameter set includes the vehicle body contour topology vector, the front grille texture feature, the vehicle lamp contour feature, and the brand logo structure feature;

[0012] Divide the video into several narrative units based on the marketing script, generate scene description text and vehicle pose description text for each narrative unit, and construct a parameterized motion template containing lens motion type, vehicle trajectory parameters, environmental dynamic effects, and prompt word weight scheduling;

[0013] Input the official photos and scene description text into the image diffusion generation model, introduce geometric constraints and depth constraints through ControlNet architecture, and batch generate scene static images with consistent brand features;

[0014] Perform brand feature consistency verification on the generated scene static images, extract the vehicle region feature vector in the generated image using a feature matching algorithm, and perform multi-dimensional similarity evaluation with the brand logo feature parameter set. If the evaluation result does not meet the preset consistency standard, it is regenerated to form a generation-verification closed loop control;

[0015] Encode the scene static images that pass the verification and the parameterized motion template into spatiotemporal latent variables that can be analyzed by the video diffusion model, input them into the video diffusion generation model, constrain the temporal consistency through the cross-frame attention mechanism, and generate video clips corresponding to the narrative units;

[0016] Use an adaptive transition decision mechanism to intelligently splice the video clips, automatically select the transition mode based on the scene structure similarity and motion vector continuity of adjacent clips, synthesize the preliminary marketing video and add standardized brand elements, and output the final vehicle marketing video product.

[0017] Further, the extraction step of the brand logo feature parameter set specifically includes:

[0018] Perspective correction and illumination normalization are performed on the official photos to eliminate the interference of shooting angle and light; a target detection model is used to accurately locate the vehicle main body area and extract the region of interest image; a residual neural network is used to extract the deep semantic features of the region of interest image to generate high-dimensional feature representation; principal component analysis is performed on the high-dimensional features to reduce the dimension, and the core dimensions whose feature contribution rate meets the preset requirements are retained to construct the brand logo feature parameter set.

[0019] Further, the parameterized motion template adopts a structured data format description, including the following core fields:

[0020] camera_motion: defines the lens motion mode, including basic types of motion such as static, horizontal pan, vertical tilt, zoom in and out, etc.

[0021] car_trajectory: describes the vehicle motion trajectory parameters, covering the starting spatial coordinates, terminal spatial coordinates, motion speed curve and path smoothness.

[0022] environment_effect: configures environmental dynamic effects, including wind, sand, rain, snow intensity, light change gradient, particle special effect density.

[0023] prompt_weight_schedule: sets the timing prompt word weight scheduling sequence, used to dynamically adjust the weight ratio of scene features and vehicle features in the video generation process.

[0024] Further, the closed-loop control mechanism for brand feature consistency verification is as follows:

[0025] The semantic segmentation model is used to automatically separate the vehicle foreground and scene background in the generated image, and extract the vehicle foreground area. The vehicle foreground area is rotated and transformed at multiple angles to generate a set of multi-angle feature vectors. The cosine similarity between the multi-angle feature vector set and the brand logo feature parameter set is calculated, and the average value is taken as the consistency score. If the consistency score is lower than the preset threshold, a re-generation request is triggered and the constraint strength parameter of the image diffusion generation model is adjusted until the score meets the standard.

[0026] Further, the method further comprises:

[0027] A video clip material library is established, and the historical generated video clips are semantically labeled and vector fingerprint extracted according to scene type, vehicle brand and motion feature. In a new video generation task, the existing clips matching the current narrative unit are preferentially retrieved from the material library, and the historical clips meeting the preset requirements are directly reused, and only the unmatched clips are executed for the generation process. The material library is continuously optimized in an incremental learning manner, and the high-quality clips generated and passing the quality evaluation are updated into the library, and the vector fingerprint extraction model is dynamically adjusted based on user feedback.

[0028] Further, the logic of the adaptive transition decision mechanism is as follows:

[0029] Calculate the structural similarity of the last frame and the first frame of the adjacent video segment. If the similarity is higher than a first preset threshold, a direct cut transition is adopted; if the similarity is between the first threshold and a second threshold, a dissolve transition is adopted, and the dissolve duration is adaptively calculated according to the motion vector difference; if the similarity is lower than the second threshold or the scene switching flag is activated, a sliding transition is adopted, and the sliding direction is geometrically consistent with the overall motion vector field direction of the vehicle.

[0030] In a second aspect, the application further discloses a car marketing video generation system based on a diffusion model, comprising the following functional modules:

[0031] A brand feature constraint source management module is used to store and preprocess standardized official photos of a target vehicle, extract and maintain a set of brand signature feature parameters;

[0032] An intelligent split-screen script analysis module is used to decompose a marketing script into narrative units, generate scene description text and vehicle pose description text;

[0033] A batch image generation engine integrates an image diffusion model and a control network to perform a batch generation task of scene static images;

[0034] A brand consistency closed-loop verification module has a built-in feature matching algorithm and a similarity evaluator to perform consistency verification and feedback control on generated static images;

[0035] A parameterized motion template configurator provides a visual interface for users to edit shot motion, vehicle trajectory, environmental effects and prompt word weight scheduling parameters;

[0036] A video diffusion generation engine integrates a video diffusion model and a timing control framework to convert static images and motion templates into video segments;

[0037] An intelligent splicing and post-synthesis module has a built-in adaptive transition decision maker to perform segment splicing, narration synthesis, watermark superposition and subtitle rendering;

[0038] A material library management and optimization module is used for vector fingerprint storage, similarity search, reuse scheduling and incremental learning update of video segments.

[0039] Further, the batch image generation engine further comprises:

[0040] A distributed task scheduling unit realizes multi-task parallel processing and load balancing based on a message queue; a dynamic video memory management unit supports large-scale batch generation under limited video memory resources through model layering offloading and gradient checkpointing technology; a generation quality monitoring unit calculates image quality evaluation indicators in real time and outputs a quality report.

[0041] Further, the video diffusion generation engine further comprises:

[0042] The timing consistency maintaining unit constrains the inter-frame vehicle feature drift and motion jitter through a cross-frame attention mechanism and an optical flow re-projection loss function.

