AIGC image content intelligent generation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an AIGC image content intelligent generation method and system based on artificial intelligence, and the method comprises the steps: obtaining an image generation instruction; generating a first feature representation through an image generation model based on the image generation instruction; identifying and separating second feature information of which the feature value is lower than a preset half-precision floating-point number stable representation threshold value from the first feature representation; performing numerical value amplification processing on the second feature information, and encoding the processed second feature information into third feature representation used for representing image detail information; and integrating the third feature representation with the first feature representation to obtain a fourth feature representation, and generating a target image according to the fourth feature representation, thereby effectively solving the problem of detail information loss caused by numerical value underflow in an image generation model in a semi-precision floating-point number calculation environment, and improving the accuracy of the image generation model by performing identification, separation and amplification processing on low-feature-value information. And the detail richness and the overall quality of the generated image are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to an AIGC image content intelligent generation method and system based on artificial intelligence. Background Technology

[0002] In the digital creative industry, AIGC (AI-Generated Content) image content intelligent generation systems are widely used to quickly generate images based on text instructions. However, as model size increases and the demands for generation detail grow, the fine numerical processing within large and complex generation models faces challenges when computed using half-precision floating-point numbers (FP16). To run large image generation models with limited hardware resources and meet the requirements of high throughput and low latency, the industry commonly uses half-precision floating-point numbers for computation. Compared to full-precision floating-point numbers (FP32), FP16 halves the data bit width, allowing for the loading of larger models or the processing of more batches of data, and leverages the tensor cores of modern GPUs to accelerate computation. This optimization is crucial for scenarios requiring fast response and large-scale content generation.

[0003] However, these large-scale models have complex internal structures, including multiple layers of nonlinear transformations, intricate attention mechanisms, and large-scale parameter matrices. During key computational stages such as image style feature extraction and cross-referencing textual semantics with image content, the model generates a large number of intermediate computational results with an extremely wide dynamic range. For example, when capturing subtle textures, light and shadow transitions, or artistic brushstroke details, the corresponding activation values ​​or gradients may be very small, close to the lower limit of floating-point representation; while when identifying main objects, strong edge features, or processing high-weight subject concepts, the corresponding values ​​may be very large. This broad numerical dynamic range is an inherent characteristic of high-fidelity, high-detail-capturing models.

[0004] When using half-precision floating-point numbers for inference calculations, their limited numerical representation range becomes a significant constraint. The effective number of bits and the exponent range of the FP16 format are both smaller than those of FP32. For extremely small values ​​generated during model calculations, if they fall below the smallest non-zero positive number that FP16 can represent, they are forced to zero, resulting in the complete loss of fine-grained information—a phenomenon known as underflow. For example, when depicting an image in the style of an oil painting, information encoded by minute differences in brushstrokes, subtle gradations of light, or the graininess of materials—is lost after underflow occurs. Conversely, for extremely large values ​​generated during model calculations, if they exceed the maximum finite value that FP16 can represent, overflow occurs. These values ​​are typically represented as infinity or non-numeric. This overflow also undermines the validity of the numerical value, distorting the image or semantic information it represents. This mismatch between the limited numerical range of half-precision floating-point numbers and the inherently wide dynamic range of intermediate calculation results in complex models is the direct cause of information loss and numerical inaccuracy.

[0005] This numerical imprecision directly leads to a decline in the quality of generated images. When processing image style features, underflow causes the loss of minute values ​​representing fine information such as brushstrokes, lighting, and texture. The model cannot maintain global consistency when reconstructing image style, resulting in style transfer distortion and making the generated images appear fragmented and inconsistent in overall artistic style. Simultaneously, numerical overflow and underflow phenomena severely interfere with the model's accurate parsing of the association between textual instructions and image content, leading to poor cross-modal semantic consistency and a significant deviation between the generated image content and the user's textual intent.

[0006] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this application provides an AIGC image content intelligent generation method and system based on artificial intelligence.

[0008] Firstly, this application provides an AIGC image content intelligent generation method based on artificial intelligence, including: Obtain image generation instructions; Based on the image generation instructions, the first feature representation is generated through the image generation model; Identify and separate second feature information from the first feature representation whose feature value is lower than a preset half-precision floating-point stable representation threshold; The second feature information is numerically amplified, and the processed second feature information is encoded into a third feature representation used to characterize image detail information; The third feature representation is integrated with the first feature representation to obtain the fourth feature representation. The target image is then generated based on the fourth feature representation.

[0009] Through the above scheme, this application can effectively solve the problem of loss of detail information caused by numerical underflow in the image generation model under half-precision floating-point calculation environment. By identifying, separating and amplifying low feature value information and encoding it into independent detail feature representation, and finally integrating it with the original features, the detail richness and overall quality of the generated image are significantly improved, overcoming the shortcomings of style transfer distortion and poor cross-modal semantic consistency in the prior art.

