Video tail frame transition method and device, equipment and storage medium

Abstract

The embodiment of the invention discloses a video tail frame transition method and device, equipment and a storage medium, and the method comprises the steps: obtaining a to-be-spliced tail frame image and a to-be-spliced first frame image, and generating first prompt words corresponding to the tail frame image and the first frame image; performing style construction processing according to a main body vector obtained by performing main body element analysis on the first cue word to generate a first target image and a second target image of corresponding styles; generating a second cue word according to feature data obtained by performing feature extraction on the first target image and the second target image and the determined associated transition parameters of the feature data; and generating an intermediate video frame according to the second prompt word, taking the tail frame image as a starting frame, taking the first frame image as a tail frame, and integrating the starting frame, the intermediate video frame and the tail frame to generate a target transition video. According to the scheme, the problem of low efficiency of video tail frame transition can be solved, the target transition video is automatically generated through the head frame image and the tail frame image, and the efficiency of video tail frame transition is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of video processing technology, and in particular to a video end-frame transition method, apparatus, device, and storage medium. Background Technology

[0002] In video production and animation, end-frame transitions are a crucial technical element connecting preceding and following scenes or actions, and their importance extends across multiple fields, including film and television creation, animation production, and short video content creation. High-quality end-frame transitions can effectively enhance the visual smoothness and artistic appeal of videos by smoothly connecting different shots, and can even convey narrative logic through transition design, becoming one of the core elements for improving the overall quality of a work. Traditional end-frame transition techniques have long relied on manual design and adjustment, resulting in a work mode highly dependent on the designer's subjective creativity and operational skills. In practice, designers need to manually design transition effects for different video content. This includes configuring parameters for basic transition effects such as image dissolve, wipe, and zoom, as well as adjusting complex effects like character deformation and scene reconstruction frame by frame—all done manually. This work mode not only requires designers to have solid artistic skills and rich video editing experience, but also a deep understanding of the precise operation of software tools, leading to a strong dependence on specialized talent in the implementation of end-frame transitions. The entire process from creative conception to implementation consumes a significant amount of time and manpower, especially since the production of complex transition effects often requires hours or even days of repeated debugging, making it difficult to meet the time-sensitive demands of large-scale content production in the video era. Therefore, the efficiency of traditional end-frame transition methods is relatively low. Summary of the Invention This application provides a video end-frame transition method, apparatus, device, and storage medium, which can solve the technical problem of low efficiency in video end-frame transition. By automatically generating the corresponding target transition video through the first and last frame images, the efficiency of video end-frame transition is significantly improved.

[0003] In a first aspect, embodiments of this application provide a video end-frame transition method, comprising: Obtain the last frame image and the first frame image to be stitched together, and generate the first prompt word corresponding to each of the last frame image and the first frame image; The main element is parsed from the first prompt word to obtain the main vector, and style construction is performed based on the main vector to generate the first target image and the second target image with the corresponding style; Feature data is obtained by extracting features from the first target image and the second target image, and a second prompt word is generated based on the feature data and the correlation transition parameters of the determined feature data. The intermediate video frame is generated based on the second prompt word. The starting frame is the last frame, and the ending frame is the first frame. The starting frame, intermediate video frame and ending frame are integrated to generate the target transition video.

[0004] Furthermore, the last frame image and the first frame image to be stitched are obtained, and the first prompt words corresponding to the last frame image and the first frame image are generated respectively, including: The last frame image of the first video to be stitched is obtained, and feature analysis is performed on the last frame image to obtain the last frame features. The last frame features include one or more of the following: first subject features, first scene features, first action features, first tone features, and first style features. The first frame image of the second video to be stitched is obtained, and feature analysis is performed on the first frame image to obtain the first frame features. The first frame features include one or more of the following: second subject features, second scene features, second action features, second tone features, and second style features. Generate the first prompt word corresponding to the last frame image based on the features of the last frame, and generate the first prompt word corresponding to the first frame image based on the features of the first frame.

[0005] Furthermore, the main element is parsed from the first prompt word to obtain the main vector, including: The first subject vector is obtained by parsing the subject element of the first prompt word corresponding to the last frame image, and the second subject vector is obtained by parsing the subject element of the first prompt word corresponding to the first frame image. Accordingly, style construction processing is performed based on the subject vector to generate a first target image and a second target image with corresponding styles, including: The first style feature is obtained by extracting style keywords from the first prompt word corresponding to the last frame image, and the second style feature is obtained by extracting style keywords from the first prompt word corresponding to the first frame image. A first target image with a first style is generated by style construction processing based on a first subject vector and a first style feature, and a second target image with a second style is generated by style construction processing based on a second subject vector and a second style feature.

[0006] Furthermore, feature data is obtained by extracting features from the first target image and the second target image, including: The first feature is obtained by extracting features from the first target image, wherein the first feature includes one or more of the following: third subject feature, third scene feature, third action feature, third tone feature, and third style feature; The second target image is subjected to feature extraction to obtain the second feature, wherein the second feature includes one or more of the following: fourth subject feature, fourth scene feature, fourth action feature, fourth tone feature, and fourth style feature; Accordingly, a second prompt word is generated based on the feature data and the determined association transition parameters of the feature data, including: A correlation analysis was performed on the first feature and the second feature to obtain the correlation transition parameters; A second prompt word is generated based on the first feature, the second feature, and the associated transition parameters.

[0007] Furthermore, correlation analysis is performed based on the first and second features to obtain correlation transition parameters, including: Correlation analysis is performed on the third subject feature and the fourth subject feature to obtain subject transition parameters; correlation analysis is performed on the third scene feature and the fourth scene feature to obtain scene transition parameters. Correlation analysis is performed on the third and fourth action features to obtain action transition parameters, and correlation analysis is performed on the third and fourth hue features to obtain hue transition parameters. A correlation analysis was performed on the third and fourth style features to obtain the style transition parameter.

[0008] Furthermore, based on the first feature, the second feature, and the associated transition parameters, a second prompt word is generated, including: Based on the third subject feature, the fourth subject feature, and the subject transition parameter, generate subject change prompt words; based on the third scene feature, the fourth scene feature, and the scene transition parameter, generate scene change prompt words. Based on the third action feature, the third action feature and the action transition parameter, generate action change prompt words; based on the third tone feature, the fourth tone feature and the tone transition parameter, generate tone change prompt words. Based on the third style feature, the fourth style feature, and the style transition parameter, a style fusion prompt word is generated. The subject change prompt word, scene change prompt word, action change prompt word, color tone change prompt word, and style fusion prompt word are integrated to obtain the second prompt word.

[0009] Furthermore, the target transition video is generated by integrating the start frame, intermediate video frames, and end frame, including: Get the set video sound effects and extract the beat points and emotional features of the video sound effects; The target video frame is obtained by integrating the start frame, middle video frames, and end frame. Identify the visual keyframes in the target video frame, bind the beat points and visual keyframes, and generate the first target transition video; The hue and saturation of the first target transition video are adjusted based on emotional characteristics to obtain the second target transition video.

[0010] In a second aspect, embodiments of this application provide a video end-frame transition device, comprising: The image acquisition module is used to acquire the last frame image and the first frame image to be stitched together; The first prompt word generation module is used to generate the first prompt words corresponding to the last frame image and the first frame image, respectively. The main vector generation module is used to parse the main elements of the first prompt word to obtain the main vector; The target image generation module is used to perform style construction processing based on the subject vector to generate a first target image and a second target image with corresponding styles. The feature data extraction module is used to extract features from the first target image and the second target image to obtain feature data; The second prompt word generation module is used to generate second prompt words based on feature data and the associated transition parameters of the determined feature data; The intermediate video frame generation module is used to generate intermediate video frames based on the second prompt word; The video generation module is used to generate a target transition video by integrating the start frame, intermediate video frames, and end frame, with the last frame image as the start frame and the first frame image as the end frame.

[0011] In a third aspect, embodiments of this application provide a video end-frame transition device, comprising: Memory and one or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement the video end frame transition method as described in the first aspect.

[0012] In a fourth aspect, embodiments of this application provide a storage medium for storing computer-executable instructions, which, when executed by a computer processor, are used to perform a video end-frame transition method as described in the first aspect.