[0043] The multi-resolution generation pipeline supports multiple output resolutions from standard definition to ultra-high definition.

[0044] The failure retry and parameter adaptive unit automatically adjusts the motion amplitude decay factor and triggers re-generation when the generated video segment quality score does not meet the standard, and the number of retries is limited to a preset upper limit.

[0045] Further, the system adopts a micro-service architecture deployment, and each functional module communicates through standardized API interfaces and asynchronous message queues for loose coupling, supports containerized orchestration and elastic scaling, and completes multiple style version automobile marketing video generation tasks within a specified time.

[0046] The beneficial effects of the present application are:

[0047] 1. Fundamental improvement of brand safety

[0048] Through standardized extraction and closed-loop verification mechanism of brand feature parameter set, consistency control is upgraded from artificial sampling inspection to algorithm hard constraint. The vehicle brand logo elements are always consistent with the official photos during the generation process, effectively avoiding the risk of feature drift, meeting the strict requirements of brand compliance of automobile enterprises, and providing safety guarantee for the commercial application of generative AI in automobile marketing field.

[0049] 2. Revolutionary improvement of motion control accuracy and reusability

[0050] Innovatively, the lens language specification of the film and television industry is deconstructed into a structured JSON template, realizing accurate quantitative configuration of lens motion, vehicle trajectory, environmental effect and timing weight. Creative experience can be deposited as programmable and version-managed technical assets, and non-professionals can also batch reuse professional-level camera operation schemes, significantly reducing the creative threshold and personnel dependence.

[0051] 3. Significant optimization of generation efficiency and cost structure

[0052] By constructing the spatio-temporal semantic fingerprint and vector retrieval system of the video segment, the historical generated content can be intelligently reused. When creating the same type of vehicle, existing assets are preferentially called to avoid repeated reasoning and calculation, significantly shortening the production cycle; combined with the distributed parallel capability of micro-service, multiple versions can be batched and output within the business time limit, realizing the scale economy effect of continuously decreasing marginal cost.

[0053] 4. Systematic enhancement of video quality and professionalism

[0054] The adaptive transition decision algorithm automatically selects the optimal transition method based on scene structure similarity to ensure natural and smooth segment transitions; the cross-frame attention mechanism and optical flow constraints effectively suppress temporal jitter, and the motion smoothness and visual coherence of the generated video segments reach commercial-grade standards, significantly improving overall quality.

[0055] 5. Comprehensive construction of engineering implementation capabilities

[0056] The microservice architecture designed for the high memory usage and long latency characteristics of diffusion models achieves high concurrency and low latency industrial deployment capabilities under limited hardware resources through optimization techniques such as dynamic memory management, model layered unloading, and continuous batch processing. This provides a solid engineering foundation for the technology solution to move from the laboratory to large-scale production.

[0057] 6. Outstanding cross-scenario scalability and commercial value

[0058] Although this invention is optimized for automotive marketing scenarios, its core modules, such as brand feature management, parameterized templates, and material library reuse, are universal and can be quickly migrated to marketing fields of high-value goods such as real estate, luxury goods, and electronic products, forming a reusable technology platform with broad commercial application prospects and industrial driving effects. Attached Figure Description

[0059] Figure 1 This is a flowchart of a method for generating automotive marketing videos based on a diffusion model, according to the present invention.

[0060] Figure 2 This is an architecture diagram of a car marketing video generation system based on a diffusion model, according to the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] Firstly, referring to Figure 1 A method for generating automotive marketing videos based on a diffusion model is disclosed, comprising the following steps:

[0063] S1. Obtain standardized official photos of the target vehicle and analyze the vehicle brand's signature feature parameter set through a deep feature extraction network. The feature parameter set includes the vehicle body outline topology vector, front grille texture features, headlight outline features, and brand logo structure features.

[0064] This step specifically involves converting official vehicle photos into quantifiable digital assets representing brand characteristics, providing an immutable constraint benchmark for subsequent generation processes. First, standardized official photos of the target vehicle provided by the automaker are input. These photos typically meet requirements such as standardized shooting angles, uniform lighting conditions, and solid background colors, ensuring that feature extraction is not affected by external interference. The system performs basic preprocessing on the photos, eliminating perspective distortion through geometric transformations to present a standard orthographic projection view of the vehicle body; simultaneously, it performs lighting normalization to balance brightness differences in different areas, preventing shadows or highlights from obscuring key details.

[0065] After preprocessing, a deep feature extraction network performs layered parsing of the vehicle image. This network employs a multi-layered perceptual architecture, with the bottom layer focusing on local detail textures, aggregating layer by layer to form component-level semantics, and finally constructing an abstract representation of the entire vehicle. During the parsing process, the system automatically locks the main vehicle region, removes background noise interference, and focuses on processing effective pixels within the vehicle's dimensions.

[0066] After forward computation, the original pixel information is mapped into a compact set of feature parameters. This set contains four core elements: the vehicle body contour topological vector depicts the overall shape curves and posture; the front grille texture features encode the air intake arrangement pattern and chrome trim details; the headlight contour features capture the edge geometry of the headlight assembly and the internal LED light strip structure; and the brand logo structural features accurately record the outline of the logo and the text layout. These parameters together constitute the vehicle brand's digital identity card. Any output deviating from this parameter set during subsequent generation will be judged as inconsistent and needs to be regenerated, thus eliminating the risk of brand feature distortion or copyright disputes at the algorithmic level.

[0067] S2. Based on the marketing script, the video is divided into several narrative units. Scene description text and vehicle posture description text are generated for each narrative unit. A parameterized motion template is constructed that includes camera movement type, vehicle trajectory parameters, environmental dynamic effects and cue word weight scheduling.

[0068] This step specifically involves converting the marketing creative script into a set of parameterized instructions that can be executed by the machine. The system receives a natural language script written by marketers, covering elements such as the video theme, style, tone, and plot progression. The intelligent parsing engine performs semantic understanding on the script and automatically divides it into multiple narrative units based on scene transitions, emotional shifts, or logical progressions. Each unit corresponds to an independent segment in the video.