[0010] Furthermore, the step of generating the first feature representation through the image generation model based on the image generation instructions includes: The image generation instructions are converted into a digital representation by a text encoder, and a first feature representation, including the internal feature representation of the image, is generated based on the digital representation by an initial image feature extraction network.

[0011] Through the above scheme, this application can ensure that the image generation instructions are accurately converted into a digital form that the model can understand, and on this basis, extract comprehensive initial image features, laying a solid foundation for subsequent detail enhancement processing, thereby improving the accuracy and efficiency of the image generation process.

[0012] Further, the steps of numerically amplifying the second feature information and encoding the processed second feature information into a third feature representation for characterizing image detail information include: Calculate the difference between each feature information in the second feature information and its local statistical value. Based on the difference and the preset amplification factor, amplify the feature value corresponding to the feature information so as to increase the numerical range of the second feature information to the level of the preset half-precision floating-point stable representation threshold. The processed feature information is transformed and compressed by an encoding neural network to generate a third feature representation.

[0013] Through the above scheme, this application can precisely amplify low-value features to a range that can be stably represented by half-precision floating-point numbers, and efficiently transform them into compact and expressive detailed feature representations through an encoding neural network, thereby effectively restoring and enhancing the fine details of the image, avoiding noise that may be introduced by blind amplification, and improving the effectiveness of detailed information.

[0014] Furthermore, before performing numerical amplification processing on the second feature information, the following is included: Calculate the detail confidence score of each feature in the second feature information; Numerical amplification is performed only on feature information whose detail confidence score exceeds a first preset threshold.

[0015] Through the above scheme, this application can intelligently select feature information with real detail value for amplification processing, avoid ineffective processing of low confidence or noise information, thereby improving processing efficiency and the purity of the generated image, and ensuring the effectiveness and accuracy of detail enhancement.

[0016] Furthermore, the detail confidence score is calculated based on at least one or more of the following factors: the consistency of local visual features of the feature information, the structural correlation of the feature information in feature maps at different levels of the image generation model, and the semantic correlation between the feature information and the image generation instructions.

[0017] Through the above scheme, this application can comprehensively evaluate the confidence level of feature information from multiple dimensions, ensuring that the amplified details are not only visually consistent and related in the internal structure of the model, but also highly matched with the user's intent, thereby comprehensively improving the accuracy and effectiveness of detail enhancement.

[0018] Furthermore, the local visual feature consistency of the feature information is calculated through the following steps: Multidimensional feature extraction is performed on the local region surrounding the feature information; Based on the extracted multidimensional features, a local structure description pattern is constructed. Compare the local structure description pattern with the preset real structure pattern library, and calculate the degree of matching between the local structure description pattern and the real structure pattern in the preset real structure pattern library; A local visual feature consistency score is generated based on the degree of matching between the local structure description pattern and the real structure patterns in the preset real structure pattern library.

[0019] Through the above scheme, this application can accurately assess the visual coherence and realism of feature information in local areas, effectively identify details with high visual consistency, and thus ensure that the enhanced details are more natural and realistic visually.

[0020] Furthermore, the structural correlation of feature information in feature maps at different levels of the image generation model is calculated through the following steps: Extracting feature information from feature maps at different levels of an image generation model, identifying structural clues present in the feature maps at different levels; The structural cues and feature information present in feature maps at different levels are spatially aligned, and the structural correlation score between the structural cues and feature information present in feature maps at different levels is calculated.

[0021] Through the above approach, this application can deeply explore the structural connections of feature information at different abstraction levels within the model, ensuring that the enhanced details have stable structural support throughout the image generation process, thereby improving the robustness of the details and the overall structural rationality of the image.

[0022] Furthermore, the semantic correlation between feature information and image generation instructions is calculated through the following steps: The image generation instructions are semantically encoded, textual semantic features are extracted, and the similarity score between the feature information and the textual semantic features is calculated to obtain the semantic relevance score.

[0023] By employing the above approach, this application can ensure that the enhanced details are highly semantically consistent with the user's text instructions, avoiding the generation of details that do not conform to the user's intent, thereby significantly improving the semantic accuracy of the generated image and user satisfaction.

[0024] Furthermore, the step of integrating the third feature representation with the first feature representation to obtain the fourth feature representation includes: The third feature representation and the first feature representation are input into a preset fusion weight calculation network to obtain the output fusion weights; The third feature representation is multiplied element-wise with the fusion weights, and then added element-wise with the first feature representation to obtain the fourth feature representation.

[0025] Through the above scheme, this application can intelligently balance the contributions of detailed features and original features by using an adaptive fusion weight calculation network. This ensures that when detailed information is integrated into the overall features, it can effectively enhance the image quality without destroying the integrity of the original features, thereby achieving a more refined and natural image generation effect.