[0013] This application embodiment generates first prompt words for each of the last and first frames during the video end-frame transition, based on the last and first frames to be stitched. Subject element parsing is performed on the first prompt words to obtain subject vectors, and style construction processing is performed on the subject vectors to generate a first and second target image with corresponding styles. Feature extraction is performed on the first and second target images to obtain feature data, and a second prompt word is generated based on the feature data and the determined associated transition parameters. An intermediate video frame is generated based on the second prompt word. Using the last frame as the starting frame and the first frame as the ending frame, the starting frame, intermediate video frames, and ending frame are integrated to generate the target transition video. By employing the above technical means, the target transition video can be automatically generated from the last and first frames to be stitched, thus avoiding the technical problem of low efficiency in video end-frame transitions. It eliminates the need for manual design, improves the automation level of video end-frame transitions, and significantly enhances the efficiency of video end-frame transitions.

[0014] The beneficial effects of the video end-frame transition device, video end-frame transition equipment, and storage medium provided above can be referenced in relation to the beneficial effects of the video end-frame transition method. Attached Figure Description

[0015] Figure 1 This is a flowchart of a video end-frame transition method provided in an embodiment of this application; Figure 2 This is a flowchart of another video end-frame transition method provided in the embodiments of this application; Figure 3 This is a flowchart illustrating another video end-frame transition method provided in the embodiments of this application; Figure 4 This is a flowchart illustrating another video end-frame transition method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a video end-frame transition device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a video end-frame transition device provided in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0017] Traditional end-frame transition techniques have long relied on manual design and adjustment, resulting in a work mode highly dependent on the designer's subjective creativity and operational skills. In practice, designers need to manually design transition effects for different video content. This includes configuring parameters for basic transition effects such as image dissolve, wipe, and zoom, as well as adjusting complex effects like character deformation and scene reconstruction frame by frame—all done manually. This work mode not only requires designers to have solid artistic skills and rich video editing experience, but also a deep understanding of the precise operation of software tools, leading to a strong dependence on specialized talent in the implementation of end-frame transitions. The entire process from creative conception to implementation consumes a significant amount of time and manpower, especially since the production of complex transition effects often requires hours or even days of repeated debugging, making it difficult to meet the time-sensitive demands of large-scale content production in the video era. Therefore, the efficiency of traditional end-frame transition methods is relatively low.

[0018] Based on this, this application provides a video end-frame transition method, apparatus, device, and storage medium. The aim is to generate first prompt words corresponding to the end-frame image and the first frame image respectively during video end-frame transition, based on the end-frame image to be stitched and the first frame image; to obtain a subject vector by parsing the subject element based on the first prompt words; to generate a first target image and a second target image with corresponding styles based on the subject vectors; to obtain feature data by feature extraction based on the first target image and the second target image; to generate a second prompt word based on the feature data and the determined associated transition parameters of the feature data; to generate an intermediate video frame based on the second prompt word; and to integrate the start frame, intermediate video frame, and end frame with the end-frame image as the start frame and the first frame image as the end frame to generate a target transition video. Using the above technical means, a target transition video can be automatically generated from the end-frame image to be stitched and the first frame image. Compared with the traditional manually designed and adjusted end-frame transition method, this embodiment eliminates the need for manual design, improving the automation level of video end-frame transition and thus significantly improving the efficiency of video end-frame transition.

[0019] Figure 1 A flowchart of a video end-frame transition method provided in this application embodiment is given. The video end-frame transition method provided in this embodiment can be executed by a video end-frame transition device, which can be implemented by software and / or hardware. The video end-frame transition device can be composed of two or more physical entities, or it can be composed of a single physical entity. Generally, the video end-frame transition device can be a computer device.

[0020] The following description uses a computer device as the main entity executing the video end-frame transition method. (Refer to...) Figure 1 The video end-frame transition method specifically includes: S11. Obtain the tail frame image and the first frame image to be stitched together, and generate the first prompt word corresponding to each of the tail frame image and the first frame image.

[0021] The process involves acquiring the last frame image of the first video to be stitched together and the first frame image of the second video to be stitched together. Image recognition is performed on the last frame image, and a corresponding first prompt word is generated based on the recognized content. Similarly, image recognition is performed on the first frame image, and a corresponding first prompt word is generated based on the recognized content. Here, the first video is the preceding video to be stitched together, and the second video is the following video to be stitched together. After the last frame image of the first video and the first frame image are stitched together to create a transition, the first and second videos can be stitched together into a complete video. For example, suppose the recognized last frame image of the first video is an image of a "fox," and the first frame image of the second video is an anthropomorphic image of a "fox spirit." A corresponding first prompt word can be generated based on the content recognized in the last frame image, such as "cunning fox," and a corresponding first prompt word can be generated based on the content recognized in the first frame image, such as "mysterious and beautiful fox spirit."

[0022] In one embodiment, the obtained first cue word may also include descriptions of multiple dimensions such as subject, scene, action, tone and style. For example, the first cue word corresponding to the last frame image may be "Subject: Red fox, with golden-red fur and motion blur, and round amber pupils; Scene: Autumn dusk in a forest with fallen leaves, with slanted tree trunks and golden light spots in the background, and the sky is orange-red gradient, with the subject located slightly to the left of the center of the image; Action: Dynamic running; Style: Realism; Tone: Strong light and shadow contrast, with depth of field focused on the fox's head, ISO (sensitivity) of 400, and f (aperture value) of 2.8." The first prompt corresponding to the first frame image can be: "Main subject: fox spirit, silver-white dress with ink-wash pattern, nine fluffy tails with moonlight luster, holding a bamboo talisman in the right hand; Scene: Japanese shrine in the rain, red torii gate wet with rain, stone lanterns emitting warm light, blue foxfire in the background; Action: standing; Style: Eastern fantasy; Color tone: cinematic CG rendering, soft lighting, cool blue color tone, dynamic blur coefficient 0.3."

[0023] As described above, generating corresponding first prompt words based on the last frame image to be stitched and the first frame image can convert the images into text, providing reference prompt words for subsequent video last frame transitions, and improving the matching degree between the target transition video and the last frame image to be stitched and the first frame image in subsequent video last frame transitions generated based on the first prompt words.

[0024] In one embodiment, when acquiring the tail frame image and the first frame image to be stitched, a multi-frame sampling strategy can be adopted. This involves sampling multiple frames (e.g., the last 10 frames) of the tail segment of the first video to obtain the clearest frame as the tail frame image, and sampling multiple frames (e.g., the first 10 frames) of the head segment of the second video to obtain the clearest frame as the first frame image. This multi-frame sampling strategy avoids single-frame motion blur or noise interference, thereby improving the accuracy of generating the target transition video corresponding to the tail frame transition in subsequent videos.

[0025] In one embodiment, the acquired tail frame image and first frame image are subjected to size normalization processing to unify the two images to the same size. For example, the resolution of the tail frame image and the first frame image can be uniformly adjusted to 1920×1080, which can be adjusted by interpolation to avoid jagged edges. The normalized tail frame image and the first frame image are then subjected to noise reduction processing to remove compression artifacts and preserve details such as hair and texture.

[0026] In one embodiment, after generating the first prompt word, user-inputted modification information can be received, and the first prompt word can be modified according to the modification information to meet the user's needs. For example, "holding a bamboo talisman in the right hand" in the first prompt word corresponding to the last frame image can be modified to "holding a bamboo talisman in the left hand"; or "blue foxfire in the background" in the first prompt word corresponding to the first frame image can be modified to "red foxfire in the background". As described above, after generating the first prompt word, the first prompt words corresponding to the last frame image and the first frame image are modified by receiving user-input modification information to meet the user's creative needs, providing personalized design for the video's end frame transition, thereby improving the user experience.

[0027] S12. Based on the first prompt word, the main element is parsed to obtain the main vector, and the style is constructed based on the main vector to generate the first target image and the second target image with the corresponding style.

[0028] The first subject vector is obtained by parsing the first cue word corresponding to the last frame image into subject elements, and the second subject vector is obtained by parsing the first cue word corresponding to the first frame image into subject elements. The first subject vector includes subject elements, scene elements, and action elements, for example, first subject vector = [subject, scene, action]. The second subject vector also includes subject elements, scene elements, and action elements, for example, second subject vector = [subject, scene, action]. Style keywords are extracted from the first cue word corresponding to the last frame image to obtain a first style feature, for example, "realism"; style keywords are also extracted from the first cue word corresponding to the first frame image to obtain a second style feature, for example, "Eastern fantasy". For example, corresponding style feature parameters can be determined based on the first and second style features, including color space, texture mode, and lighting model. For example, the first style feature is "realism," with a corresponding color space of "sRGB color gamut, high saturation (S=80%), warm tone tendency," a corresponding texture mode of "hair noise σ=2.5, PBR material reflectivity 0.3," and a corresponding lighting model of "physical lighting (physically based shadow calculation)." The second style feature is "Oriental fantasy," with a corresponding color space of "CMYK color gamut, low saturation (S=30%), cool tone tendency," a corresponding texture mode of "silk texture σ=0.3, ink wash radius 5px," and a corresponding lighting model of "soft light effect (Gaussian blur radius 8px)." Style construction processing is performed based on the first subject vector and the first style feature to generate the first target image corresponding to the first style. Style construction processing is then performed based on the second subject vector and the second style feature to generate the second target image corresponding to the second style. As described above, by generating corresponding target images through cross-style mapping of subject vectors, the core features of the subject in images of different styles are consistent, avoiding the technical problem of "subject deformation" in style transfer, improving style transfer capability, ensuring subject consistency between the first target image after style transfer and the last frame image before transfer, and ensuring subject consistency between the second target image after style transfer and the first frame image before transfer, thereby improving the accuracy and visual appeal of the subsequently generated last frame transition video.