[0069] For each narrative unit, the system generates two types of descriptive text: scene description text uses natural language to depict the atmosphere of the environment, such as the vastness of the desert, the prosperity of the city, or the treacherousness of the mountain road; vehicle posture description text defines the specific presentation of the vehicle in the unit, including the initial position, orientation angle, driving action, and spatial relationship with other elements.

[0070] Building upon this foundation, a parametric motion template is constructed, transforming artistic creative needs into structured control parameters. This template comprises four core dimensions: camera motion type, describing the camera's physical movement and encompassing basic camera movements such as stillness, panning, tilting, and zooming; vehicle trajectory parameters, defining the vehicle's path and rhythm within the frame using coordinates and velocity curves; environmental dynamic effects, quantifying the intensity and evolution of natural phenomena such as wind, rain, snow, and changes in lighting; and cue word weighting, setting the relative importance of each element at different times to achieve a smooth transition of focus. This template-based design solidifies creative experience into programmable, reusable, and version-manageable technical assets.

[0071] S3. Input the official photos and scene description text into the image diffusion generation model, and introduce dual control of geometric constraints and depth constraints through the ControlNet architecture to generate static scene images with consistent brand characteristics in batches.

[0072] This step specifically involves: merging brand constraints with creative needs to drive the diffusion model to batch-produce static scene images that meet the requirements. The system simultaneously receives two types of input: first, pre-processed standardized official photos, carrying precise brand visual characteristics; and second, scene description text, defining the style, lighting, atmosphere, and other artistic elements of the target image in natural language. These two elements are deeply integrated within the model through a cross-attention mechanism, ensuring that the generation process both adheres to brand characteristics and achieves creative expression.

[0073] The image diffusion generation model is based on a latent diffusion architecture, performing denoising within the latent space. To ensure the output vehicle maintains a high degree of consistency with the official photograph in terms of structure and spatial relationships, a ControlNet architecture is introduced to implement dual control. The geometric constraint branch performs edge detection on the official photograph, extracting key geometric lines such as body contour lines, window creases, and trim segmentation lines to form hard edge constraints, preventing distortion or disproportion of the vehicle's shape during generation. The depth constraint branch analyzes the official photograph through a depth estimation network, constructing a 3D spatial depth map to clarify the front-rear positional relationships of components such as the front, doors, and roof, ensuring that the generated image of the vehicle has correct spatial depth and a sense of three-dimensionality.

[0074] Under the combined effect of dual constraints, the model simultaneously references the semantic guidance of the text description, the structural boundaries of the geometric constraints, and the spatial hierarchy of the depth constraints in each denoising step, generating multiple candidate scene static images in batches. The vehicle brand features in these images maintain topological consistency with the official photos, while the scene environments are diverse and generated by the algorithm, avoiding copyright issues and providing high-quality, highly consistent visual material for subsequent video production.

[0075] S4. Perform brand feature consistency verification on the generated static scene image. Use feature matching algorithm to extract vehicle area feature vectors in the generated image and perform multi-dimensional similarity evaluation with the brand's iconic feature parameter set. If the evaluation result does not meet the preset consistency standard, regenerate to form a generation-verification closed-loop control.

[0076] This step specifically involves: establishing an automated quality control mechanism to create a closed-loop system that ensures brand characteristics are not distorted. The system reviews each static scene image output in batches by the image diffusion model. First, a high-precision semantic segmentation model automatically identifies and separates the vehicle foreground area from the background environment in the image, accurately locating the vehicle's pixel range and eliminating background interference.

[0077] After obtaining the vehicle region, a feature matching algorithm is used to perform multi-dimensional deep analysis of the region, extracting high-dimensional feature vectors covering contour, texture, and structure. The contour dimension captures the curvature and continuity of the body lines; the texture dimension encodes the fine arrangement and gloss reflection of the front grille; and the structure dimension analyzes the geometric shape of the headlight contours and the spatial arrangement of the brand logo. These features together constitute the digital fingerprint of the generated vehicle.

[0078] The system performs a multi-dimensional similarity assessment between the fingerprint and the brand identity feature parameter set established in Step 1, comparing them layer by layer from global shape to local details. The assessment process comprehensively considers the degree of feature similarity and the completeness of key element preservation, forming a quantitative consistency score. This score is compared with a preset strict quality standard; if it is lower than the standard threshold, it indicates that the brand features of the generated image have deviated or been distorted.

[0079] At this point, the system automatically triggers a regeneration request and dynamically adjusts the constraint strength parameters of the image diffusion model to enhance adherence to the features of the official photo. It then returns to the generation stage to re-execute. This closed-loop process iterates continuously until the similarity assessment of the output image meets the standards, ensuring that every static image entering subsequent processes strictly conforms to the brand's visual specifications, thus eliminating the risk of damage to brand assets from an algorithmic perspective.

[0080] S5. The verified static scene image and the parameterized motion template are jointly encoded into spatiotemporal latent variables that can be parsed by the video diffusion model. The video diffusion generation model is then input into the model, and the temporal consistency is constrained through the cross-frame attention mechanism to generate video segments corresponding to the narrative units.

[0081] This step specifically involves: imbuing static images with dynamic vitality, transforming them into temporally coherent video clips. The system selects a verified static image of the scene as the first frame of the video, ensuring the accuracy of brand characteristics. The camera motion type, vehicle trajectory parameters, environmental dynamic effects, and temporal weighting scheduling information from the parametric motion template are uniformly encoded and converted into control signals describing temporal evolution. These signals are deeply integrated with the spatial features of the static image in a deep network, jointly constructing a spatiotemporal latent variable containing spatial details and temporal evolution. This variable fully characterizes the entire cycle of change in the video from the first frame to the last frame.

[0082] A spatiotemporal latent variable input video diffusion generation model is proposed. This model is based on a latent diffusion architecture and performs an iterative denoising process in a spatiotemporal joint latent space. To ensure the inter-frame coherence of the generated video, the model introduces a cross-frame attention mechanism, so that the generation of the current frame not only depends on its own noise state, but can also actively refer to the feature information of the preceding and following frames. This inter-frame information interaction effectively constrains the frame-by-frame consistency of the vehicle's appearance, avoiding appearance flickering or distortion; at the same time, it smooths the camera motion trajectory, eliminating jumps or jitter; the environmental dynamic effects also follow the temporal logic to evolve naturally, and the changes in wind and sand particles and lighting present physically reasonable continuity.