[0026] Secondly, this application also provides an AIGC image content intelligent generation system based on artificial intelligence, used to perform the above method, including: The instruction acquisition module is used to acquire image generation instructions; The first feature representation generation module is used to generate a first feature representation based on the image generation instructions and through the image generation model. The second feature information recognition and separation module is used to identify and separate second feature information from the first feature representation whose feature value is lower than the preset half-precision floating-point stable representation threshold; The third feature representation generation module is used to perform numerical amplification processing on the second feature information and encode the processed second feature information into a third feature representation used to characterize image detail information. The fourth feature representation and target image generation module is used to integrate the third feature representation with the first feature representation to obtain the fourth feature representation, and generate the target image based on the fourth feature representation.

[0027] This application discloses an AIGC image content intelligent generation method based on artificial intelligence. It obtains image generation instructions and generates a first feature representation based on these instructions using an image generation model. Addressing the potential numerical underflow problem in half-precision floating-point computation environments, it innovatively identifies and separates second feature information from the first feature representation, with feature values ​​below a preset half-precision floating-point stable representation threshold. This effectively solves the problems of image detail loss, style transfer distortion, and poor cross-modal semantic consistency caused by numerical underflow in existing large-scale AIGC models using half-precision floating-point computation. By identifying and amplifying low feature value information, this application can restore and enhance details in the image that might otherwise be lost due to underflow, such as fine textures, light and shadow transitions, and artistic brushstrokes, significantly improving the detail richness and visual quality of the generated image. Simultaneously, by encoding the detail information into an independent third feature representation and integrating it with the original features, it ensures the effective integration of details and avoids damage to the overall image structure. This application overcomes the mismatch between the limited numerical range of half-precision floating-point numbers and the wide dynamic range of intermediate calculation results of complex models in the prior art, thereby significantly improving the generation quality of AIGC image content and user experience while ensuring computational efficiency. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating an AIGC image content intelligent generation method based on artificial intelligence, provided in an embodiment of this application.

[0029] Figure 2 This is a schematic diagram of the structure of an AIGC image content intelligent generation system based on artificial intelligence, provided in an embodiment of this application.

[0030] Labeling explanation: 210, instruction acquisition module; 220, first feature representation generation module; 230, second feature information recognition and separation module; 240, third feature representation generation module; 250, fourth feature representation and target image generation module. Detailed Implementation

[0031] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0032] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0033] Traditional AIGC image content intelligent generation systems, when using half-precision floating-point numbers for calculations, suffer from limited numerical expression ranges. This causes extremely small numerical information generated within the model, such as subtle textures and lighting transitions in the image, to easily underflow to zero, while extremely large numerical information may overflow. This results in the loss of fine details and numerical inaccuracies, directly leading to stylistic distortion in the generated images and a deterioration in the semantic relationship between the image content and the text instructions.

[0034] Regarding this, firstly, see... Figure 1 This application proposes an AIGC image content intelligent generation method based on artificial intelligence, including: S1. Obtain image generation instructions; S2. Based on the image generation instructions, generate the first feature representation through the image generation model; S3. Identify and separate the second feature information from the first feature representation whose feature value is lower than the preset half-precision floating-point stable representation threshold; S4. Perform numerical amplification on the second feature information and encode the processed second feature information into a third feature representation used to characterize image detail information; S5. Integrate the third feature representation with the first feature representation to obtain the fourth feature representation, and generate the target image based on the fourth feature representation.

[0035] This application identifies, separates, and numerically amplifies minute feature information in the image generation model that is susceptible to half-precision floating-point numbers, encodes it into independent detail feature representations, and then integrates it with the original feature representations. This effectively avoids the information underflow and overflow problems caused by half-precision floating-point number calculations, and significantly improves the detail representation and overall quality of the generated image.

[0036] Image generation instructions refer to textual descriptions or semantic commands input by the user to guide the image generation model in creating specific image content. For example, a user could input a golden Labrador running at sunset as the image generation instruction. An image generation model is a deep learning-based model, such as a diffusion model or generative adversarial network, capable of generating images based on input instructions or feature representations. The first feature representation is an intermediate feature vector or feature map initially generated by the image generation model based on the image generation instructions, containing the main content and style information of the image. The half-precision floating-point stable representation threshold refers to the smallest non-zero positive value that can be stably and accurately represented in a half-precision floating-point (FP16) computing environment. Values ​​below this threshold are prone to underflow in FP16 calculations, leading to information loss. Second feature information refers to feature information identified from the first feature representation whose feature values ​​are below the half-precision floating-point stable representation threshold. This information typically represents fine details in the image. Numerical amplification processing refers to numerically boosting the second feature information so that its feature value reaches or exceeds the half-precision floating-point stable representation threshold, thereby preventing underflow in subsequent FP16 calculations. The third feature representation is the second feature information after numerical amplification, and it is encoded into a feature vector or feature map specifically used to characterize image detail information. The fourth feature representation is the final feature representation obtained by integrating the third feature representation with the first feature representation, and it contains complete image content, style, and fine detail information. The target image refers to the high-quality image ultimately generated by the image generation model based on the fourth feature representation.