[0029] As described above, the main element is parsed from the first prompt word to obtain the main vector, and the style is constructed based on the main vector to generate the first target image and the second target image of the corresponding style, thereby realizing the automatic generation of style transfer images. Compared with the traditional manual drawing method, this greatly improves the efficiency of dual-style image generation, and thus improves the overall efficiency of the end frame transition.

[0030] S13. Feature data is obtained by extracting features from the first target image and the second target image, and a second prompt word is generated based on the feature data and the associated transition parameters of the determined feature data.

[0031] Feature extraction is performed on the first and second target images to obtain their respective feature data. This feature data includes one or more of the following: subject features, scene features, action features, tone features, and style features. The feature data corresponding to the first and second target images are compared to determine the correlation transition parameters between them. These correlation transition parameters are a series of structured parameters used to quantify and control the feature change logic, time allocation, and visual effects from the original image (e.g., the first target image) to the target image (e.g., the second target image) during video frame transition generation. Their core characteristic is dynamic generation through cross-frame feature association (such as differences in subject shape, scene, tone, etc.), rather than fixed values, ensuring that the transition effect is consistent with the semantics of the image content. Correlation transition parameters include subject correlation transition parameters, scene correlation transition parameters, action transition parameters, tone transition parameters, and style transition parameters. Based on the feature data corresponding to the first and second target images, and the correlation transition parameters corresponding to both types of feature data, a second prompt word is generated. For example, the generated second prompt is: "A realistic fox melancholically transforms into a fantastical fox spirit in a warm yellow forest. Its fur melts into flowing silk, and dust gradually turns into luminous particles that surround the moonlit bamboo forest. The color tone changes from warm yellow to cool blue, blending realistic textures with soft light effects. The whole process lasts 3.5 seconds, accompanied by the gradual emergence of the shrine torii gate from the shadows of the trees."

[0032] As described above, by performing difference analysis and causal reasoning on the feature data corresponding to the first target image and the feature data corresponding to the second target image, a logically consistent transitional description is generated, i.e., a second prompt word is obtained. This improves the narrative coherence of the video's end-frame transition, thereby enhancing the smoothness and visual appeal of the target transition video generated for subsequent video end-frame transitions. Furthermore, the entire process from image features to prompt words is automated, thus improving the overall efficiency of the video end-frame transition.

[0033] In one embodiment, after generating the second prompt word, modification information input by the user can be received, and the second prompt word can be modified according to the modification information to meet the user's needs. For example, "melancholy transforming into a fantastic fox demon" in the second prompt word can be modified to "joyfully transforming into a fantastic fox demon." As described above, after generating the second prompt word, modifying it by receiving modification information input by the user satisfies the user's creative needs, provides personalized design for the video's end frame transition, and thus improves the user experience.

[0034] S14. Generate intermediate video frames based on the second prompt word, take the last frame image as the starting frame and the first frame image as the ending frame, and integrate the starting frame, intermediate video frames and ending frame to generate the target transition video.

[0035] The generated second prompt word is structurally decomposed into executable video generation parameters. These parameters are then time-allocated according to the set duration, and the corresponding intermediate video frames are generated based on the allocated time. Starting with the last frame and ending with the first frame, the target transition video is generated by stitching the first, intermediate, and last frames together in that order. This target transition video begins with the last frame of the first video, passes through the intermediate frames, and ends with the first frame of the second video. The transition between the first and second videos is achieved based on this target transition video.

[0036] As described above, during the transition of the video's end frame, based on the end frame image to be stitched and the first frame image, a first prompt word is generated for each of the end frame image and the first frame image. Subject element parsing is performed based on the first prompt word to obtain a subject vector. Style construction processing is then performed based on the subject vector to generate a first target image and a second target image with corresponding styles. Feature extraction is performed on the first and second target images to obtain feature data. A second prompt word is generated based on the feature data and the determined correlation transition parameters of the feature data. An intermediate video frame is generated based on the second prompt word. Using the end frame image as the starting frame and the first frame image as the ending frame, the starting frame, intermediate video frames, and ending frame are integrated to generate the target transition video. Using the above technical means, the target transition video can be automatically generated from the end frame image to be stitched and the first frame image. Compared to the traditional manual design and adjustment of end frame transitions, this embodiment eliminates the need for manual design, improving the automation level of the video end frame transition and thus significantly improving the efficiency of the video end frame transition.

[0037] Figure 2 This is a flowchart of another video end-frame transition method provided in the embodiments of this application, referred to... Figure 2 The video end-frame transition method specifically includes: S111. Obtain the last frame image of the first video to be stitched, perform feature analysis on the last frame image to obtain the last frame features, wherein the last frame features include one or more of the following: first subject features, first scene features, first action features, first tone features, and first style features.

[0038] The process begins by acquiring the last frame image of the first video to be stitched together. Image recognition and feature analysis are then performed on the last frame image to obtain its features. These features include one or more of the following: a first subject feature, a first scene feature, a first action feature, a first tone feature, and a first style feature. For example, subject feature analysis is performed on the last frame image to obtain the first subject feature. This can be achieved by instance segmentation of the last frame image, locating the subject's outline and outputting mask coordinates, mapping the subject's visual features to a semantic space, and generating corresponding subject feature labels. For instance, the first subject feature might be "red fox, golden-red fur with motion blur, and round amber pupils." Scene feature analysis is then performed on the last frame image to obtain the first scene feature. This can be achieved by image recognition of the last frame image to determine the scene type, such as a forest. Key objects in the identified scene are detected, and an object element table is generated, which may include elements such as sunset, fallen leaves, tree trunks, light spots, and the sky. Finally, a depth map of the identified scene is obtained, and the spatial position of the subject relative to its environment is analyzed. Based on the aforementioned scene type, object element table, and spatial position of the subject and environment, the first scene feature is obtained. For example, the first scene feature is "autumn dusk in a forest with fallen leaves, a background with slanted tree trunks and golden light spots, a sky with an orange-red gradient, and the subject located slightly to the left of the center of the frame." Motion feature analysis is performed on the last frame image to obtain the first motion feature. For example, the motion vector of the subject in the last frame image is calculated, where the motion vector includes direction and velocity. The joint coordinates of the subject are extracted, for example, the joint angles of the limbs of a fox running, such as forelimb θ=120° and hindlimb θ=90°. Motion analysis is performed on the last 10 consecutive frames of the first video to determine the motion type (e.g., running or standing). Based on the aforementioned motion vector, joint coordinates, and motion type, the first motion feature is obtained. For example, the first motion feature is "dynamic running." Tone feature analysis is performed on the last frame image to obtain the first tone feature. For example, the RGB three-channel distribution of the last frame image is calculated, for example, the RGB three-channel distribution of a warm yellow forest is R=200±10, G=180±15, B=100±5. The RGB three-channel distribution is converted to HSV space. In HSV space, the dominant hue and saturation are extracted; for example, the dominant hue is H=30°±5° (warm yellow), and the saturation is S=80%±5%. Parameters such as light intensity, depth of field, ISO, and aperture value are identified in the last frame image. Based on the obtained dominant hue, saturation, light intensity, depth of field, ISO, and aperture value, the first hue feature is obtained. For example, the first hue feature is "strong light and shadow contrast, depth of field focused on the fox's head, ISO (sensitivity) 400, f (aperture value) 2.8". Style feature analysis is performed on the last frame image to obtain the first style feature. For example, text-image alignment processing is performed on the obtained first subject feature, first scene feature, first action feature, and first hue feature, and style keywords are matched to obtain the corresponding first style feature.For example, if the style keyword is "realism", then the primary style characteristic is "realism".