[0083] Based on the duration and frame rate set by the motion template, the model generates a complete video sequence frame by frame, ultimately outputting a video clip that perfectly matches the duration, rhythm, and camera movement requirements of the narrative unit. In this clip, the vehicles strictly maintain their brand characteristics, their movement trajectories are smooth and natural, and the scene dynamics conform to physical laws, making it directly usable for subsequent stitching.

[0084] S6. Adopt an adaptive transition decision mechanism to intelligently splice video segments. Based on the scene structure similarity and motion vector continuity of adjacent segments, the transition method is automatically selected to synthesize a preliminary marketing video and add standardized brand elements to output the final automotive marketing video product.

[0085] This step specifically involves: seamlessly connecting video clips and post-production branding to generate a ready-to-use marketing product. After receiving video clips generated by each narrative unit, the system first performs intelligent analysis on adjacent clips. By calculating the structural similarity between the last frame of the previous clip and the first frame of the next clip at multiple scales, it assesses the coherence of the two in visual elements such as composition, color, and texture; simultaneously, it analyzes the continuity of motion vectors between the two frames to determine whether the direction and speed of camera movement transition smoothly.

[0086] Based on the above analysis, the system's built-in adaptive transition decision-making mechanism automatically selects the optimal transition method. If the scene similarity is high and the movement is continuous, a direct cut is used to achieve a crisp, instantaneous transition; if the similarity is moderate, a dissolve transition is enabled, using transparency transitions to achieve a smooth connection, with the dissolve duration dynamically adjusted according to the difference in movement; if the similarity is low or the scene changes significantly, a sliding transition is executed, with the sliding direction geometrically consistent with the overall movement trend of the vehicle to create smooth visual guidance.

[0087] After determining the transition method, the system strings together all segments in narrative order to form a preliminary marketing video. Standardized post-production compositing then follows: a text-to-speech engine generates professional narration synchronized with the visuals; a dynamic brand watermark with varying transparency is overlaid in the lower right corner of the screen; and timecode-aligned subtitles are intelligently rendered based on the audio content. The final output is a video file that meets commercial placement standards, ensuring that the encoding format, resolution, and color space meet the requirements for multi-platform distribution, resulting in a finished automotive marketing video ready for direct dissemination.

[0088] Furthermore, the specific steps for extracting the brand's iconic feature parameter set include:

[0089] Perspective correction and illumination normalization are performed on official photos to eliminate interference from shooting angle and lighting; a target detection model is used to accurately locate the main body area of ​​the vehicle and extract the region of interest (ROI) image; deep semantic features of the ROI image are extracted through a residual neural network to generate a high-dimensional feature representation; principal component analysis is performed on the high-dimensional features to reduce dimensionality, retaining the core dimensions whose feature contribution rate meets the preset requirements, and constructing a set of brand-identical feature parameters.

[0090] Furthermore, the parametric motion template is described using a structured data format, containing the following core fields:

[0091] camera_motion: Defines the camera motion mode, including basic camera movement types such as stationary, horizontal pan, vertical tilt, zoom in and zoom out;

[0092] car_trajectory: Describes the vehicle's trajectory parameters, including the starting spatial coordinates, ending spatial coordinates, velocity curve, and path smoothness.

[0093] environment_effect: Configures dynamic environmental effects, including wind, sand, rain and snow intensity, light change gradient, and particle effect density;

[0094] prompt_weight_schedule: Sets the timing prompt word weight scheduling sequence, used to dynamically adjust the weight ratio of scene features and vehicle features during video generation.

[0095] Furthermore, the closed-loop control mechanism for brand characteristic consistency verification is as follows:

[0096] The semantic segmentation model is used to automatically separate the vehicle foreground and scene background in the generated image, and the vehicle foreground region is extracted. The vehicle foreground region is augmented by multi-angle rotation transformation to generate a set of multi-view feature vectors. The cosine similarity between the set of multi-view feature vectors and the set of brand identity feature parameters is calculated, and the average value is taken as the consistency score. If the consistency score is lower than the preset threshold, a regeneration request is triggered and the constraint strength parameters of the image diffusion generation model are adjusted until the score meets the standard.

[0097] Furthermore, the method further includes:

[0098] A video clip library is established, and historically generated video clips are semantically labeled and vector fingerprints are extracted according to scene type, vehicle brand, and motion features. In new video generation tasks, existing clips that match the current narrative unit are retrieved from the library first. Historical clips that meet the preset matching requirements are directly reused, and the generation process is only performed on unmatched clips. The library is continuously optimized using an incremental learning approach, and newly generated high-quality clips that have passed quality assessment are updated into the library. The vector fingerprint extraction model is dynamically adjusted based on user feedback.

[0099] Furthermore, the logic of the adaptive transition decision-making mechanism is as follows:

[0100] Calculate the structural similarity between the last frame and the first frame of adjacent video segments. If the similarity is higher than the first preset threshold, a direct cut transition is used. If the similarity is between the first threshold and the second threshold, a dissolve transition is used, and the dissolve duration is adaptively calculated based on the difference in motion vectors. If the similarity is lower than the second threshold or the scene switching flag is activated, a sliding transition is used, and the sliding direction is geometrically consistent with the direction of the overall motion vector field of the vehicle.

[0101] Secondly, referring to Figure 2 The illustrated automotive marketing video generation system based on a diffusion model includes the following functional modules:

[0102] The Brand Feature Constraint Source Management Module is used to store and preprocess standardized official photos of target vehicles, and to extract and maintain a set of brand signature feature parameters.

[0103] The intelligent storyboard parsing module is used to break down the marketing script into narrative units and generate scene description text and vehicle posture description text.

[0104] A batch image generation engine that integrates an image diffusion model and a control network to perform batch generation tasks of static scene images;

[0105] The brand consistency closed-loop verification module has a built-in feature matching algorithm and similarity evaluator to perform consistency verification and feedback control on the generated static graph;

[0106] The parametric motion template configurator provides a visual interface for users to edit parameters such as camera movement, vehicle trajectory, environmental effects, and cue word weighting.