[0037] The core of the AIGC image content intelligent generation method proposed in this application lies in the special processing of detailed information that is susceptible to the precision limitations of half-precision floating-point numbers during image generation, ensuring that these critical details are not lost during computation. First, image generation instructions need to be obtained. These instructions can be text descriptions input through a user interface, such as a unicorn walking in a forest, or other forms of semantic input. The instructions can be obtained directly from the user input interface or read from a pre-defined task queue. Next, based on the obtained image generation instructions, a first feature representation is generated through an image generation model. The image generation model can be any advanced AIGC model architecture, such as a Transformer-based diffusion model or a generative adversarial network. In one implementation, the image generation model can directly take the image generation instructions as input and generate a feature representation containing the main content and style information of the image through its internal encoder and feature extraction layer. For example, the model may include a text encoder that converts the text instructions into a numerical representation, and then, through an initial image feature extraction network, generate a first feature representation based on this numerical representation. The first feature representation is typically a high-dimensional vector or a multi-channel feature map that abstractly describes the overall structure and semantics of the target image.

[0038] Subsequently, second feature information whose feature values ​​are lower than a preset half-precision floating-point stable representation threshold is identified and separated from the first feature representation. Within the image generation model, especially when using half-precision floating-point numbers for computation, many feature values ​​representing fine image details, such as texture, lighting, and microstructures, are often very small, easily falling below the minimum representable non-zero value of FP16, thus being truncated to underflow and zero during computation. To identify these potentially lost details, all feature values ​​in the first feature representation can be traversed and compared with the preset half-precision floating-point stable representation threshold. Any feature value lower than this threshold and its corresponding feature information are marked as second feature information and separated.

[0039] After identifying and separating the second feature information, it needs to be numerically amplified and encoded into a third feature representation for characterizing image details. The purpose of numerical amplification is to increase these small feature values ​​to a range that can be stably represented by half-precision floating-point numbers, thus avoiding information loss. The difference between each feature in the second feature information and its local statistical value, such as the local mean or median, can be calculated. Then, based on these differences and a preset amplification factor, the feature values ​​are amplified. In this way, relatively small feature values ​​can be selectively amplified, increasing their numerical range to a preset threshold for stable half-precision floating-point representation. After numerical amplification, the processed second feature information needs to be encoded into a third feature representation. This can be achieved using a dedicated encoding neural network that performs feature transformation and compression on the amplified feature information to generate a detailed feature representation. For example, this encoding neural network could be a small convolutional neural network or a multilayer perceptron, whose function is to transform the amplified values ​​into more semantically meaningful detailed feature vectors.

[0040] Finally, the third feature representation is integrated with the first feature representation to obtain the fourth feature representation. The target image is then generated based on this fourth feature representation. The purpose of integration is to reintegrate the specially processed detail information into the main image feature representation, forming a complete and high-precision feature representation. One integration method is to input the third and first feature representations into a pre-defined fusion weight calculation network, which outputs fusion weights. Then, the third feature representation is multiplied element-wise with the fusion weights, and then added element-wise with the first feature representation to obtain the fourth feature representation. This weighted fusion method ensures that detail information is integrated into the overall features with appropriate strength, avoiding over-emphasis or under-emphasis. After obtaining the fourth feature representation, the image generation model, such as the decoder part of a diffusion model or the generator of a GAN, will generate the final target image based on this fourth feature representation containing rich detail information.

[0041] The AIGC image content intelligent generation method proposed in this application works by accurately identifying and effectively processing the fine details that image generation models are prone to lose in half-precision floating-point calculations. Traditional methods, when processing image generation instructions, directly generate a first feature representation through the image generation model and then generate the target image based on this representation. However, when the model generates a large amount of extremely small feature information, this information is very prone to underflow to zero in half-precision floating-point calculations, leading to the loss of image details.

[0042] After generating the first feature representation, this application does not directly use it for image generation. Instead, it further identifies and separates second feature information from the first feature representation, those feature values ​​being lower than a preset half-precision floating-point stable representation threshold. These second feature information are precisely the image details that are easily ignored or lost in traditional methods. Subsequently, these second feature information undergoes numerical amplification processing, increasing their numerical range to a level that can be stably represented by half-precision floating-point numbers, effectively avoiding underflow. The amplified second feature information is encoded into a third feature representation, specifically used to characterize the image's detail information. Integrating this detail-rich third feature representation with the original first feature representation yields a more complete and accurate fourth feature representation. Based on this fourth feature representation, the image generation model can generate target images with higher detail fidelity and stronger semantic consistency.

