Denoising video generation method based on multi-condition control

Abstract

The invention discloses a de-noised video generation method based on multi-condition control, and relates to the technical field of computer vision and image processing, and the method comprises the steps: specifying a motion track, converting the motion track into a motion mark of a target object, setting a text and an emotion label, obtaining condition information representation based on the text and the emotion label, and generating a de-noised video. Specifying a lens control option as a camera motion mark, obtaining a first Gaussian noise, obtaining a submerged space representation of the Gaussian noise based on the first Gaussian noise, inputting the motion mark of the target object, the time step, the condition information representation and the submerged space representation of the Gaussian noise into a video generation model, and outputting predicted noise, the subsurface space representation of the video is obtained based on the predicted noise, and the subsurface space representation of the video is processed by the decoder to obtain the specified video, so that multi-condition control video generation is realized, and the video generation quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and image processing, and in particular to a denoising video generation method based on multi-condition control. Background Art

[0002] Video generation technology is an important research direction in the fields of computer vision, artificial intelligence, and multimedia. It aims to automatically generate high-quality video content through algorithms and models. With the rapid development of deep learning technology, video generation technology has been widely used in film and television production, game development, virtual reality (VR), augmented reality (AR), and advertising and marketing.

[0003] Traditional video generation methods usually rely on rule-based systems or simple machine learning models. Due to the limitations of model complexity and computing resources, the video quality generated by traditional methods is often not high, with problems such as blurred images and loss of details, which makes it difficult to meet the needs of high-quality video generation. At the same time, traditional methods find it difficult to simultaneously handle the joint control of multiple conditions such as text, emotions, target motion and camera motion. Summary of the Invention

[0004] In order to overcome the defects of the above-mentioned prior art in that the video generation quality is low and multi-condition control cannot be performed, the present invention proposes a denoising video generation method based on multi-condition control.

[0005] To achieve the above objectives, the present invention adopts the following technical solutions, a denoised video generation method based on multi-condition control, comprising:

[0006] S1: Draw the specified motion trajectory on the canvas and convert it into the motion mark of the target object;

[0007] S2: Set text and emotion labels, and obtain conditional information representation based on the text and emotion labels;

[0008] S3: specifying a lens control option as a camera motion marker, obtaining a first Gaussian noise, and obtaining a latent space representation of the Gaussian noise based on the first Gaussian noise; wherein the latent space representation of the Gaussian noise is obtained by compressing the first Gaussian noise through an encoder;

[0009] S4: Input the motion label, time step, conditional information representation, and Gaussian noise latent space representation of the target object into the video generation model and output the predicted noise;

[0010] S5: Obtain a latent space representation of the video based on the predicted noise, and process the latent space representation of the video through a decoder to obtain a specified video.

[0011] Preferably, the training process of the video generation model includes:

[0012] SA1: Collect and process public video data to obtain a video emotion classification dataset;

[0013] SA2: Use encoder and decoder to compress and decompress the videos in the video emotion classification dataset to obtain the latent space representation z0;

[0014] SA3: Perform forward noise processing on the latent space representation to obtain the diffused latent space noise representation z t ;

[0015] SA4: Obtain conditional information representation c based on text and sentiment labels;

[0016] SA5: Using the target module of the video generation model, the motion mark TR of the target object is converted into the target motion feature representation OM;

[0017] SA6: Target motion feature representation OM, time step t, conditional information representation c, latent space noise representation z t ,The camera motion marker RT is input into the U-net denoising module of the video generation model, and is transformed and iterated through multiple convolutional layers, spatial layers, and temporal layers to obtain the predicted noise.