[0107] The video diffusion generation engine integrates a video diffusion model and a timing control framework to convert static images and motion templates into video clips.

[0108] The intelligent splicing and post-production compositing module has a built-in adaptive transition decision-maker that performs segment splicing, narration compositing, watermark overlay, and subtitle rendering.

[0109] The media library management and optimization module is used for vector fingerprint storage, similarity retrieval, reuse scheduling, and incremental learning and updating of video clips.

[0110] This system constructs a complete engineering pipeline from brand constraints to finished product output, with each module working together to achieve the automated generation of automotive marketing videos.

[0111] The brand feature constraint source management module, acting as the system input gateway, is responsible for storing and preprocessing standardized official photos of the target vehicle. This module first performs perspective correction and illumination normalization to eliminate interference from shooting angles and lighting conditions. Then, it invokes a deep feature extraction network to perform layered analysis of the vehicle image, progressively aggregating from low-level texture to high-level semantics, precisely identifying iconic elements such as the vehicle body outline, front grille, headlight design, and brand logo, transforming them into a compact set of digital feature parameters. This parameter set, after persistent storage, serves as the brand constraint benchmark for all subsequent generation stages, ensuring that the core features of the vehicle cannot be tampered with.

[0112] The intelligent storyboard parsing module is responsible for structuring and deconstructing marketing creatives. Upon receiving a natural language marketing script, this module automatically identifies scene transitions and emotional shifts based on semantic understanding, dividing the long script into several narrative units. For each unit, the module generates two types of descriptive text: scene description text depicting the environment and visual style, and vehicle posture description text defining the vehicle's spatial position, orientation, and dynamic behavior within the unit. The parsing results are output in a standardized format, providing precise instructions for subsequent generation.

[0113] The batch image generation engine is the core power unit for static image production. This engine integrates an image diffusion model and a control network. After receiving official photos and scene description text, it performs iterative denoising in the latent space. The control network introduces dual control through geometric and depth constraints: geometric constraints extract hard edge information such as vehicle body contours and window folds, while depth constraints analyze the spatial depth relationships of components. These dual constraints and textual prompts are deeply integrated under a cross-attention mechanism, batch-outputting arrays of static images that both adhere to the brand's DNA and match the scene's creative concept.

[0114] The brand consistency closed-loop verification module acts as a quality gatekeeper. For each generated image, this module automatically separates the vehicle foreground and background using a high-precision semantic segmentation model. After eliminating environmental interference, it employs a feature matching algorithm to perform multi-dimensional in-depth analysis of the vehicle region. The contour dimension captures the curvature of the body lines, the texture dimension encodes the grille arrangement pattern, and the structural dimension analyzes the geometric shape of the headlights. These features are compared with a pre-stored brand parameter set for multi-level similarity evaluation. If the evaluation result does not meet the preset standard, a regeneration request is automatically triggered, and the constraint strength of the control network is dynamically enhanced, forming an iterative optimization closed loop.

[0115] The parametric motion template configurator provides a visual human-computer interaction interface, transforming artistic creation experience into structured technical assets. Users can edit four core parameters on the interface: camera movement type (selecting basic camera movement modes such as still, pan, tilt, and zoom); vehicle trajectory parameters (drawing motion path curves on an interactive canvas and configuring speed variation rhythm); environmental dynamic effects (adjusting the intensity of wind, sand, rain, snow, and the gradient of light evolution); and cue word weight scheduling (setting temporal feature intensity changes to achieve a gradual presentation of brand elements from weak to strong). All configurations are stored in a standardized data format, supporting version management and reuse.

[0116] The video diffusion generation engine is responsible for giving static images dynamic life. This engine integrates a video diffusion model and a temporal control framework, taking a validated static image as the first frame and encoding motion templates as control signals describing temporal evolution. In the spatiotemporal joint latent space, the engine performs iterative denoising and introduces a cross-frame attention mechanism, ensuring that the generation of each frame actively references the features of preceding and following frames, constraining the temporal consistency of the vehicle's appearance, smoothing camera movement trajectories, and ensuring that environmental dynamics conform to physical laws. The final output is a video clip that strictly matches the duration and rhythm of the narrative unit.

[0117] The intelligent splicing and post-production compositing module enables seamless segment integration and brand packaging. This module calculates the structural similarity and motion vector continuity of the first and last frames of adjacent segments, adaptively selecting transition methods: direct cuts for high-similarity scenes, dynamic duration dissolves for medium-similarity scenes, and sliding transitions for low-similarity scenes or scene changes. After splicing, the module uses a text-to-speech engine to generate narration, overlays a dynamic brand watermark with varying transparency, renders timecode-aligned subtitles, and outputs a finished video that meets commercial advertising standards.

[0118] The asset library management and optimization module constructs an intelligent management system for generated assets. Newly generated video clips are fed into this module, where a spatiotemporal feature extraction network generates compact vector fingerprints that capture the clip's semantic content, brand attributes, and motion style. The fingerprints and the original clip data are stored together in a vector database, supporting multi-tag-based similarity retrieval. When a new task starts, the module prioritizes retrieving historical clips with high matching degrees for direct reuse; clips that do not match are regenerated and added to the database. The system also continuously optimizes the search ranking through incremental learning, automatically adjusting fingerprint weights based on user feedback to achieve continuous growth in asset value.

[0119] Furthermore, the batch image generation engine further includes:

[0120] The distributed task scheduling unit enables parallel processing and load balancing of multiple tasks based on message queues; the dynamic memory management unit supports large-scale batch generation with limited memory resources through model hierarchical offloading and gradient checkpointing techniques; and the generation quality monitoring unit calculates image quality evaluation indicators in real time and outputs quality reports.

[0121] Furthermore, the video diffusion generation engine further includes:

[0122] The temporal consistency preservation unit constrains inter-frame vehicle feature drift and motion jitter through a cross-frame attention mechanism and an optical flow reprojection loss function.

