16 results about "Diffusion loss" patented technology

The invention relates to the technical field of vision and artificial intelligence, and discloses a point cloud completion method and device based on text prompt, equipment and a storage medium, and the method comprises the steps: generating the diffusion loss of a diffusion model according to a first loss model, and generating the completion loss of a completion model according to a second loss model; according to the diffusion loss of the diffusion model, the completion loss of the completion model and the total loss model, generating the total loss of the overall model, updating the model parameters of the overall model, and storing the updated overall model; determining text semantic features of a preset object based on text prompt information corresponding to the target object, determining fusion features of the target object based on a plurality of non-overlapping local areas in the current incomplete point cloud and the text semantic features of the preset object, and updating the overall model based on the fusion features of the target object. And complementing the current incomplete point cloud to generate a complete point cloud of the target object. According to the invention, the complementation efficiency of the complete point cloud of the target object can be improved.

The application discloses a target three-dimensional temperature field prediction method based on a generated diffusion model. The method steps are as follows: a simplified geometric model of the target is established, and key parameters such as target material, power system and meteorology are determined; an outer boundary heat balance equation of the target is established, and the surface temperature field under different working conditions is calculated as basic data; a multi-modal feature extraction module is established to perform deep feature extraction and fusion on three-dimensional coordinates, node temperatures and key parameters; a generated diffusion model is established to train the fused features, learn the mapping from feature noise to real temperature features; according to input parameters and three-dimensional shape constraints, the predicted node temperature value is output through a reverse denoising process, and the diffusion loss is calculated, and finally a complete three-dimensional temperature field is generated. The application effectively solves the problem of low efficiency of traditional simulation and realizes fast and accurate prediction of the target temperature field.

A condition controllable image sample expansion method and system for surface defect detection, wherein the method comprises: collecting multiple negative sample and positive sample images, and constructing an image sample expansion model; inputting a template and an abnormal type into a condition coding module to obtain an output condition; an encoder compresses the negative sample image to a latent feature space, inputs the obtained latent space vector into a diffusion model and a decoder, combines the output condition, and obtains a reconstructed negative sample image; a total loss function and a diffusion loss function are constructed, and the two loss functions are minimized until a preset stopping condition is reached, and the model weight is adjusted during the training process; the model is deployed to a device end, a positive sample image is input, a controllable condition is designed, and corresponding samples are generated to realize image sample expansion. The present application realizes controllable region shape, size, position and type, can generate new samples according to the controllable condition, and provides a data basis for subsequent modeling.

The invention provides an image generation method and system based on a diffusion personalized generation model, and a storage medium, and the method comprises the following steps: obtaining an adversarial disturbance image, building a diffusion personalized generation model, taking the adversarial disturbance image as the input of the diffusion personalized generation model, and generating an adversarial disturbance image; acquiring potential features of the anti-disturbance image according to an auto-encoder of the diffusion personalized generation model; performing calculation based on a first preset rule according to the potential features; calculating based on a second preset rule according to random noise of the diffusion personalized generation model; performing denoising processing on the potential features according to the alignment loss and the potential diffusion loss; and cutting the de-noised features and decoding the de-noised features through a decoder which diffuses the personalized generation model. Compared with the prior art, the method has the advantages that the damaged mapping relation between the confrontation disturbance image and the potential space is repaired through the alignment loss, the potential diffusion loss and the dynamic weight changing along with the time step, and the generation quality of the image is improved.

The application relates to the visual technology field and the artificial intelligence technology field, and discloses a point cloud completion method and device based on a text prompt, equipment and a storage medium. The method comprises the following steps: generating a diffusion loss of a diffusion model according to a first loss model, and generating a completion loss of a completion model according to a second loss model; generating a total loss of an overall model according to the diffusion loss of the diffusion model, the completion loss of the completion model and a total loss model, updating model parameters of the overall model, and saving the updated overall model; determining text semantic features of a preset object based on text prompt information corresponding to a target object, determining fusion features of the target object based on a plurality of non-overlapping local regions in a current incomplete point cloud and the text semantic features of the preset object, and completing the current incomplete point cloud based on the fusion features of the target object by using the updated overall model, so as to generate a complete point cloud of the target object. The application can improve the completion efficiency of the complete point cloud of the target object.