[0043] Through the above scheme, this application ensures that key details such as subtle textures, light and shadow transitions, and artistic brushstrokes of images can be effectively preserved and accurately expressed in a half-precision floating-point computing environment, thereby significantly improving the generation quality and user experience of AIGC images. The AIGC image content intelligent generation method based on artificial intelligence proposed in this application aims to solve the problem of image detail loss and generation quality degradation caused by numerical underflow and overflow in existing AIGC systems under a half-precision floating-point computing environment. Traditional methods, when using half-precision floating-point numbers for calculation, often sacrifice numerical precision in pursuit of computational efficiency and model size, causing tiny feature values ​​generated within the model to be truncated to zero. This results in poor performance in terms of fine textures and light and shadow transitions in the generated images, and even stylistic distortion and semantic deviation.

[0044] Unlike existing technologies that directly use half-precision floating-point numbers for end-to-end computation, this application, after generating a preliminary feature representation, can proactively identify fragile details—the second feature information—with feature values ​​below the stable representation threshold of half-precision floating-point numbers. By selectively amplifying these second feature information values ​​and encoding them into independent third feature representations, this application effectively avoids the loss of these critical details due to underflow in subsequent half-precision floating-point calculations. The enhanced detail features are then integrated with the original features to form a fourth feature representation containing richer and more accurate information, thereby guiding the image generation model to generate high-quality target images.

[0045] This method offers significant advantages over existing technologies. While maintaining the high efficiency and low resource consumption of half-precision floating-point computation, it greatly enhances the detail representation of the generated images, solving the problems of blurred image details and inconsistent styles in traditional methods. Through precise processing of detail information, this application can better capture and express the subtle semantics in image generation instructions, thereby improving the semantic consistency between the generated image and the user's intent. For example, when generating images with complex textures or subtle lighting effects, this application can more faithfully reproduce these details, making the generated images more artistic and realistic. Therefore, this application significantly improves the generation quality of AIGC image content and the user experience while ensuring computational efficiency.

[0046] Furthermore, the step of generating the first feature representation through the image generation model based on the image generation instructions includes: The image generation instructions are converted into a digital representation by a text encoder, and a first feature representation, including the internal feature representation of the image, is generated based on the digital representation by an initial image feature extraction network.

[0047] The image generation instructions are typically given in the form of natural language text, such as a cat running in a grassy field. A text encoder is configured to receive these natural language instructions and convert them into a machine-understandable digital representation. This digital representation can be a high-dimensional vector, embedding, or other form of numerical data, capturing the semantic information and key concepts of the instructions. Further, an initial image feature extraction network receives the digital representation generated by the text encoder and aims to transform this semantic information into a preliminary image feature representation, or first feature representation. The first feature representation includes not only the overall structure and layout information of the image but also representations of internal image features, such as macroscopic and microscopic visual elements like the main object, background, color, and texture. The initial image feature extraction network can employ various deep learning architectures, such as convolutional neural networks, Transformers, or variants thereof, to efficiently extract and construct these image features from the digital representation.

[0048] This application achieves an efficient transformation from abstract image generation instructions to concrete image feature representations by introducing a text encoder and an initial image feature extraction network. The text encoder is responsible for converting human-readable natural language instructions into computer-processable semantic vectors, ensuring accurate understanding of the instructions. Subsequently, the initial image feature extraction network uses these semantic vectors as input and, through its complex internal structure and learned mapping relationships, gradually constructs a first feature representation containing rich visual information. This process ensures that the generated first feature representation accurately reflects the intent of the original instruction, laying a solid foundation for subsequent image generation processes.

[0049] The above technical solutions enable accurate semantic understanding and efficient feature extraction of image generation instructions, thereby generating an initial image feature representation that is highly consistent with the user's intent. This helps improve the ability to control details and the overall image quality in the subsequent image generation process, ensuring that the generated target image has higher accuracy and richness in both semantics and visuals.

[0050] Further, the steps of numerically amplifying the second feature information and encoding the processed second feature information into a third feature representation for characterizing image detail information include: Calculate the difference between each feature information in the second feature information and its local statistical value. Based on the difference and the preset amplification factor, amplify the feature value corresponding to the feature information so as to increase the numerical range of the second feature information to the level of the preset half-precision floating-point stable representation threshold. The processed feature information is transformed and compressed by an encoding neural network to generate a third feature representation.