[0123] A multi-resolution generation pipeline supports multiple output resolutions from standard definition to ultra-high definition;

[0124] The failure retry and parameter adaptation unit automatically adjusts the motion amplitude attenuation factor and triggers regeneration when the quality score of the generated video clip fails to meet the standard. The number of retries is limited to a preset upper limit.

[0125] Furthermore, the system adopts a microservice architecture, with each functional module communicating loosely through standardized API interfaces and asynchronous message queues. It supports containerized orchestration and elastic scaling, enabling the generation of multiple styles of automotive marketing videos within a specified time.

[0126] Implementation Environment Configuration

[0127] This embodiment is deployed on a server cluster equipped with eight NVIDIA A100 graphics processors (80GB of video memory per card), a dual-socket Intel Xeon Platinum 8480+ CPU, 1TB of system memory, and a 10TB all-flash NVMe SSD storage array. The software environment is based on Ubuntu 22.04 LTS operating system, developed using Python 3.10, with core dependencies including the PyTorch 2.1 framework, Diffusers library, Transformers library, and Kubernetes container orchestration platform.

[0128] Input data preparation

[0129] We selected an upcoming mid-size SUV as the target model and obtained three official photos of it: front, side, and rear 45-degree angle. Each photo has a resolution of 4096×2160 pixels and has been reviewed by the OEM's brand department to ensure there are no copyright issues. The marketing script is as follows: "The new vehicle conquers all terrains, from city to wilderness. Scene 1: City skyline in the morning light, the vehicle quietly displays its elegant lines; Scene 2: Driving onto suburban mountain roads, dynamically showcasing the suspension system and handling performance; Scene 3: Desert off-roading, kicking up sand to demonstrate power and freedom."

[0130] Step 1: Extraction of Brand Feature Parameter Set

[0131] Import the three official photos into the brand feature constraint source management module in batches. During the preprocessing stage, perspective correction is performed. The perspective transformation matrix is ​​calculated using vehicle body feature point detection to correct the side view to a standard frontal side view angle, eliminating distortion caused by shooting tilt. Illumination normalization employs a white balance algorithm and histogram equalization to ensure that the three photos taken under different lighting conditions achieve a unified standard in brightness and color.

[0132] In the feature extraction stage, the YOLOv8x model detects the vehicle body with a confidence threshold of 0.95 and a non-maximum suppression threshold of 0.45 to accurately locate the vehicle's bounding box. After cropping the region of interest (ROI) of the vehicle, it is uniformly scaled to 512×512 pixels and input into a ResNet50 network. This network is pre-trained on the ImageNet dataset and followed by a global average pooling layer, outputting a 2048-dimensional feature vector. Principal component analysis retains the first 128 dimensions, with a cumulative contribution rate of 95.3%, forming a brand-identical feature parameter set. This parameter set maintains consistency across the three images, with a cosine similarity greater than 0.92.

[0133] Step Two: Intelligent Storyboard and Template Configuration

[0134] The intelligent storyboard analysis module divides the marketing script into three narrative units. The first unit is 5 seconds long, with the scene description text "Modern city morning, glass curtain walls reflect golden sunlight, vehicles parked in an empty square"; the vehicle posture description text is "Static display, front of the car facing one-thirtieth of the way towards the camera." The second unit is 8 seconds long, with the scene text "Winding mountain road, shaded by green trees on both sides, the road surface is wet with puddles reflecting light"; the posture text is "Driving at a constant speed of 40 kilometers per hour, the camera follows the side and rear of the vehicle." The third unit is 10 seconds long, with the scene text "Magnificent desert, undulating dunes, strong sunlight creating deep shadows"; the posture text is "Accelerating and sprinting up the hill, wheels kicking up dust, the camera zooms in from a low angle."

[0135] The parametric motion template configurator generates a structured configuration for each unit.

[0136] The first unit sets `camera_motion` to `static`, `car_trajectory` to maintain stationary coordinates, and `environment_effect` to configure gradual changes in lighting intensity. The second unit sets `camera_motion` to `pan_right` for pan-right following, `car_trajectory` to define a smooth path from the left side of the screen to the right side, and the speed curve uses an ease-in-out easing function. The third unit sets `camera_motion` to `dolly_in` for dolly-in advancement, `car_trajectory` to define the sprint trajectory from the bottom of the slope to the top, `environment_effect` to set the number and spread of sand particles, and `prompt_weight_schedule` to reduce the brand logo's weight in the first half of the video to highlight the scene, gradually increasing its visibility in the second half.

[0137] Step 3: Constraint Graph Generation

[0138] The batch image generation engine receives official photos and scene description text, and starts the Stable Diffusionv2.1 model. Positive prompts integrate scene and vehicle descriptions, while negative prompts exclude interference such as "blur, distortion, watermark, and text". ControlNet uses a dual-branch parallel processing of Canny edge detection and depth estimation. The Canny branch has a low threshold of 100 and a high threshold of 200, while the depth branch uses the Midas large model.

[0139] The style intensity coefficient α was set to 0.4, the feature consistency preservation coefficient β to 0.8, the CFG scale value to 7.5, the sampling steps to 40, and the Euler a sampler to be selected. Fifteen candidate images were generated per unit, with a resolution of 1024×576. The batch size was set to 4, and xformers memory optimization was used to enable each A100 card to process 4 sets of tasks in parallel. During the generation process, the brand feature parameter set was injected as a conditional embedding into the intermediate layer of the UNet network, and the constraints were strengthened through a cross-attention mechanism.

[0140] Step 4: Brand Consistency Closed-Loop Verification

[0141] The brand consistency closed-loop verification module reviewed each of the 45 candidate images. The SAM segmentation model, using the ViT-Huge backbone network, automatically separated the vehicle foreground from the backgrounds of buildings, mountains, and deserts. The SIFT algorithm extracted 128 feature points from each image, performed an 8-angle rotation transformation augmentation, and calculated the mean cosine similarity with the brand feature parameter set. The threshold θ was set to 0.85. In the first unit of 15 images, 2 images were regenerated due to headlight contour deviation; after adjustment, β was increased to 0.85, and the second generation met the standard. In the second unit, 1 image was regenerated due to grille texture distortion. All images in the third unit passed on the first attempt. Finally, 42 images entered the video conversion stage, with an overall pass rate of 93.3%.