The invention relates to the technical field of voice processing of financial and medical scenes, and discloses a voice synthesis method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring a text sequence and a corresponding voice sequence; inputting the voice sequence into a diffusion model for denoising, generating a middle hidden state, and outputting and calculating diffusion loss; introducing CTC (connectivity-dominant time classification) loss as first alignment loss in the middle layer of the model, and calculating the first alignment loss based on the text sequence and the middle hidden state; calculating a second alignment loss based on the speech sequence and an output of the diffusion model; the total loss is calculated by combining the three losses so as to update parameters of the diffusion model, and the text sequence to be synthesized is processed through the updated diffusion model and voice output is obtained. According to the method, the CTC loss and the second alignment loss are introduced as explicit guidance, a learning direction is provided for the model at the initial stage of training, the model convergence process is remarkably accelerated, and computing resources needed for achieving the target performance are reduced.

The application relates to the technical field of intelligent vehicle control, and discloses an automobile active suspension control method based on reinforcement learning and diffusion strategy, which comprises the following steps: performing asymmetric random mask processing on an input suspension state vector and a road elevation observation vector respectively, projecting the masked inputs to a unified latent space, and forming a joint representation based on a policy network; based on the joint representation, performing reverse denoising by using DDIM to generate a diffusion trajectory, and sampling multiple candidate actions in the later stage of the diffusion trajectory; conservatively evaluating the candidate actions by using an integrated value network, and obtaining a final control instruction by weighted fusion through softmax; in the training process, an experience replay buffer and an expert buffer are maintained, an end-to-end online training is performed by combining reinforcement learning loss, diffusion loss and reconstruction loss. The application can stably and efficiently train the diffusion control strategy, and significantly improves the adaptive control capability and ride comfort of the active suspension under complex road conditions.

Embodiments of the present specification provide a target large language model training method and device, equipment and medium, wherein the method comprises: obtaining a mask sequence set; the proportion of masked word units in each mask sequence of the mask sequence set is positively correlated with the time step; inputting each mask sequence into the target large language model to obtain the prediction distribution of the masked word units in each mask sequence; determining a reweighting factor of each mask sequence based on the time step of each mask sequence; the reweighting factor monotonically decreases with the increase of the time step within a preset time step interval; calculating a basic mask diffusion loss of each mask sequence based on the prediction distribution; calculating a target loss value based on the reweighting factor of each mask sequence and the basic mask diffusion loss of each mask sequence; and training the target large language model based on the target loss value.

The application belongs to the technical field of optical fiber materials, and particularly relates to a preparation method of a PMN-PT ferroelectric crystal core glass cladding composite optical fiber. The application pre-increases the PMN-PT concentration in the cladding glass, that is, PMN-PT crystals are pre-filled in the cladding glass, diffusion reaction is carried out under certain high-temperature conditions, ions in the PMN-PT crystals are pre-diffused into the inner hole of the cladding glass, the concentration gradient between the PMN-PT ferroelectric crystal core and the cladding glass is reduced, the further diffusion of the fiber core crystal to the cladding glass can be effectively inhibited, and the diffusion loss of the fiber core is correspondingly reduced. The pre-diffusion layer is combined with the cladding glass closely, the pre-diffusion layer can isolate the fiber core and the cladding glass, inhibit the diffusion reaction from occurring in the subsequent heat drawing and single crystallization process, protect the material form and functional integrity of the fiber core, ensure that the material composition of the fiber core is unchanged, and make the composite optical fiber maintain the original function.

The application discloses an anti-voice conversion watermark embedding method and program product based on a diffusion model, which comprises inputting real-time carrier voice into a trained anti-voice conversion watermark embedding network based on a diffusion model to generate watermark-containing voice; the training method of the anti-voice conversion watermark embedding network comprises: preprocessing received historical carrier voice to generate logarithmic mel spectrum and diffusion prior, and sending the historical carrier voice into the anti-voice conversion watermark embedding network together with the logarithmic mel spectrum and the diffusion prior to generate watermark-containing voice; the model parameters of the anti-voice conversion watermark embedding network are constantly updated with the minimum total loss function as the target, the total loss function comprises an adversarial loss function between the conversion results of the carrier voice and the watermark-containing voice, a diffusion loss function between the predicted noise and the actual added noise, and a frequency weighting loss function of the watermark-containing voice and the carrier voice under multi-scale short-time Fourier transform. The application can guarantee the imperceptibility of the watermark, improve the anti-voice conversion performance and robustness of the watermark-containing voice.