[0039] As described above, feature analysis of the end-frame image identifies subject features, scene features, action features, tone features, and style features, enabling the image to be transcribed into text. This provides a semantic basis for generating the target transition video corresponding to the subsequent end-frame transition. The fully automated generation of end-frame features significantly improves processing efficiency compared to manual annotation, thereby enhancing the overall efficiency of the video end-frame transition.

[0040] S112. Obtain the first frame image of the second video to be stitched, perform feature analysis on the first frame image to obtain the first frame features, wherein the first frame features include one or more of the following: second subject, second scene, second action, second tone, and second style.

[0041] The process begins by acquiring the first frame of the second video to be stitched together. Image recognition and feature analysis are then performed on the first frame to obtain its features. These features include one or more of the following: second subject features, second scene features, second action features, second tone features, and second style features. For example, subject feature analysis is performed on the first frame to obtain the second subject features. This can be achieved by segmenting the first frame, locating the subject's outline, outputting mask coordinates, mapping the subject's visual features to a semantic space, and generating corresponding subject feature labels. For instance, the second subject feature might be described as "a fox spirit, wearing a silver-white dress with ink-wash patterns, nine fluffy tails with a moonlit sheen, holding a bamboo talisman in her right hand." Scene feature analysis is then performed on the first frame to obtain the second scene features. This can be achieved by performing image recognition on the first frame to determine the scene type, such as a shrine. Key objects in the identified scene are detected, generating an object element table, which may include elements such as a shrine, birds, rain, stone lanterns, and foxfire. Finally, a depth map of the identified scene is obtained, and the spatial position of the subject relative to its environment is analyzed. Based on the aforementioned scene type, object element table, and spatial position of the subject and environment, a second scene feature is obtained. For example, the second scene feature is "a Japanese shrine in the rain, with red torii gates wet from the rain, stone lanterns emitting warm light, and blue foxfire in the background." Motion feature analysis is performed on the first frame image to obtain the second motion feature. For example, the motion vector of the subject in the first frame image is calculated, where the motion vector includes direction and velocity. The joint coordinates of the subject are extracted; for example, the angle θ of the right hand holding the talisman is 120°, and the center of gravity offset of the standing posture is ≤5 pixels. Motion analysis is performed on the first 10 consecutive frames of the second video to determine the motion type (e.g., running or standing). Based on the aforementioned motion vectors, joint coordinates, and motion type, the second motion feature is obtained. For example, the second motion feature is "standing." Tonal feature analysis is performed on the first frame image to obtain the second tonal feature. For example, the RGB three-channel distribution of the first frame image is calculated, and then converted to HSV space. The dominant hue and saturation are extracted in HSV space; for example, the dominant hue might be H=210°, S=30%, V=70% (cool blue tone). The light intensity, depth of field, ISO, and aperture value in the last frame image are identified. Based on the aforementioned dominant hue, saturation, light intensity, depth of field, ISO, and aperture value, a second hue feature is obtained. For example, the second hue feature might be "cinematic CG rendering, soft lighting, cool blue tone, dynamic blur coefficient 0.3". Style feature analysis is then performed on the first frame image to obtain a second style feature. For example, text-image alignment is performed on the previously obtained second subject feature, second scene feature, second action feature, and second hue feature to match style keywords, resulting in the corresponding second style feature. For example, if the matched style keyword is "Eastern fantasy," then the second style feature is "Eastern fantasy".

[0042] As described above, feature analysis of the first frame image identifies subject features, scene features, action features, tone features, and style features, thus translating the image into text and providing a semantic basis for generating the target transition video corresponding to the subsequent video's end frame. The fully automated generation of first frame features significantly improves processing efficiency compared to manual annotation, thereby enhancing the overall efficiency of the video's end frame transition.

[0043] S113. Generate the first prompt word corresponding to the tail frame image based on the tail frame features, and generate the first prompt word corresponding to the first frame image based on the first frame features.

[0044] The first prompt word for the last frame image is generated by parsing the features using a syntax tree to extract core elements, supplementing feature information based on these core elements, generating extended keywords, and integrating the core elements and extended keywords. The generated structured prompt word can then be refined using a corresponding language model to produce a fluent first prompt word. For example, the first prompt word for the last frame image could be: "In a realistic style, a red fox with golden-red fur runs vigorously in an autumn forest bathed in warm yellow sunset light, its bushy tail kicking up fallen leaves. The motion blur perfectly captures its agile posture. ISO is 400, f / 2.8 aperture, focusing on the fox's shimmering amber pupils." Similarly, the first prompt word for the first frame image is generated by parsing the features using a syntax tree to extract core elements, supplementing feature information based on these core elements, generating extended keywords, and integrating the core elements and extended keywords. The generated structured prompt word can then be refined using a corresponding language model to produce a fluent first prompt word. For example, the first prompt corresponding to the first frame image is "A nine-tailed fox spirit in a silver-white ink-patterned ruqun stands in front of a shrine on a rainy night, holding a bamboo talisman in her right hand, with a solemn posture."

[0045] As described above, based on the features of the final frame image and the first frame image to be stitched together, corresponding first prompt words are automatically generated for each, realizing image textification. This provides reference prompt words for subsequent video end-frame transitions, improving the matching accuracy between the target transition video generated based on the first prompt words and the final and first frame images to be stitched together. Furthermore, the entire generation process is automated, greatly improving processing efficiency and thus enhancing the overall efficiency of video end-frame transitions.

[0046] Figure 3 This is a flowchart illustrating another video end-frame transition method provided in the embodiments of this application, referred to... Figure 3 The video end-frame transition method specifically includes: S131. Extract features from the first target image to obtain a first feature, wherein the first feature includes one or more of the following: third subject feature, third scene feature, third action feature, third tone feature, and third style feature.

[0047] After generating the first target image in S12, feature extraction is performed on the first target image to obtain the first feature. The first feature includes one or more of the following: third subject feature, third scene feature, third action feature, third tone feature, and third style feature. For example, subject feature analysis is performed on the first target image to obtain the third subject feature. For instance, instance segmentation is performed on the first target image to locate the subject outline and output mask coordinates. The subject's visual features are mapped to the semantic space to generate a corresponding semantic vector, for example, ["golden red fox", "wild animal"]. The fur texture parameters (roughness σ=2.3) are then retrieved to obtain the third subject feature. Scene feature analysis is performed on the first target image to obtain the third scene feature. For example, image recognition is performed on the first target image to determine the scene type, for example, determining the scene type as "autumn forest". Key objects in the identified scene are detected to generate an object element table, for example, object elements include elements such as sunset, fallen leaves, tree trunks, light spots, and sky. The depth map of the identified scene is obtained, and the spatial position of the subject and environment is analyzed, for example, the fox is located in the foreground at a depth value of 0.3. Based on the aforementioned scene type, object element table, and spatial position of the subject and environment, a third scene feature is obtained. Action feature analysis is performed on the first target image to obtain a third action feature. For example, the motion vector of the subject in the first target image is calculated, where the motion vector includes direction and velocity. The joint coordinates of the subject are extracted to determine the action classification, for example, "dynamic running." Based on the aforementioned motion vector and joint coordinates, a third action feature is obtained. Tonal feature analysis is performed on the first target image to obtain a third tonal feature. For example, the RGB three-channel distribution of the first target image is calculated, and the RGB three-channel distribution is converted to HSV space. The dominant hue and saturation are extracted in the HSV space; for example, the dominant hue H=35° (warm yellow), saturation S=75%, and RGB mean (190, 170, 90) match the semantic meaning of "autumn" in the aforementioned scene features. Based on the aforementioned dominant hue and saturation, a third tonal feature is obtained. Style feature analysis is performed on the first target image to obtain a third style feature. For example, by performing text-image alignment processing on the aforementioned third subject feature, third scene feature, third action feature, and third tone feature, and matching style keywords, the corresponding third style feature is obtained. For example, if the matched style keyword is "realism," then the third style feature is "realism."

[0048] As described above, by performing feature analysis on the first target image, the main features, scene features, action features, tone features, and style features in the first target image are determined, so as to convert the image into text, providing a semantic basis for the generation of the target transition video corresponding to the subsequent video end frame transition. The entire process of automatically generating the first feature is much more efficient than manual annotation, thereby improving the overall efficiency of the video end frame transition.

[0049] S132. Extract features from the second target image to obtain a second feature, wherein the second feature includes one or more of the following: fourth subject feature, fourth scene feature, fourth action feature, fourth tone feature, and fourth style feature.