[0142] Step 5: Video diffusion timing generation

[0143] The video diffusion generation engine encodes compliant static images and motion templates into spatiotemporal latent variables. Stable VideoDiffusion v1.1 accepts 512×512 resolution input at a frame rate of 30fps. The first unit generates 150 frames, the second 240 frames, and the third 300 frames. The CrossFrame Attention mechanism sets a 5-frame temporal window, a motion amplitude attenuation factor λ of 0.85, a motion_bucket_id parameter of 127, and a cond_aug value of 0.02.

[0144] The temporal control framework parses motion templates, maps camera motion types to camera parameter matrices, converts vehicle trajectory parameters into Bézier curve control points, encodes environmental dynamic effects as temporal noise perturbations, and uses cue word weight scheduling to generate frame-level feature intensity curves. During generation, cross-frame attention constrains vehicle appearance drift, optical flow reprojection loss suppresses motion jitter, and tile block control ensures stable memory usage within 12GB. Three video clips are output, with Flow Magnusion smoothness scores all greater than 0.75 and CLIP text-image matching scores greater than 0.68.

[0145] Step Six: Adaptive Transition and Composition

[0146] The intelligent splicing and post-production compositing module analyzes the connection between segments. The SSIM structural similarity between the last frame of the first unit and the first frame of the second unit is 0.72, and a 1.5-second dissolve transition is used; the SSIM between the last frame of the second unit and the first frame of the third unit is 0.48, and a left-sliding transition is used, with the direction consistent with the vehicle's exit direction in the second unit. The transition duration is adaptively calculated through motion vector difference to ensure a natural rhythm.

[0147] During the synthesis phase, the narration utilizes Azure Neural TTS speech synthesis, featuring a standard male voice at a rate of 180 words per minute. The first unit's narration is titled "Urban Lights and Shadows, Awaiting Departure," the second "Winding Mountain Roads, Driving with Ease," and the third "Boundless Desert, Conquering Heaven and Earth." The brand watermark is a dynamic logo with a fade-in / fade-out effect in the lower right corner, 30% transparency, lasting for 3 seconds. The subtitles are rendered in ASS format, using Source Han Sans font, size 36, with a white outline, and automatically aligned with the audio track timecode.

[0148] Step 7: Incremental Learning of the Resource Library

[0149] The three generated video clips were sent to the media library management module. The VideoMAE model (ViT-B backbone, 16-frame input window) extracted a 1024-dimensional spatiotemporal vector fingerprint. The FAISS index used the IndexIVFFlat type, with 1000 inverted lists. The clips were labeled with multiple tags: brand "X-SUV" and scene category.

[0150] "urban / mountain / desert", movement type "static / dynamic / dynamic", quality rating 4.5 / 5.0.

[0151] When generating new city night scene videos of the same car model for subsequent tasks, the media library found a segment with a matching degree of 0.82 for the first unit. This segment was directly reused, and only the second and third units were generated. The overall time was reduced from 23 minutes to 9 minutes, an efficiency improvement of 61%. User click playback data was fed back to the Q-learning optimizer with a learning rate of 0.01 and a reward function R = completion rate × 0.7 + click rate × 0.3. The retrieval weights were dynamically adjusted, increasing the reuse accuracy from the initial 58% to 76%.

[0152] Effect verification

[0153] The output marketing video is 23 seconds long, with a resolution of 1920×1080, H.264 encoding, and a bitrate of 8Mbps, meeting the upload standards of mainstream social media platforms. After review by the brand department, the similarity between the vehicle's front grille, headlight outline, and logo and the official photos is greater than 0.9, with no feature drift. The marketing team reported that the total time from script submission to final product delivery was 47 minutes, a significant improvement in efficiency compared to traditional live-action shooting (3 days) and purely AI-generated content (2.5 hours). Furthermore, multiple versions can be generated in parallel, with a daily production capacity of over 15 videos, meeting the intensive promotional needs during the new car launch period.

[0154] By adopting the above-disclosed technical solution of this invention, the following beneficial effects are obtained:

[0155] 1. Brand safety fundamentally improved

[0156] By standardizing the extraction of brand feature parameters and implementing a closed-loop verification mechanism, consistency control is upgraded from manual sampling to algorithmic hard constraints. Throughout the generation process, the vehicle brand's iconic elements consistently maintain a high degree of consistency with official photos, effectively mitigating feature drift risks and meeting the stringent compliance requirements of automakers. This provides a safety guarantee for the commercial application of generative AI in the automotive marketing field.

[0157] 2. Revolutionary improvements in motion control precision and reusability

[0158] It innovatively deconstructs the cinematic language standards of the film and television industry into structured JSON templates, enabling precise quantitative configuration of camera movement, vehicle trajectories, environmental effects, and temporal weights. Creative experience can be transformed into programmable and version-manageable technical assets, allowing non-professionals to reuse professional-grade camera movement solutions in batches, significantly reducing the creative threshold and reliance on personnel.

[0159] 3. Significantly optimized production efficiency and cost structure.

[0160] By constructing a spatiotemporal semantic fingerprint and vector retrieval system for video clips, historically generated content can be intelligently reused. When creating secondary works for similar models, existing assets are prioritized to avoid repetitive reasoning and calculations, significantly shortening the production cycle. Combined with the distributed parallel capabilities of microservices, multiple versions can be batch-output within commercial time constraints, achieving economies of scale with continuously decreasing marginal costs.