[0018] Preferably, in step SA1, collecting public video data, processing the public video data, and obtaining a video emotion classification dataset include:

[0019] SA11: Get a public video dataset D = {X1, X2, ..., X i ,...,X N}, perform camera motion labeling on public video data to obtain the camera motion labeling dataset D t :

[0020] D t ={(X1,RT1),(X2,RT2),...,(X i ,RT i ),...,(X N ,RT N )};

[0021] Among them, i is the video number, X i is the i-th video, N is the total number of videos, RT i is the camera motion marker of the i-th video, RT i ={rt0,rt1,...,rt l ,...,rt L-1}, rt l is the camera motion representation of the lth frame of the video, RT∈R L×12 , L is the frame length of the video;

[0022] SA12: Camera Motion Labeling Dataset D t Perform target object motion labeling to obtain the target object motion labeling dataset D o :

[0023] D o ={(X1,RT1,TR1),(X2,RT2,TR2),...,(X i ,RT i ,TR i ),...,(X N ,RT N ,TR N )};

[0024] Among them, TR i is the motion mark of the target object in the i-th video;

[0025] SA13: Based on the target object motion labeling dataset, a multimodal pre-trained model is used to synthesize text from video to obtain the video text dataset D T :

[0026] D T ={(X1,RT1,TR1,Te1),(X2,RT2,TR2,Te2),...,(X i ,RT i ,TR i ,Te i ),...,(X N ,RT N ,TR N ,Te N )};

[0027] Among them, Te i is the text of the i-th video;

[0028] SA14: Based on video text dataset D T , perform sentiment classification on the text of each video and obtain the video sentiment classification dataset De:

[0029] De={(X1,RT1,TR1,Te1,E1),(X2,RT2,TR2,Te2,E2),...,(X i ,RT i ,TR i ,Te i ,E i ),...,(X N ,RT N ,TR N ,Te N ,EN )};

[0030] Among them, E i is the sentiment classification label corresponding to the i-th video.

[0031] Preferably, the latent space noise representation z t The calculation formula is:

[0032]

[0033] Where t is the time step number, T is the total number of time steps, ∈ represents the second Gaussian noise, α t represents the scheduling weight for the second Gaussian noise at time step t, and z0 is the latent space representation.

[0034] Preferably, the calculation formula of the loss function L of the video generation model is:

[0035]

[0036] Among them, ∈ θ (z t ,c,RT,TR) is the model output, i.e., the predicted noise, t is the time step number, θ is the model parameter, ∈ represents the second Gaussian noise, RT is the camera motion marker, TR is the target object motion marker, E represents the mathematical expectation, is the L2 norm.

[0037] Preferably, the calculation formula of the condition information representation c is:

[0038] c=E CLIP (concat(Te i ,E i ))

[0039] Among them, E CLIP Indicates encoding operation through the CLIP module, and concat indicates concatenation processing.

[0040] Preferably, the video generation model includes a target module and a U-net denoising module, the target module is used to convert the motion mark of the target object into a target motion feature representation, and the U-net denoising module is used to obtain predicted noise.

[0041] Preferably, the U-net denoising module includes a downsampling stage, an intermediate stage, and an upsampling stage, each stage includes at least one block, and each block includes a convolutional layer, a spatial layer, and a temporal layer connected in sequence; the target module includes several blocks, each block includes at least one convolutional layer, and the total number of blocks in the target module is equal to the sum of the total number of blocks in the downsampling stage and the intermediate stage in the U-net denoising module.

[0042] Preferably, the conditional information representation and camera motion labeling are input into each block of each stage of the U-net denoising module; the output of each block in the target module is input into each block of the downsampling stage and the intermediate stage of the U-net denoising module.

[0043] Preferably, in step S5, the latent space representation of the video is obtained by subtracting the latent space representation of Gaussian noise from the predicted noise.

[0044] The advantages of the present invention are:

[0045] (1) The present invention specifies the motion trajectory, sets text and emotion tags to obtain conditional information representation, specifies lens control options, inputs the motion mark, time step, conditional information representation, and Gaussian noise latent space representation of the target object into the video generation model, outputs predicted noise, and obtains the specified video based on the predicted noise. Compared with the traditional video generation method, the present invention realizes multi-condition control to generate video, thereby improving the quality of video generation.

[0046] (2) The present invention integrates a target module based on the U-Net denoising module to convert the motion mark of the target object into a target motion feature representation, thereby achieving the control of the target motion trajectory at a lower computational cost.

[0047] (3) The present invention integrates lens control options based on the U-Net denoising module to achieve video lens motion control at a lower computational cost.