[0051] The second feature information refers to the feature information identified and separated from the first feature representation, whose feature values ​​are lower than a preset half-precision floating-point stable representation threshold. This feature information typically carries fine details of the image, but is easily lost or truncated in low-precision floating-point representation. To effectively preserve these details, numerical amplification is required. Specifically, the purpose of numerical amplification is to increase the numerical range of the second feature information to the level of the preset half-precision floating-point stable representation threshold, thereby ensuring that it can be accurately represented and utilized in subsequent processing and storage. This can be achieved by calculating the difference between each feature information and its local statistical values, such as local mean, local median, or local maximum. Based on this difference, and combined with a preset amplification factor, the feature values ​​corresponding to the feature information can be amplified in a targeted manner. For example, the larger the difference, the more likely the feature information is to represent important details, and therefore a larger amplification factor can be applied.

[0052] After numerical upscaling, the processed feature information needs to be encoded into a third feature representation to characterize image details. An encoding neural network is designed to perform feature transformation and compression on the processed feature information, thereby reducing data redundancy while preserving key details. This neural network can employ various combinations of structures such as convolutional layers, pooling layers, and fully connected layers to learn and extract the most compact representation that best represents image details. In this way, even if the original second feature information has low values, after upscaling and encoding, a stable and expressive third feature representation can be formed, supporting the final generation of a high-quality target image.

[0053] This application identifies feature points with significant variations or uniqueness within local regions by calculating the difference between the second feature information and its local statistical values. These feature points often correspond to fine textures, edges, or minute structures in the image. Amplification based on this difference ensures that features truly representing details are enhanced, rather than blindly amplifying all low-value features, thus avoiding the introduction of unnecessary noise. The amplified feature information is then input into an encoding neural network for feature transformation and compression, enabling these enhanced details to be efficiently encoded into a compact and stable third feature representation. By learning the inherent structure and patterns of the data, the encoding neural network can transform high-dimensional, potentially noisy, amplified features into low-dimensional, semantically rich representations, thereby maximizing the preservation of image details within a limited representation space and effectively integrating them with the first feature representation to jointly guide the generation of the target image.

[0054] Through the above technical solution, this application effectively solves the problem of low feature value information being easily lost or lacking precision in half-precision floating-point representation in traditional methods. By using difference calculation and targeted numerical amplification, the numerical range of image detail information can be accurately improved, bringing it above the threshold for stable half-precision floating-point representation, thereby avoiding truncation or quantization errors of detail information. Simultaneously, the introduction of the encoding neural network ensures that this enhanced detail information can be efficiently and robustly encoded, forming a high-quality third feature representation. Therefore, this application can significantly improve the ability to capture and reproduce fine details during AIGC image generation, making the generated target image visually clearer, more realistic, and more layered, effectively improving the overall image quality and user experience.

[0055] Furthermore, before performing numerical amplification processing on the second feature information, the following is included: Calculate the detail confidence score of each feature in the second feature information; Numerical amplification is performed only on feature information whose detail confidence score exceeds a first preset threshold.

[0056] The detail confidence score is a metric used to quantify the contribution of each feature in the second feature information to the final image detail. This score can reflect factors such as local contrast, texture complexity, edge strength, or semantic relevance to the image generation instructions. For example, features with high local contrast or complex textures are typically assigned higher detail confidence scores, indicating they are more likely to represent important image details. The first preset threshold is a pre-defined value used to filter out features with sufficiently high detail confidence. Only when the detail confidence score of a feature exceeds this threshold is the feature considered a valid detail worthy of numerical upscaling. This threshold can be obtained empirically, trained using machine learning methods, or dynamically adjusted based on specific application scenarios and image quality requirements.

[0057] This application introduces a detail confidence assessment mechanism to intelligently filter second feature information before numerical amplification. When the detail confidence score of a feature in the second feature information exceeds a first preset threshold, it indicates that the feature is highly likely to represent important details in the image that need to be accurately preserved. Numerical amplification of this feature effectively increases its numerical range, ensuring that it will not lose precision due to the limitations of half-precision floating-point representation when subsequently encoded as a third feature representation, thus guaranteeing the integrity of image details. Conversely, feature information with a detail confidence score below the first preset threshold is considered to be noise or unimportant information, and is not numerically amplified to avoid amplifying noise and reduce unnecessary computational resource consumption.

[0058] Through the above technical solution, this application can more intelligently identify and process detailed information in images. By selectively amplifying high-confidence detailed features, it ensures the effective preservation and enhancement of key image details, resulting in a richer and more accurate representation of details in the final generated image. On the other hand, by avoiding the amplification of low-confidence features, it effectively suppresses the introduction and amplification of noise, thereby improving the overall quality and clarity of the generated image. Furthermore, this selective processing mechanism also helps reduce unnecessary computational burden and improves the efficiency of the entire image generation method.

[0059] Furthermore, the details confidence score is calculated based on at least one or more of the following factors: The consistency of local visual features of feature information, the structural correlation of feature information in feature maps at different levels of the image generation model, and the semantic correlation between feature information and image generation instructions.