[0161] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating automotive marketing videos based on a diffusion model, characterized in that, Includes the following steps: Obtain standardized official photos of the target vehicle, and use a deep feature extraction network to analyze the vehicle brand's iconic feature parameter set, which includes the vehicle body outline topology vector, front grille texture features, headlight outline features, and brand logo structure features. Based on the marketing script, the video is divided into several narrative units, and scene description text and vehicle posture description text are generated for each narrative unit. A parameterized motion template is constructed that includes camera movement type, vehicle trajectory parameters, environmental dynamic effects and cue word weight scheduling. The official photos and scene description text are input into the image diffusion generation model. The model is then used to introduce dual control of geometric and depth constraints through the ControlNet architecture to generate static scene images with consistent brand characteristics in batches. The generated static scene image is verified for brand feature consistency. The feature matching algorithm is used to extract the vehicle area feature vector in the generated image and perform a multi-dimensional similarity evaluation with the brand's iconic feature parameter set. If the evaluation result does not meet the preset consistency standard, the image is regenerated, forming a generation-verification closed-loop control. The verified static scene image and the parameterized motion template are jointly encoded into spatiotemporal latent variables that can be parsed by the video diffusion model. These variables are then input into the video diffusion generation model, and temporal consistency is constrained through a cross-frame attention mechanism to generate video segments corresponding to the narrative units. An adaptive transition decision mechanism is used to intelligently stitch video clips together. The transition method is automatically selected based on the scene structure similarity and motion vector continuity of adjacent clips. A preliminary marketing video is synthesized and standardized brand elements are added to output the final automotive marketing video.

2. The method according to claim 1, characterized in that, The extraction steps for the brand's iconic feature parameter set specifically include: Perspective correction and illumination normalization are performed on official photos to eliminate interference from shooting angle and lighting; a target detection model is used to accurately locate the main body area of ​​the vehicle and extract the region of interest (ROI) image; deep semantic features of the ROI image are extracted through a residual neural network to generate a high-dimensional feature representation; principal component analysis is performed on the high-dimensional features to reduce dimensionality, retaining the core dimensions whose feature contribution rate meets the preset requirements, and constructing a set of brand-identical feature parameters.

3. The method according to claim 1, characterized in that, The parameterized motion template is described using a structured data format and includes the following core fields: camera_motion: Defines the camera motion mode, including basic camera movement types such as stationary, horizontal pan, vertical tilt, zoom in and zoom out; car_trajectory: Describes the vehicle's trajectory parameters, including the starting spatial coordinates, ending spatial coordinates, velocity curve, and path smoothness. environment_effect: Configures dynamic environmental effects, including wind, sand, rain and snow intensity, light change gradient, and particle effect density; prompt_weight_schedule: Sets the timing prompt word weight scheduling sequence, used to dynamically adjust the weight ratio of scene features and vehicle features during video generation.

4. The method according to claim 1, characterized in that, The closed-loop control mechanism for brand feature consistency verification is as follows: The semantic segmentation model is used to automatically separate the vehicle foreground and scene background in the generated image, and the vehicle foreground region is extracted. The vehicle foreground region is augmented by multi-angle rotation transformation to generate a set of multi-view feature vectors. The cosine similarity between the set of multi-view feature vectors and the set of brand identity feature parameters is calculated, and the average value is taken as the consistency score. If the consistency score is lower than the preset threshold, a regeneration request is triggered and the constraint strength parameters of the image diffusion generation model are adjusted until the score meets the standard.

5. The method according to claim 1, characterized in that, The method further includes: A video clip library is established, and historically generated video clips are semantically labeled and vector fingerprints are extracted according to scene type, vehicle brand, and motion features. In new video generation tasks, existing clips that match the current narrative unit are retrieved from the library first. Historical clips that meet the preset matching requirements are directly reused, and the generation process is only performed on unmatched clips. The library is continuously optimized using an incremental learning approach, and newly generated high-quality clips that have passed quality assessment are updated into the library. The vector fingerprint extraction model is dynamically adjusted based on user feedback.

6. The method according to claim 1, characterized in that, The logic of the adaptive transition decision mechanism is as follows: Calculate the structural similarity between the last frame and the first frame of adjacent video segments. If the similarity is higher than the first preset threshold, a direct cut transition is used. If the similarity is between the first threshold and the second threshold, a dissolve transition is used, and the dissolve duration is adaptively calculated based on the difference in motion vectors. If the similarity is lower than the second threshold or the scene switching flag is activated, a sliding transition is used, and the sliding direction is geometrically consistent with the direction of the overall motion vector field of the vehicle.

7. A car marketing video generation system based on a diffusion model, characterized in that, Includes the following functional modules: The Brand Feature Constraint Source Management Module is used to store and preprocess standardized official photos of target vehicles, and to extract and maintain a set of brand signature feature parameters. The intelligent storyboard parsing module is used to break down the marketing script into narrative units and generate scene description text and vehicle posture description text. A batch image generation engine that integrates an image diffusion model and a control network to perform batch generation tasks of static scene images; The brand consistency closed-loop verification module has a built-in feature matching algorithm and similarity evaluator to perform consistency verification and feedback control on the generated static graph; The parametric motion template configurator provides a visual interface for users to edit parameters such as camera movement, vehicle trajectory, environmental effects, and cue word weighting. The video diffusion generation engine integrates a video diffusion model and a timing control framework to convert static images and motion templates into video clips. The intelligent splicing and post-production compositing module has a built-in adaptive transition decision-maker that performs segment splicing, narration compositing, watermark overlay, and subtitle rendering. The media library management and optimization module is used for vector fingerprint storage, similarity retrieval, reuse scheduling, and incremental learning and updating of video clips.

8. The system according to claim 7, characterized in that, The batch image generation engine further includes: The distributed task scheduling unit enables parallel processing and load balancing of multiple tasks based on message queues; the dynamic memory management unit supports large-scale batch generation with limited memory resources through model hierarchical offloading and gradient checkpointing techniques; and the generation quality monitoring unit calculates image quality evaluation indicators in real time and outputs quality reports.

9. The system according to claim 7, characterized in that, The video diffusion generation engine further includes: The temporal consistency preservation unit constrains inter-frame vehicle feature drift and motion jitter through a cross-frame attention mechanism and an optical flow reprojection loss function. A multi-resolution generation pipeline supports multiple output resolutions from standard definition to ultra-high definition; The failure retry and parameter adaptation unit automatically adjusts the motion amplitude attenuation factor and triggers regeneration when the quality score of the generated video clip fails to meet the standard. The number of retries is limited to a preset upper limit.

10. The system according to claim 7, characterized in that, The system is deployed using a microservice architecture, with each functional module communicating loosely through standardized API interfaces and asynchronous message queues. It supports containerized orchestration and elastic scaling, and can complete the task of generating multiple styles of automotive marketing videos within a specified time.

Images