[0048] (4) The video generation model of the present invention includes a target module and a U-net denoising module. By setting text and emotion tags, conditional information representation is obtained based on the text and emotion tags, and then the conditional information representation is input into the video generation model, so that the emotion tags and text are mixed and embedded in the U-net denoising module, thereby achieving a customizable effect of the emotional atmosphere of the video.

[0049] (5) The present invention realizes a multi-condition controlled video generation framework by integrating multiple conditional input information, so that the generated video images are clear and the effect is excellent. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a model architecture diagram of the present invention;

[0051] Figure 2 The figure is a flow chart of the steps of the method of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] Traditional video generation methods primarily rely on rule-based motion interpolation or single-condition generative models (e.g., text-to-video). These methods suffer from two major technical bottlenecks: First, the generated videos are dynamic and monotonous, making it difficult to coordinate the complex relationships between target motion, camera pose, and high-level semantics (e.g., emotion); second, they suffer from insufficient interactive flexibility, preventing users from precisely controlling the generated content through intuitive methods (e.g., drawing trajectories). Existing techniques, such as purely text-driven generative models (e.g., diffusion models), can generate high-resolution videos, but their ability to fuse multimodal conditions (text, motion, and footage) is limited, resulting in significant deviations from user intent and insufficient dynamic coherence.

[0054] like Figure 1-2 As shown, the present invention proposes a denoising video generation method based on multi-condition control, including:

[0055] S1: Draw the specified motion trajectory on the canvas and convert it into the motion mark of the target object;

[0056] The motion tagging of target objects is a structured data representation used to describe the motion trajectory, posture changes or other dynamic properties of specific targets (such as people, vehicles, etc.). Its core function is to provide the video generation model with precise control conditions for the target motion, ensuring that the motion of the target object in the generated video meets the user's expectations or real physical laws.

[0057] S2: Set text and emotion labels, and obtain conditional information representation based on the text and emotion labels;

[0058] S3: specifying a lens control option as a camera motion marker; obtaining a first Gaussian noise, and obtaining a latent space representation of the Gaussian noise based on the first Gaussian noise; wherein the latent space representation of the Gaussian noise is obtained by compressing the first Gaussian noise through an encoder;

[0059] Specified lens control options are user- or system-defined camera motion parameters that define the camera's motion, trajectory, and perspective during video generation. These options are typically provided as interactive parameters, such as drop-down menus, sliders, or path drawing tools.

[0060] Camera motion tagging is a structured data representation that encodes lens control options into machine-processable tags (such as vectors, matrices, or time series) as input conditions for video generation models. Its core is to convert user intentions into computable camera motion parameters.

[0061] S4: Input the motion label, time step, conditional information representation, and Gaussian noise latent space representation of the target object into the video generation model and output the predicted noise;

[0062] S5: Obtain a latent space representation of the video based on the predicted noise, and process the latent space representation of the video through a decoder to obtain a specified video.

[0063] The latent space representation of the video is obtained by subtracting the latent space representation of Gaussian noise from the predicted noise.

[0064] The training process of the video generation model includes:

[0065] SA1: Collect and process public video data to obtain a video emotion classification dataset, including:

[0066] SA11: Get a public video dataset D = {X1, X2, ..., X i ,...,X N}, perform camera motion labeling on public video data to obtain the camera motion labeling dataset D t :

[0067] D t ={(X1,RT1),(X2,RT2),...,(X i ,RT i ),...,(X N ,RT N )};

[0068] Among them, i is the video number, X i is the i-th video, N is the total number of videos, RT i is the camera motion marker of the i-th video, RT i ={rt0,rt1,...,rt l ,...,rt L-1}, rt l is the camera motion representation of the lth frame of the video, RT∈R L×12 , L is the frame length of the video;

[0069] Camera motion representation is a structured data form that accurately describes the pose (position and orientation) of a camera in each frame of a video and its motion trajectory over time. Its core purpose is to convert camera motion in the physical world into machine-processable computational parameters, thereby achieving programmatic control of the video generation and analysis process.