[0050] After generating the second target image in S12, feature extraction is performed on the second target image to obtain the second feature. The second feature includes one or more of the following: fourth subject feature, fourth scene feature, fourth action feature, fourth tone feature, and fourth style feature. For example, subject feature analysis is performed on the second target image to obtain the fourth subject feature. For instance, instance segmentation is performed on the second target image to locate the subject outline and output mask coordinates. The subject's visual features are mapped to the semantic space to generate a corresponding semantic vector, for example, the semantic vector is ["Nine-tailed fox demon", "Eastern mythological character"]. The clothing texture (silk gloss 0.7) is then inverted to obtain the fourth subject feature. Scene feature analysis is performed on the second target image to obtain the fourth scene feature. For example, image recognition is performed on the second target image to determine the scene type, for example, determining the scene type as "Japanese shrine". Key objects in the identified scene are detected to generate an object element table, for example, object elements include shrine, birds, rain, stone lanterns, and foxfire. The depth map of the identified scene is obtained, and the spatial position of the subject and environment is analyzed. For example, the depth map shows the fox demon is located 0.3 depths in front of the torii gate. Based on the aforementioned scene type, object element table, and spatial position of the subject and environment, a fourth scene feature is obtained. Motion feature analysis is performed on the second target image to obtain a fourth motion feature. For example, the motion vector of the subject in the second target image is calculated, where the motion vector includes direction and velocity. The joint coordinates of the subject are extracted; for example, the angle of the right hand holding the talisman is θ=120°, predicted as "static spell preparation." Based on the aforementioned motion vector and joint coordinates, a fourth motion feature is obtained. Tonal feature analysis is performed on the second target image to obtain a fourth tonal feature. For example, the RGB three-channel distribution of the first frame image is calculated, and the RGB three-channel distribution is converted to HSV space. In HSV space, the dominant hue and saturation are extracted; for example, H=210° (cool blue), S=30%, RGB mean (100, 120, 180), consistent with the semantics of "night shrine." Based on the aforementioned dominant hue and saturation, a fourth tonal feature is obtained. Style feature analysis is performed on the second target image to obtain a fourth style feature. For example, the aforementioned fourth subject feature, fourth scene feature, fourth action feature, and fourth tone feature are processed with text-image alignment to match style keywords, resulting in the corresponding second style feature. For instance, if the matched style keyword is "Eastern Fantasy," the ink brushstroke parameters (blur radius 5px) are extracted, and the lighting is set to moonlight soft light, then the second style feature is "Eastern Fantasy."

[0051] As described above, feature analysis of the second target image identifies subject features, scene features, action features, tone features, and style features, enabling the image to be transcribed into text. This provides a semantic basis for generating the target transition video corresponding to the subsequent video end frame transition. The fully automated generation of first frame features significantly improves processing efficiency compared to manual annotation, thereby enhancing the overall efficiency of the video end frame transition.

[0052] S133. Perform correlation analysis on the first feature and the second feature to obtain the correlation transition parameters.

[0053] A correlation analysis is performed on the first and second features obtained above to obtain the correlation transition parameters corresponding to each feature. Correlation transition parameters refer to a series of structured parameters used to quantify and control the feature change logic, time allocation, and visual effects from the original image (e.g., the first target image) to the target image (e.g., the second target image) during the video frame transition generation process. Their core characteristic is that they are dynamically generated through cross-frame feature correlation (such as differences in elements like subject shape, scene, and tone), rather than fixed values, ensuring that the transition effect is consistent with the semantics of the image content. By comparing the first feature of the first target image and the second feature of the second target image, a correlation analysis is performed to determine the correlation transition parameters corresponding to each feature. Subsequently, the transition process can be described based on these correlation transition parameters, providing a logical strategy for the video end-frame transition, thereby improving the efficiency of the video end-frame transition.

[0054] For example, the third subject feature in the first feature and the fourth subject feature in the second feature are analyzed for correlation to determine the subject transition parameters. That is, the subject features in the first target image and the subject features in the second target image are analyzed for correlation to determine the subject transition parameters of the two. For example, the semantic similarity between the two semantic vectors can be calculated based on the semantic vector in the third subject feature and the semantic vector in the fourth subject feature, and the subject transition parameters can be determined based on the semantic similarity. For example, if the cosine similarity between the two semantic vectors is 0.78, the transition logic of "morphological gradation" can be triggered when the animal becomes an anthropomorphic character. The subject transition parameters include parameters such as morphological interpolation coefficient and tail number change. For example, the morphological interpolation coefficient is a linear interpolation of the ears changing from pointed ears (0) to round ears (1), changing by 0.05 per frame; the number of tails increases stepwise from 1 (0) to 9 (1) (0.5 seconds + 2 tails per tail). As described above, by performing correlation analysis on the third subject feature in the first feature and the fourth subject feature in the second feature, the subject transition parameters are determined. This allows the transition logic of the subject in the first target image becoming the subject in the second target image to be determined based on the subject transition parameters. This provides a logical strategy for the subject transition in the video end frame transition, thereby improving the overall efficiency of the video end frame transition.

[0055] For example, a correlation analysis is performed on the third scene feature in the first feature and the fourth scene feature in the second feature to obtain scene transition parameters. That is, a correlation analysis is performed on the scene features in the first target image and the scene features in the second target image to determine the scene transition parameters between them. For example, the structural matching degree between the third and fourth scene features is calculated, and the scene transition parameters are obtained based on the structural matching degree. For example, the matching degree of vertical elements in the third and fourth scene features is detected; for example, the matching degree between the "tree trunk" in the third scene feature and the "torii pillar" in the fourth scene feature is 0.65. The "vertical structure replacement" transition path is determined based on the matching degree of vertical elements. Scene transition parameters include element replacement weights and depth-of-field synchronization parameters. For example, the element replacement weight could be that the tree outline is distorted into a torii shape at a rate of 0.05 frames / frame, completed in 3 seconds; the depth-of-field synchronization parameter is to keep the subject depth value constant at 0.3, and the background gradually changes from a deep forest (0.7-0.9) to a shrine (0.5-0.8). As described above, by performing correlation analysis on the third scene feature in the first feature and the fourth scene feature in the second feature, scene transition parameters are determined. This allows the transition logic of the scene in the first target image becoming the scene in the second target image to be determined based on the scene transition parameters. This provides a logical strategy for scene transition in the video end frame transition, thereby improving the overall efficiency of the video end frame transition.

[0056] For example, a correlation analysis is performed on the third motion feature in the first feature and the fourth motion feature in the second feature to obtain motion transition parameters. That is, a correlation analysis is performed on the motion features in the first target image and the motion features in the second target image to determine the motion transition parameters between them. For example, the motion transition parameters are determined based on the dynamic time warping of the third and fourth motion features. For example, the time axis of the running motion sequence in the third motion feature and the standing posture in the fourth motion feature is aligned to find key transition frames, such as frame 15 where the body is fully standing. Motion transition parameters include parameters such as joint angular velocity and motion blur coefficient. For example, the joint angular velocity changes by 6° per second from running (θ=110°) to holding the arm bent (θ=120°); the dynamic blur coefficient linearly decreases from 0.8 (running) to 0.3 (standing), matching the change in motion speed. As described above, by performing correlation analysis on the third action feature in the first feature and the fourth action feature in the second feature, action transition parameters are determined. This allows the transition logic of the main action in the first target image to become the main action in the second target image to be determined based on the action transition parameters. This provides a logical strategy for the main action transition in the video end frame transition, thereby improving the overall efficiency of the video end frame transition.

[0057] For example, a correlation analysis is performed on the third hue feature in the first feature and the fourth hue feature in the second feature to obtain the hue transition parameters. That is, a correlation analysis is performed on the hue features in the first target image and the hue features in the second target image to determine the hue transition parameters between them. For example, the hue transition parameters are determined based on the Lab color space difference between the third and fourth hue features. For example, if the Lab color space difference between the third and fourth hue features is ΔE=35, then the total transition time is calculated based on ΔE≤5 per frame, with the RGB values ​​decreasing by (13,7,-14) per frame. The hue transition parameters include color temperature transition parameters and saturation transition parameters. For example, the hue transition parameter is a color temperature increase of 100K per frame from 5500K (warm yellow) to 8000K (cool blue); the saturation transition parameter is a linear decrease in saturation from 75% to 30% to avoid color banding. The above-mentioned correlation analysis of the third hue feature in the first feature and the fourth hue feature in the second feature determines the hue transition parameters. This allows the transition logic of the hue in the first target image to become the hue in the second target image to be determined based on the hue transition parameters. This provides a logical strategy for the hue transition in the video end frame transition, thereby improving the overall efficiency of the video end frame transition.