The invention discloses a target three-dimensional temperature field prediction method based on a generated diffusion model. The method comprises the following steps: establishing a simplified geometric model of a target, and determining key parameters such as a target material, a power system and weather; establishing a target outer boundary heat balance equation, and respectively calculating surface temperature fields under different working conditions as basic data; establishing a multi-modal feature extraction module, and performing deep feature extraction and fusion on the three-dimensional coordinates, the node temperature and the key parameters; establishing a generation diffusion model to train fused features, and learning mapping from feature noise to real temperature features; according to the input parameters and the three-dimensional shape constraint, a predicted node temperature value is output through a reverse denoising process, diffusion loss is calculated, and finally a complete three-dimensional temperature field is generated. According to the invention, the problem of low efficiency of traditional simulation is effectively solved, and rapid and accurate prediction of the target temperature field is realized.

The application discloses a face makeup transfer method based on an improved Diffusion model, comprising: preliminary data processing, obtaining a pseudo makeup dataset, the dataset is made by using the public MT dataset, the LADN dataset and the Wild-MT dataset by using histogram matching combined with a thin plate spline interpolation method; training the improved Diffusion model by using the obtained dataset to obtain an optimal makeup transfer model; the loss function of model training comprises a latent diffusion loss, a makeup loss and a diffusion reconstruction loss, the latent diffusion loss is used for guiding the model to learn to gradually restore a clear face makeup transfer image from a noise image, the makeup loss is used for optimizing the makeup effect of the generated image, and the diffusion reconstruction loss is used for keeping the consistency of the image; the optimal makeup transfer model is applied to process input source images and reference images to generate makeup transfer results with natural transition. The application can realize a high-quality and controllable makeup transfer effect while keeping the identity features of the source images.

The invention discloses an image generation method and system based on diffusion model time step pruning, and the method comprises the steps: firstly calculating the denoising loss of each image time step, obtaining a diffusion loss difference, and then calculating the loss change of continuous time steps through the diffusion loss difference; carrying out space average pooling on the predicted noise maps of the batch image samples and obtaining image feature representation; according to the method, the generation quality is kept while the diffusion model is efficiently compressed by adopting the representation perception structured pruning framework, and the function of dynamically adjusting the time step screening strategy according to the staged characteristics of the diffusion process by adopting the staged time step selection mechanism is also realized; according to the method, the most representative key time step in the stage can be focused, redundancy can be reduced, efficiency can be improved, meanwhile, after the time step rich in information amount is selected, a first-order Taylor expansion method based on cumulative gradient is adopted to calculate the importance score of pruning, gradient information under key signals can be accumulated, and the method is suitable for being widely popularized and used.

The application discloses an image generation method and system based on diffusion model time step pruning, first calculates the denoising loss of each image time step and obtains the diffusion loss difference, then calculates the continuous time step loss change by using the diffusion loss difference, and then performs spatial average pooling on the predicted noise map of a batch of image samples to obtain image feature representation; the application realizes efficient compression of the diffusion model by using the representation perception structured pruning framework while maintaining the generation quality, and also realizes the function of dynamically adjusting the time step screening strategy according to the phased characteristics of the diffusion process by using the phased time step selection mechanism, which not only enables focusing on the most representative key time step within the stage, but also reduces redundancy and improves efficiency, and the importance score of pruning calculated by using the first-order Taylor expansion method based on the cumulative gradient after the information-rich time step is selected can accumulate the gradient information under the key signal, so that the application is suitable for being widely promoted and used.

A system and method are disclosed. The method includes rendering an image using a three-dimensional (3D) Gaussian splatting process; processing the rendered image with a pretrained latent diffusion model to estimate noise in a latent space; generating a diffusion loss based on a difference between the estimated noise and a sampled noise; periodically applying the diffusion loss to update parameters of the 3D Gaussian splatting process; and generating a novel view synthesis image based on the updated parameters.