[0070] SA12: Camera Motion Labeling Dataset D t Perform target object motion labeling to obtain the target object motion labeling dataset D o :

[0071] D o ={(X1,RT1,TR1),(X2,RT2,TR2),...,(X i ,RT i ,TR i ),...,(X N ,RT N ,TR N )};

[0072] Among them, TR i is the motion mark of the target object in the i-th video;

[0073] SA13: Based on the target object motion labeling dataset, a multimodal pre-trained model (blip2 model) is used to synthesize text from videos to obtain the video text dataset D T :

[0074] D T ={(X1,RT1,TR1,Te1),(X2,RT2,TR2,Te2),...,(X i ,RT i ,TR i ,Te i ),...,(X N ,RT N ,TR N ,Te N )};

[0075] Among them, Te i is the text prompt of the i-th video;

[0076] The blip2 model is a ready-made multimodal pre-trained model. Its input is video data with motion tags. The blip2 model can simultaneously understand image and text information, extract video key frames, and feed them into the blip2 model's visual encoder to obtain image features. The blip2 model's Q-Former (a learnable query transformer) aligns visual and language features, automatically generates text descriptions, and forms a video text data dataset.

[0077] SA14: Based on video text dataset D T , perform sentiment classification on the text of each video and obtain the video sentiment classification dataset De:

[0078] De={(X1,RT1,TR1,Te1,E1),(X2,RT2,TR2,Te2,E2),...,(X i ,RT i ,TR i ,Te i ,E i ),...,(X N ,RT N ,TR N ,Te N ,E N )};

[0079] Among them, E i is the sentiment classification label corresponding to the i-th video.

[0080] SA2: Use encoder and decoder to compress and decompress the videos in the video emotion classification dataset to obtain the latent space representation z0;

[0081] The formula is: z0=D(X i ); where D(.) is for compression and decompression operations;

[0082] SA3: Perform forward noise processing on the latent space representation z0 to obtain the diffused latent space noise representation z t , the calculation formula is:

[0083]

[0084] Among them, t is the time step number, T is the total number of time steps, ∈ represents the second Gaussian noise and obeys the standard normal distribution, α t represents the scheduling weight for the second Gaussian noise at time step t;

[0085] SA4: Obtain conditional information representation c based on text and sentiment labels. The calculation formula is:

[0086] c=E CLIP (concat(Te i ,E i ))

[0087] Among them, E CLIP Indicates encoding operation through the public CLIP module, and concat indicates concatenation processing;

[0088] SA5: Using the target module of the video generation model, the motion mark TR of the target object is converted into the target motion feature representation;

[0089] The target motion feature is expressed as OM={m1,m2,...,m k ,...,m mid};

[0090] SA6: Target motion feature representation OM, time step t, conditional information representation c, latent space noise representation z t ,The camera motion marker RT is input into the U-net denoising module of the video generation model, and it is transformed and iterated through multiple convolutional layers, spatial layers, and temporal layers to obtain the predicted noise;

[0091] The calculation formula of the loss function L of the video generation model is:

[0092]

[0093] Among them, ∈ θ (z t ,c,RT,TR) is the model output, i.e., the predicted noise, θ is the model parameter, ∈ represents the second Gaussian noise, RT is the camera motion marker, TR is the target object motion marker, E represents the mathematical expectation, is the L2 norm.

[0094] The U-Net denoising module includes a downsampling stage, an intermediate stage, and an upsampling stage, and is composed of a block set H = {h1,h2,...,h M ,h mid ,h M+1 ,...,h 2M}, and each block includes a convolutional layer, a spatial layer, and a temporal layer connected in sequence; where h1,h2,...,h M They are the 1st block, 2nd block, ..., Mth block in the downsampling stage. The U-Net denoising module reduces the resolution while increasing the channel dimension through layer-by-layer convolution operations, gradually transitioning the low-level features to the global representation, h mid The block in the middle stage, i.e. the mid block, can capture the global features of the data, h M+1 ,...,h 2M They are the M+1th block, ..., 2Mth block in the upsampling stage, gradually restoring the spatial resolution of the features. In this embodiment, a convolutional layer, a spatial layer, and a temporal layer arranged in sequence are regarded as a block, for example, Represents the first block, including the first layer, the second layer and the third layer of the first block, which correspond to the convolutional layer, the spatial layer and the temporal layer respectively; at the same time, the target module includes several blocks, each block includes at least one convolutional layer, and the total number of blocks is equal to the sum of the total number of blocks in the downsampling stage and the intermediate stage in the U-net denoising module.