[0058] For example, a correlation analysis is performed on the third style feature in the first feature and the fourth style feature in the second feature to obtain style transition parameters. That is, a correlation analysis is performed on the tonal features in the first target image and the tonal features in the second target image to determine the style transition parameters between them. For example, the style transition parameters are obtained based on the Gram matrix similarity between the third and fourth style features. The Gram matrix similarity quantifies the degree of matching between the Gram matrices of two images (or feature maps) in terms of texture style features by calculating the similarity between their Gram matrices. For example, the Gram matrix distance between the realistic texture in the third style feature and the ink wash style in the fourth style feature is 0.72, triggering the "texture blending" strategy. Style transition parameters include style blending weights and feature triggering parameters. For example, the style blending weight is 0.7 for the first 3 seconds of realistic texture and 0.8 for the next 4 seconds of fantasy texture; the feature triggering parameter is that the ink wash effect is activated when the subject deformation is 60% complete (2.1 seconds), with the intensity changing from 0 to 0.8. As described above, by performing correlation analysis on the third style feature in the first feature and the fourth style feature in the second feature, style transition parameters are determined. This allows for the determination of the transition logic from the style in the first target image to the style in the second target image based on the style transition parameters. This provides a logical strategy for style transition in the video end frame transition, thereby improving the overall efficiency of the video end frame transition.

[0059] S134. Generate a second prompt word based on the first feature, the second feature, and the associated transition parameters.

[0060] Based on the first feature, the second feature, and the corresponding associated transition parameters, the transition logic for each feature is determined. Transitional change prompts corresponding to each feature are generated based on the transition logic, and these prompts are integrated to obtain the final second prompt. By performing correlation analysis on the first feature corresponding to the first target image and the second feature corresponding to the second target image, a logically consistent transition description is generated, i.e., the second prompt. This provides transition logic for the automatic generation of the target transition video in the subsequent video tail frame transition, thereby improving the efficiency of subsequent target transition video generation.

[0061] For example, subject change prompts are generated based on the third subject feature, the fourth subject feature, and subject transition parameters. For instance, the generated subject change prompt is: "The subject gradually changes from a running golden-red fox to a silver-white fox demon holding a talisman; the ears become rounded frame by frame from pointed (changing by 0.05 per frame); the nine tails expand in stages from one (adding two tails every 0.5 seconds); the fur roughness decreases from σ=2.3 to σ=0.3; the subject depth value remains unchanged at 0.3." Scene change prompts are generated based on the third scene feature, the fourth scene feature, and scene transition parameters. For instance, the generated scene change prompt is: "The environment gradually changes from an autumn forest to a rainy night shrine; the trees are distorted into red torii gates with a weight of 0.05 per frame; fallen leaves are replaced with falling talismans; the depth of field maintains subject focus; the background depth transitions from 0.7-0.9 to 0.5-0.8; the element replacement is completed in 3 seconds." Action change prompts are generated based on the third action feature, the fourth action feature, and action transition parameters. For example, the generated action change cue words are: "The action transitions from running 30° to the right to standing with a talisman, the forelimb joint angle changes by 6° per second, the dynamic blur coefficient linearly decreases from 0.8 to 0.3, and the body is upright in the 15th frame, matching the spell preparation posture." Based on the third hue feature, fourth hue feature, and hue transition parameters, hue change cue words are generated. For example, the generated hue change cue words are: "The hue shifts frame by frame from H=35° warm yellow (S=75%) to H=210° cool blue (S=30%), with ΔE≤5 per frame, color temperature increases from 5500K to 8000K, and saturation decreases synchronously, ensuring a smooth color transition without abrupt changes." Based on the third style feature, fourth style feature, and style transition parameters, style blending cue words are generated. For example, the generated style blending prompt is: "The style blends realism with Eastern fantasy. For the first 3 seconds, 70% of the realistic texture is retained (ISO 400, f / 2.8). For the next 4 seconds, ink wash effects (radius 5px) and soft light effects are added. Foxfire particles (50-200) are activated when the subject is 60% deformed." Integrating the aforementioned subject change prompt, scene change prompt, action change prompt, tone change prompt, and style blending prompt, a second prompt is obtained.For example, the second prompt word obtained is: "The main body gradually changes from a golden-red running red fox to a silver-white fox demon holding a talisman. The ears become rounded frame by frame from pointed ears (changing by 0.05 per frame). The nine tails expand from 1 in stages (adding 2 tails every 0.5 seconds). The hair roughness decreases from σ=2.3 to σ=0.3. The environment gradually changes from an autumn forest to a rainy night shrine. The trees are distorted into red torii gates with a weight of 0.05 / frame. The fallen leaves are replaced with falling talismans. The action changes from running 30° to the right to standing with a talisman. The forelimb joint angle changes by 6° per second. The dynamic blur coefficient decreases from 0.8 to 0.3. The color tone changes from warm yellow at H=35° to cool blue at H=210° frame by frame. ΔE≤5 per frame. The color temperature rises to 8000K. The style blends realistic textures with ink painting fantasy. The realistic weight is 0.7 for the first 3 seconds. The ink painting effect and foxfire particles (number 50-200) are activated in the last 4 seconds. The total transition time is 7 seconds."

[0062] As mentioned above, feature correlation analysis can ensure that transition parameters conform to cultural logic. For example, "a fox transforms into a human body and then grows nine tails" can reduce logical gaps and improve narrative coherence compared to traditional linear transitions. This can improve the logical coherence of the video's end frame transition and thus enhance the user's viewing experience.

[0063] The above supports coordinated transitions across five dimensions: subject, scene, action, tone, and style, expanding the creative dimensions and improving the overall visual effect of the video's end frame transition.

[0064] Figure 4 This is a flowchart illustrating another video end-frame transition method provided in an embodiment of this application. (Refer to...) Figure 4 The video end-frame transition method specifically includes: S141. Obtain the set video sound effects and extract the beat points and emotional features of the video sound effects.

[0065] After generating intermediate video frames in S14, the set video sound effects can be obtained. These sound effects can be pre-set in the system, uploaded by the user in real time, or downloaded from a corresponding server. After obtaining the set video sound effects, the beat points and emotional features of these sound effects are extracted. For example, time-frequency analysis can be performed on the video sound effects, generating a spectrogram through short-time Fourier transform, and using the periodic energy peaks of the spectrogram as beat points. For instance, strong drum beats in electronic music are positioned as key beat points to trigger visual abrupt changes. The energy value of each beat point can be calculated to distinguish between strong beat points (such as the bass drum) and weak beat points (such as the hi-hat). For example, beat points with energy >70dB are defined as "strong trigger points," corresponding to visual keyframe switching. The emotional feature vector of the video sound effects can be extracted, mapped to an emotional dimension, and the corresponding emotional labels can be determined based on the emotional dimension, such as "tense," "cheerful," and "mysterious" emotional labels. The emotional features of the video sound effects are then determined based on these emotional labels.

[0066] S142. Integrate the start frame, intermediate video frames, and end frame to obtain the target video frame.

[0067] The video frames are stitched together in the order of start frame, middle video frames, and end frame to obtain the target video frame.

[0068] S143. Determine the visual keyframes in the target video frame, bind the beat points and visual keyframes, and generate the first target transition video.

[0069] After obtaining the target video frames, visual keyframes are identified within them. These visual keyframes include frames where the subject completes its deformation, midpoint frames during scene transitions, and visually prominent frames. The semantic similarity between video frames and prompts can be calculated. Video frames with a semantic similarity greater than a preset threshold (e.g., 0.85) are identified as visual keyframes, i.e., frames where the subject completes its deformation and midpoint frames during scene transitions. The color contrast and motion vector amplitude of video frames within the target video frame can also be calculated. Video frames with a color contrast greater than a preset threshold (e.g., ΔE > 15) are identified as visual keyframes, i.e., visually prominent frames; similarly, video frames with a motion vector amplitude greater than a preset threshold (e.g., greater than 10 pixels / frame) are identified as visual keyframes, i.e., visually prominent frames.

[0070] The first target transition video is generated by binding the aforementioned determined beat points of the video sound effects with visual keyframes in the target video frames. For example, the video timeline of the target video frame and the audio timeline of the video sound effects can be linearly aligned, and visual keyframes can be assigned according to the beat point positions. For instance, the strong beat point at 1.5 seconds can be bound to the visual keyframe indicating the completion of the subject's deformation (i.e., frame 1.5 seconds). Strong beat points (energy > 70dB) can be bound to visual keyframes corresponding to visual abrupt changes (such as instantaneous scene switching); weak beat points (energy < 50dB) can be bound to visual keyframes corresponding to changes in detail (such as a slight increase in particle count).