[0060] Local visual feature consistency refers to the degree of similarity or coordination between the visual attributes of a feature and its surrounding local areas in terms of structure, texture, or color. High local visual feature consistency usually indicates that the feature is a stable and meaningful detail component in the image, rather than random noise. The structural correlation of feature information across different levels of feature maps in the image generation model can be understood as the inherent connection and consistency between the feature representation at different levels of abstraction during the image generation model's processing, such as from low-level edge texture to high-level semantic structure. Strong structural correlation means that the feature maintains its structural importance at different processing stages of the model and is an indispensable detail in the image generation process. The semantic correlation between feature information and image generation instructions refers to the degree of relevance or matching between the visual content represented by the feature and the user-input image generation instructions at the semantic level. By evaluating semantic correlation, details that are crucial to achieving the user's intent can be prioritized.

[0061] Since the second feature information with low eigenvalues ​​may contain a lot of noise or unimportant information, indiscriminately amplifying all low eigenvalue information may introduce artifacts or degrade image quality. By comprehensively considering these factors, we can more accurately select the second feature information that truly represents the key details of the image, thereby ensuring that subsequent numerical amplification processing can effectively improve the expressive power of image details, while avoiding the amplification of irrelevant noise and guaranteeing the visual quality and semantic accuracy of the generated image.

[0062] Furthermore, the local visual feature consistency of the feature information is calculated through the following steps: Multidimensional feature extraction is performed on the local region surrounding the feature information; Based on the extracted multidimensional features, a local structure description pattern is constructed. Compare the local structure description pattern with the preset real structure pattern library, and calculate the degree of matching between the local structure description pattern and the real structure pattern in the preset real structure pattern library; A local visual feature consistency score is generated based on the degree of matching between the local structure description pattern and the real structure patterns in the preset real structure pattern library.

[0063] Specifically, multidimensional feature extraction of the local region surrounding the feature information refers to the second feature information to be analyzed. Within its spatial neighborhood, various feature extractors, such as convolutional neural networks, wavelet transforms, or Gabor filters, are used to obtain multidimensional visual features including texture, edges, color gradients, and shape. These features can comprehensively describe the visual characteristics of the local region. Constructing a local structural description pattern based on the extracted multidimensional features can be understood as integrating and encoding the extracted multidimensional features to form a descriptor that represents the unique visual structure of the local region. For example, algorithms such as local binary patterns, scale-invariant feature transformations, or accelerated robust features can be used to transform complex visual features into compact and recognizable pattern representations.

[0064] In practical applications, the local structure description pattern is compared with a pre-set library of real structure patterns. The degree of matching between the local structure description pattern and the real structure patterns in the pre-set library is calculated to evaluate the similarity between the visual structure of the current local region and the known high-quality, real image structure. The library of real structure patterns can be pre-constructed by learning and modeling typical detailed structures in a large number of high-quality images. The degree of matching can be calculated using cosine similarity, Euclidean distance, or a learning-based metric.

[0065] Therefore, based on the degree of matching between the local structure description pattern and the real structure pattern in the preset real structure pattern library, a local visual feature consistency score is generated. This score quantifies the degree of conformity between the visual structure of the local area where the feature information is located and the real structure pattern. The higher the score, the more realistic and consistent the visual structure of the area is.

[0066] This application comprehensively captures the visual details of a region by extracting multidimensional features from the local area surrounding the feature information. By constructing a local structural description pattern, these complex visual features are abstracted into comparable structural representations. By comparing this local structural description pattern with a pre-set library of real structural patterns, the visual realism and structural rationality of the local area represented by the current feature information can be objectively evaluated. This pattern-matching mechanism enables the system to identify details that are visually more consistent with human perception habits and the physical laws of the real world, thus providing a reliable basis for subsequent detail confidence calculations. Through the above technical solution, the consistency of local visual features of each feature in the second feature information can be accurately quantified, helping to more accurately identify details in the image with high visual realism and structural integrity, avoiding misjudging low-quality or unnatural details as high-confidence details, thereby improving the targeting and effectiveness of subsequent numerical amplification processing, and ultimately contributing to the generation of AIGC image content with better visual effects and more realistic detail representation.

[0067] Furthermore, the structural correlation of feature information in feature maps at different levels of the image generation model is calculated through the following steps: Extracting feature information from feature maps at different levels of an image generation model, identifying structural clues present in the feature maps at different levels; The structural cues and feature information present in feature maps at different levels are spatially aligned, and the structural correlation score between the structural cues and feature information present in feature maps at different levels is calculated.