[0095] In the denoising process, TR i After the target module is processed, the target motion feature representation OM = {m1,m2,...,m k ,...,m mid}, the output of the corresponding block in the target module will be input into the downsampling stage and the intermediate stage of the U-net denoising module; c,RT i It will be input into each block of each stage of the U-net denoising module;

[0096] The target motion feature representation OM is embedded in the U-net denoising module as follows:

[0097]

[0098] in, They represent the output and input of the first layer of the kth block in the U-net denoising module respectively;

[0099] The text condition information c is embedded in the U-net denoising module as follows:

[0100]

[0101] in, They represent the input and output of the second layer of the kth block respectively, CrossAttn(.) is the cross attention mechanism, and Conv(.) is the convolution operation;

[0102] The camera motion marker RT i The embedding representation of the U-net denoising module is:

[0103]

[0104] in, They represent the output and input of the third layer of the kth block in the U-net denoising module, MLP(.) represents a three-layer feedforward neural network, and reshape(.) is to mark the camera motion RT i The dimension is expanded, and SelfAttn(.) is the self-attention mechanism.

[0105] The method of the present invention (Mymethod) is compared with the AnimateDiff method and the VideoComposer method. The results are shown in Table 1.

[0106] Table 1 Comparison of effectiveness of various methods

[0107] AnimateDiff VideoComposer Mymethod CLIP 0.2252 0.2223 0.2307 FID 154.69 132.02 124.26 FVD 1834.14 954.39 873.12

[0108] CLIP represents semantic similarity. A larger value indicates a higher similarity between the generated result and the text. FID represents image authenticity. A smaller value indicates a more realistic generated result. FVD also represents video authenticity. A smaller value indicates a more realistic generated result. Note: The FID evaluation result is evaluated by extracting frames from the video to obtain images.

[0109] From the above indicators, it can be seen that the method of the present invention, under the background of multi-condition controllable generation, not only improves the semantic consistency of the generated video, but also significantly improves the realism and dynamic performance of the video, which is superior to the current mainstream methods.

[0110] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0111] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0112] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A denoising video generation method based on multi-condition control, characterized in that: include: S1: Draw the specified motion trajectory on the canvas and convert it into the motion mark of the target object; S2: Set text and emotion labels, and obtain conditional information representation based on the text and emotion labels; S3: specifying a lens control option as a camera motion marker, obtaining a first Gaussian noise, and obtaining a latent space representation of the Gaussian noise based on the first Gaussian noise; wherein the latent space representation of the Gaussian noise is obtained by compressing the first Gaussian noise through an encoder; S4: Input the motion label, time step, conditional information representation, and Gaussian noise latent space representation of the target object into the video generation model and output the predicted noise; S5: Obtain a latent space representation of the video based on the predicted noise, and process the latent space representation of the video through a decoder to obtain a specified video.

2. The denoising video generation method based on multi-condition control according to claim 1, characterized in that: The training process of the video generation model includes: SA1: Collect and process public video data to obtain a video emotion classification dataset; SA2: Use encoder and decoder to compress and decompress the videos in the video emotion classification dataset to obtain the latent space representation z0; SA3: Perform forward noise processing on the latent space representation to obtain the diffused latent space noise representation z t ; SA4: Obtain conditional information representation c based on text and sentiment labels; SA5: Using the target module of the video generation model, the motion mark TR of the target object is converted into the target motion feature representation OM; SA6: Target motion feature representation OM, time step t, conditional information representation c, latent space noise representation z t ,The camera motion marker RT is input into the U-net denoising module of the video generation model, and is transformed and iterated through multiple convolutional layers, spatial layers, and temporal layers to obtain the predicted noise.