[0071] S144. Adjust the hue and saturation of the first target transition video according to the emotional characteristics to obtain the second target transition video.

[0072] After generating a first target transition video with sound effects, the hue and saturation of the first target transition video are adjusted according to the aforementioned emotional characteristics to obtain a second target transition video. For example, based on the emotional characteristics and a preset mapping relationship, corresponding hue and saturation adjustment values ​​can be determined. The hue of the first target transition video is adjusted according to the hue adjustment values ​​to obtain a target hue value; and the saturation of the first target transition video is adjusted according to the saturation adjustment values ​​to obtain a target saturation value, corresponding to the adjusted second target transition video. For example, in the preset mapping relationship, if the emotional characteristic is "cheerful," the corresponding hue adjustment value is: shifted towards a warmer tone by a preset value (e.g., H-30°), and the corresponding saturation adjustment value is: increased by 20%-30%. If the emotional characteristic is "mysterious," the corresponding hue adjustment value is: shifted towards a cooler tone by a preset value (e.g., H+60°), and the corresponding saturation adjustment value is: decreased by 15%-25%.

[0073] As described above, by binding the beat points of video sound effects with visual keyframes, users can obtain a synergistic viewing experience, thereby improving the overall audiovisual experience of the target transition video corresponding to the end frame transition. Furthermore, by guiding the color adjustment of the target transition video through emotional features, the emotional matching of audiovisual elements is ensured, thus enhancing the user's audiovisual experience. In addition, the entire process is automated, significantly improving the efficiency of audiovisual synchronization processing for the end frame transition compared to manual audio-visual synchronization methods.

[0074] As described above, by automating the generation of prompts and target transition videos, manual intervention is greatly reduced, the automation level of video end-frame transitions is improved, and the production efficiency of target transition videos in video end-frame transitions is significantly increased. Automated generation of the first and second prompts provides novel and personalized content, increasing the attractiveness and uniqueness of the target transition videos.

[0075] Based on the above embodiments, Figure 5 This is a schematic diagram of a video end-frame transition device provided in an embodiment of this application. (Reference) Figure 5 The video end-frame transition device provided in this embodiment specifically includes: an image acquisition module 21, a first prompt word generation module 22, a main vector generation module 23, a target image generation module 24, a feature data extraction module 25, a second prompt word generation module 26, an intermediate video frame generation module 27, and a video generation module 28.

[0076] Among them, the image acquisition module 21 is used to acquire the tail frame image and the first frame image to be stitched together; The first prompt word generation module 22 is used to generate the first prompt word corresponding to the tail frame image and the first frame image, respectively. The main vector generation module 23 is used to parse the main elements of the first prompt word to obtain the main vector; The target image generation module 24 is used to perform style construction processing based on the subject vector to generate a first target image and a second target image with corresponding styles; Feature data extraction module 25 is used to extract features from the first target image and the second target image to obtain feature data; The second prompt word generation module 26 is used to generate a second prompt word based on the feature data and the associated transition parameters of the determined feature data; Intermediate video frame generation module 27 is used to generate intermediate video frames based on the second prompt word; The video generation module 28 is used to generate a target transition video by integrating the start frame, intermediate video frames and the end frame, with the end frame image as the start frame and the first frame image as the end frame.

[0077] In one embodiment, the first prompt word generation module 22 includes: a tail frame feature determination submodule, a first frame feature determination submodule, and a first prompt word generation submodule; The image acquisition module 21 is also used to acquire the last frame image of the first video to be stitched together; The tail frame feature determination submodule is used to perform feature analysis on the tail frame image to obtain tail frame features, wherein the tail frame features include one or more of the following: first subject features, first scene features, first action features, first tone features, and first style features. The image acquisition module 21 is also used to acquire the first frame image of the second video to be stitched together; The first frame feature determination submodule is used to perform feature analysis on the first frame image to obtain the first frame features, wherein the first frame features include one or more of the following: second subject features, second scene features, second action features, second tone features, and second style features. The first prompt word generation submodule is used to generate the first prompt word corresponding to the last frame image based on the features of the last frame, and to generate the first prompt word corresponding to the first frame image based on the features of the first frame.

[0078] In one embodiment, the subject vector generation module 23 includes: a first subject vector generation submodule and a second subject vector generation submodule; The first main vector generation submodule is used to parse the main elements of the first prompt word corresponding to the last frame image to obtain the first main vector. The second day's main vector generation submodule is used to parse the main elements of the first prompt word corresponding to the first frame image to obtain the second main vector; The target image generation module 24 includes: a first style feature determination submodule, a second style feature determination submodule, a first target image generation submodule, and a second target image generation submodule; The first style feature determination submodule is used to extract style keywords from the first prompt word corresponding to the last frame image to obtain the first style feature; The second style feature determination submodule is used to extract style keywords from the first prompt word corresponding to the first frame image to obtain the second style feature; The first target image generation submodule is used to perform style construction processing based on the first subject vector and the first style features to generate a first target image with a first style. The second target image generation submodule is used to generate a second target image with a second style by performing style construction processing based on the second subject vector and the second style features.

[0079] In one embodiment, the feature data extraction module 25 includes: a first feature extraction submodule and a second feature extraction submodule; The first feature extraction submodule is used to extract features from the first target image to obtain a first feature, wherein the first feature includes one or more of the following: third subject feature, third scene feature, third action feature, third tone feature, and third style feature; The second feature extraction submodule is used to extract features from the second target image to obtain the second features, wherein the second features include one or more of the fourth subject features, fourth scene features, fourth action features, fourth tone features, and fourth style features; The second prompt word generation module 26 includes: an associated transition parameter determination submodule and a second prompt word generation submodule; The correlation transition parameter determination submodule is used to perform correlation analysis on the first feature and the second feature to obtain the correlation transition parameters; The second prompt word generation submodule is used to generate a second prompt word based on the first feature, the second feature, and the associated transition parameters.

[0080] In one embodiment, the associated transition parameter determination submodule includes: a main body transition parameter determination unit, a scene transition parameter determination unit, an action transition parameter determination unit, a tone transition parameter determination unit, and a style transition parameter determination unit; The main transition parameter determination unit is used to perform correlation analysis on the third main feature and the fourth main feature to obtain the main transition parameters; The scene transition parameter determination unit is used to perform correlation analysis on the third scene features and the fourth scene features to obtain the scene transition parameters; The motion transition parameter determination unit is used to perform correlation analysis on the third motion feature and the fourth motion feature to obtain the motion transition parameters; The tone transition parameter determination unit is used to perform correlation analysis on the third tone feature and the fourth tone feature to obtain the tone transition parameter; The style transition parameter determination unit is used to perform correlation analysis between the third style feature and the fourth style feature to obtain the style transition parameter.

[0081] In one embodiment, the second prompt word generation submodule includes: a subject change prompt word generation unit, a scene change prompt word generation unit, an action change prompt word generation unit, a tone change prompt word generation unit, a style fusion prompt word generation unit, and a second prompt word generation unit; The main body change prompt word generation unit is used to generate main body change prompt words based on the third main body feature, the fourth main body feature, and the main body transition parameter; The scene change prompt word generation unit is used to generate scene change prompt words based on the third scene features, the fourth scene features, and scene transition parameters; The action change cue word generation unit is used to generate action change cue words based on the third action feature, the third action feature and the action transition parameters; The tone change cue word generation unit is used to generate tone change cue words based on the third tone feature, the fourth tone feature, and the tone transition parameter; The style fusion prompt generation unit is used to generate style fusion prompts based on the third style feature, the fourth style feature, and the style transition parameter; The second prompt word generation unit is used to integrate subject change prompt words, scene change prompt words, action change prompt words, color tone change prompt words, and style fusion prompt words to obtain the second prompt word.

[0082] In one embodiment, the video generation module 28 includes: a sound effect acquisition submodule, a beat emotion extraction submodule, a video frame integration submodule, a beat binding submodule, a first video generation submodule, and a second video generation submodule; The audio effects acquisition submodule is used to acquire the set video audio effects; The beat emotion extraction submodule is used to extract the beat points and emotional features of video sound effects; The video frame integration submodule is used to integrate the start frame, middle video frames, and end frame to obtain the target video frame; The beat binding submodule is used to determine the visual keyframes in the target video frame and bind the beat points to the visual keyframes; The first video generation submodule is used to generate the first target transition video; The second video generation submodule is used to adjust the hue and saturation of the first target transition video based on emotional features to obtain the second target transition video.