3. The denoising video generation method based on multi-condition control according to claim 2, characterized in that: In step SA1, public video data is collected and processed to obtain a video emotion classification dataset, including: SA11: Get a public video dataset D = {X1, X2, ..., X i ,...,X N }, perform camera motion labeling on public video data to obtain the camera motion labeling dataset D t : D t ={(X1,RT1),(X2,RT2),...,(X i ,RT i ),...,(X N ,RT N )}; Among them, i is the video number, X i is the i-th video, N is the total number of videos, RT i is the camera motion marker of the i-th video, RT i ={rt0,rt1,...,rt l ,...,rt L-1 }, rt l is the camera motion representation of the lth frame of the video, RT∈R L×12 , L is the frame length of the video; SA12: Camera Motion Labeling Dataset D t Perform target object motion labeling to obtain the target object motion labeling dataset D o : D o ={(X1,RT1,TR1),(X2,RT2,TR2),...,(X i ,RT i ,TR i ),...,(X N ,RT N ,TR N )}; Among them, TR i is the motion mark of the target object in the i-th video; SA13: Based on the target object motion labeling dataset, a multimodal pre-trained model is used to synthesize text from video to obtain the video text dataset D T : D T ={(X1,RT1,TR1,Te1),(X2,RT2,TR2,Te2),...,(X i ,RT i ,TR i ,Te i ),...,(X N ,RT N ,TR N ,Te N )}; Among them, Te i is the text of the i-th video; SA14: Based on video text dataset D T , perform sentiment classification on the text of each video and obtain the video sentiment classification dataset De: De={(X1,RT1,TR1,Te1,E1),(X2,RT2,TR2,Te2,E2),...,(X i ,RT i ,TR i ,Te i ,E i ),...,(X N ,RT N ,TR N ,Te N ,E N )}; Among them, E i is the sentiment classification label corresponding to the i-th video.

4. The method for generating denoised video based on multi-condition control according to claim 2, wherein: Latent space noise representation z t The calculation formula is: Where t is the time step number, T is the total number of time steps, ∈ represents the second Gaussian noise, α t represents the scheduling weight for the second Gaussian noise at time step t, and z0 is the latent space representation.

5. The method for generating denoised video based on multi-condition control according to claim 3, wherein: The calculation formula of the loss function L of the video generation model is: Among them, ∈ θ (z t ,c,RT,TR) is the model output, i.e., the predicted noise, t is the time step number, θ is the model parameter, ∈ represents the second Gaussian noise, RT is the camera motion marker, TR is the target object motion marker, E represents the mathematical expectation, is the L2 norm.

6. The method for generating denoised video based on multi-condition control according to claim 3, wherein: The calculation formula of condition information representation c is: c=E CLIP (concat(Te i ,It i )) Among them, E CLIP Indicates encoding operation through CLIP module, and concat indicates concatenation processing.

7. The method for generating denoised video based on multi-condition control according to claim 1, wherein: The video generation model includes a target module and a U-net denoising module. The target module is used to convert the motion mark of the target object into the target motion feature representation, and the U-net denoising module is used to obtain the predicted noise.

8. The method for generating denoised video based on multi-condition control according to claim 7, wherein: The U-net denoising module includes a downsampling stage, an intermediate stage, and an upsampling stage, each stage includes at least one block, and each block includes a convolutional layer, a spatial layer, and a temporal layer connected in sequence; the target module includes several blocks, each block includes at least one convolutional layer, and the total number of blocks in the target module is equal to the sum of the total number of blocks in the downsampling stage and the intermediate stage in the U-net denoising module.

9. The method for generating denoised video based on multi-condition control according to claim 8, wherein: The conditional information representation and camera motion labeling are input into each block of each stage of the U-net denoising module; The output of each block in the target module is input to each block in the downsampling stage and the intermediate stage of the U-net denoising module.

10. The method for generating denoised video based on multi-condition control according to claim 1, wherein: In step S5, the latent space representation of the video is obtained by subtracting the latent space representation of Gaussian noise from the predicted noise.

Images