[0083] The video end-frame transition device provided in this application embodiment can be used to execute the video end-frame transition method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0084] This application provides a video end-frame transition device, referring to... Figure 6 The video end-frame transition device includes a processor 31, a memory 32, a communication module 33, an input device 34, and an output device 35. The number of processors and the number of memories in the video end-frame transition device can be one or more. The processor, memory, communication module, input device, and output device of the video end-frame transition device can be connected via a bus or other means.

[0085] The memory 32, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the video end-frame transition method in any embodiment of this application (e.g., the image acquisition module, first prompt word generation module, main vector generation module, target image generation module, feature data extraction module, second prompt word generation module, intermediate video frame generation module, and video generation module in the video end-frame transition device). The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created according to the use of the device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the device via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0086] The communication module 33 is used for data transmission.

[0087] The processor 31 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory, thereby realizing the above-mentioned video end frame transition method.

[0088] Input device 34 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 35 may include display devices such as a display screen.

[0089] The video end-frame transition device provided above can be used to execute the video end-frame transition method provided in the above embodiments, and has corresponding functions and beneficial effects.

[0090] This application embodiment also provides a storage medium for storing computer-executable instructions. When executed by a computer processor, the computer-executable instructions are used to execute a video end-frame transition method. The video end-frame transition method includes: acquiring an end-frame image to be stitched and a first-frame image; generating a first prompt word corresponding to each of the end-frame image and the first-frame image; parsing the first prompt word to obtain a subject vector, and performing style construction processing based on the subject vector to generate a first target image and a second target image with corresponding styles; extracting features from the first target image and the second target image to obtain feature data, and generating a second prompt word based on the feature data and the associated transition parameters of the determined feature data; generating an intermediate video frame based on the second prompt word, and integrating the start frame, intermediate video frame, and end frame with the end-frame image as the start frame and the first-frame image as the end frame to generate a target transition video.

[0091] Storage medium – any type of memory device or storage device. The term “storage medium” is intended to include: mounting media, such as CD-ROM, floppy disk, or magnetic tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements, etc. Storage medium may also include other types of memory or combinations thereof. Furthermore, storage medium may reside in a first computer system in which the program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term “storage medium” can include two or more storage media residing in different locations (e.g., in different computer systems connected via a network). Storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0092] Of course, the storage medium for storing computer-executable instructions provided in the embodiments of this application is not limited to the video end-frame transition method described above, but can also execute related operations in the video end-frame transition method provided in any embodiment of this application.

[0093] The video end-frame transition device, storage medium, and video end-frame transition equipment provided in the above embodiments can execute the video end-frame transition method provided in any embodiment of this application. For technical details not described in detail in the above embodiments, please refer to the video end-frame transition method provided in any embodiment of this application.

[0094] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the claims.

Claims

1. A method for transitioning the end frame of a video, characterized in that, include: Obtain the tail frame image and the first frame image to be stitched together, and generate the first prompt word corresponding to each of the tail frame image and the first frame image; The first prompt word is parsed to obtain a main vector, and style construction processing is performed based on the main vector to generate a first target image and a second target image with corresponding styles; Feature data is obtained by extracting features from the first target image and the second target image, and a second prompt word is generated based on the feature data and the determined associated transition parameters of the feature data. An intermediate video frame is generated based on the second prompt word. The starting frame is the last frame image, and the ending frame is the first frame image. The starting frame, the intermediate video frame, and the ending frame are then integrated to generate the target transition video.

2. The method according to claim 1, characterized in that, The step of acquiring the tail frame image and the first frame image to be stitched together, and generating the first prompt word corresponding to each of the tail frame image and the first frame image, includes: The last frame image of the first video to be stitched is obtained, and feature analysis is performed on the last frame image to obtain the last frame features. The last frame features include one or more of the following: first subject features, first scene features, first action features, first tone features, and first style features. The first frame image of the second video to be stitched is obtained, and feature analysis is performed on the first frame image to obtain the first frame features. The first frame features include one or more of the following: second subject features, second scene features, second action features, second tone features, and second style features. A first prompt word corresponding to the tail frame image is generated based on the tail frame features, and a first prompt word corresponding to the first frame image is generated based on the first frame features.

3. The method according to claim 1, characterized in that, The step of parsing the main elements of the first prompt word to obtain the main vector includes: The first subject vector is obtained by parsing the first prompt word corresponding to the last frame image and the second subject vector is obtained by parsing the first prompt word corresponding to the first frame image. Accordingly, the step of generating a first target image and a second target image with corresponding styles based on the subject vector through style construction processing includes: A first style feature is obtained by extracting style keywords from the first prompt word corresponding to the last frame image, and a second style feature is obtained by extracting style keywords from the first prompt word corresponding to the first frame image. A first target image with a first style is generated by performing style construction processing based on the first subject vector and the first style feature, and a second target image with a second style is generated by performing style construction processing based on the second subject vector and the second style feature.

4. The method according to claim 1, characterized in that, The step of extracting feature data based on the first target image and the second target image includes: The first feature is obtained by extracting features from the first target image. The first feature includes one or more of the following: third subject feature, third scene feature, third action feature, third tone feature, and third style feature. The second feature is obtained by performing feature extraction on the second target image. The second feature includes one or more of the following: fourth subject feature, fourth scene feature, fourth action feature, fourth tone feature, and fourth style feature. Accordingly, generating the second prompt word based on the feature data and the determined associated transition parameters of the feature data includes: A correlation analysis is performed on the first feature and the second feature to obtain the correlation transition parameters; The second prompt word is generated based on the first feature, the second feature, and the associated transition parameter.

5. The method according to claim 4, characterized in that, The step of performing correlation analysis based on the first feature and the second feature to obtain correlation transition parameters includes: The third subject feature and the fourth subject feature are subjected to correlation analysis to obtain subject transition parameters; the third scene feature and the fourth scene feature are subjected to correlation analysis to obtain scene transition parameters. The third action feature and the fourth action feature are correlated to obtain action transition parameters, and the third hue feature and the fourth hue feature are correlated to obtain hue transition parameters. A correlation analysis is performed on the third style feature and the fourth style feature to obtain the style transition parameter.

6. The method according to claim 5, characterized in that, The step of generating the second prompt word based on the first feature, the second feature, and the associated transition parameter includes: Based on the third subject feature, the fourth subject feature, and the subject transition parameter, generate subject change prompt words; based on the third scene feature, the fourth scene feature, and the scene transition parameter, generate scene change prompt words. Based on the third action feature, the third action feature and the action transition parameter, generate action change prompt words; based on the third hue feature, the fourth hue feature and the hue transition parameter, generate hue change prompt words. Based on the third style feature, the fourth style feature, and the style transition parameter, a style fusion prompt word is generated. The subject change prompt word, the scene change prompt word, the action change prompt word, the tone change prompt word, and the style fusion prompt word are integrated to obtain the second prompt word.

7. The method according to any one of claims 1-6, characterized in that, The process of integrating the starting frame, the intermediate video frames, and the ending frame to generate the target transition video includes: Obtain the set video sound effects, and extract the beat points and emotional features of the video sound effects; The target video frame is obtained by integrating the starting frame, the intermediate video frames, and the ending frame. Determine the visual keyframes in the target video frame, bind the beat points and the visual keyframes, and generate the first target transition video; The hue and saturation of the first target transition video are adjusted based on emotional characteristics to obtain the second target transition video.

8. A video end-frame transition device, characterized in that, include: The image acquisition module is used to acquire the last frame image and the first frame image to be stitched together; The first prompt word generation module is used to generate the first prompt word corresponding to the tail frame image and the first frame image respectively; The main vector generation module is used to parse the main elements of the first prompt word to obtain the main vector; The target image generation module is used to perform style construction processing based on the subject vector to generate a first target image and a second target image with the corresponding style. The feature data extraction module is used to extract features from the first target image and the second target image to obtain feature data; The second prompt word generation module is used to generate a second prompt word based on the feature data and the determined associated transition parameters of the feature data; An intermediate video frame generation module is used to generate intermediate video frames based on the second prompt word; The video generation module is used to generate a target transition video by integrating the start frame, the intermediate video frames, and the end frame, with the end frame image as the start frame and the first frame image as the end frame.

9. A video end-frame transition device, characterized in that, include: Memory and one or more processors; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A storage medium for storing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a processor, are used to perform the method as described in any one of claims 1-7.

Images