[0068] Image generation models typically consist of multiple layers, each responsible for extracting features at different levels of abstraction. Feature maps at different levels refer to the feature representations output by these layers, containing rich information ranging from low-level texture to high-level semantics. Structural cues can be understood as edges, corners, texture patterns, or more complex structured information related to the features to be analyzed within these different level feature maps. These cues reflect the morphology and organization of feature information at different scales. Extracting these structural cues can be achieved through convolution operations, pooling operations, attention mechanisms, or specific feature detection algorithms. For example, edge detection operators can be used to extract edge information from low-level feature maps, or more complex network structures can be used to capture semantic structure from deep feature maps.

[0069] Spatial alignment of structural cues and feature information in feature maps at different levels refers to establishing a spatial correspondence between structural cues extracted from different levels and the original feature information through some transformation or mapping. This can be achieved through interpolation, upsampling, downsampling, or feature fusion using attention mechanisms. For example, if the size of a lower-level feature map differs from the original feature information, it can be upsampled to the same size using bilinear interpolation, thus achieving spatial alignment. Calculating the structural correlation score between structural cues and feature information in feature maps at different levels quantifies the similarity or correlation between these aligned structural cues and feature information. This can be accomplished by calculating cosine similarity, Euclidean distance, mutual information, or through feature matching and score prediction using a specially designed neural network module. A higher score indicates a more consistent structural representation of the feature information across different levels of feature maps, and a higher confidence level in detail.

[0070] This application extracts structural cues related to feature information from feature maps at different levels of an image generation model, spatially aligns these cues with the feature information, and then calculates a structural correlation score. This allows for a comprehensive, multi-scale evaluation of the structural integrity and consistency of feature information. This multi-level structural correlation analysis enables the system to more accurately determine whether feature information represents real and important details in the image, rather than noise or unstable representations. This approach effectively avoids misjudgments caused by insufficient information at a single level, thus providing a more reliable basis for subsequent numerical amplification processing.

[0071] The above technical solutions enable more precise quantification of the structural stability of feature information within the image generation model, thereby improving the accuracy of detail confidence score calculation. This helps the system more intelligently identify secondary feature information that truly needs to be magnified to preserve detail, avoiding unnecessary processing of unimportant or noisy information, thus improving the overall efficiency of the AIGC image content intelligent generation method and the quality of the generated images.

[0072] Furthermore, the semantic correlation between feature information and image generation instructions is calculated through the following steps: The image generation instructions are semantically encoded, textual semantic features are extracted, and the similarity score between the feature information and the textual semantic features is calculated to obtain the semantic relevance score.

[0073] Specifically, semantic encoding of image generation instructions refers to converting user-inputted natural language image generation instructions into machine-understandable vector representations with semantic information. For example, pre-trained text encoders, such as Transformer-based models like BERT, GPT series, or T5, can be used to map the semantic information of words, phrases, and even entire sentences in the instructions into a high-dimensional vector space. This process aims to capture the deep semantics, such as image content, style, and emotion, contained within the instructions.

[0074] Extracting textual semantic features refers to extracting core semantic information directly related to image content generation from the semantically encoded digital representation. This can include dimensionality reduction of the encoded vector, feature selection, or focusing on key entities and descriptive words in the instruction through attention mechanisms to form a refined textual semantic feature representation. In practical applications, calculating the similarity score between feature information and textual semantic features quantifies the degree of proximity between image features and textual semantic features in the semantic space. For example, cosine similarity, the reciprocal of Euclidean distance, or other neural network-based similarity metrics can be used. This similarity score reflects the extent to which image features conform to or embody the semantic content described by the image generation instruction. Thus, a semantic relevance score can be obtained, which directly reflects the degree of semantic matching between specific feature information and the image generation instruction. The higher the score, the stronger the semantic relevance between the feature information and the instruction, and the more likely it is to represent important details in the instruction.

[0075] This application transforms abstract textual descriptions into computable digital representations by semantically encoding image generation instructions and extracting textual semantic features. By calculating the similarity score between image feature information and these textual semantic features, the degree of matching between image feature information and user intent can be quantitatively evaluated. This semantic relevance-based evaluation mechanism enables the system to identify image detail features that are semantically highly relevant to the user instruction. For example, when the instruction requests the generation of a cat wearing a red hat, image feature information related to the red hat will receive a high semantic relevance score, thus being assigned higher detail confidence in subsequent processing. This ensures that these key details are accurately amplified and encoded, ultimately reflected in the generated target image.

[0076] Through the above technical solution, this application can more accurately identify and process image detail information that is highly semantically related to the user's image generation instructions. Traditional detail processing methods may only be based on local visual features or structural correlations, while ignoring the semantic intent of the user's instructions, resulting in deviations in the detail representation of the generated image from the user's expectations. This application introduces semantic correlation as an important dimension in the calculation of detail confidence, enabling the system to prioritize and strengthen those detail features that are semantically crucial to the user's instructions, thereby significantly improving the semantic accuracy of the generated image and user satisfaction, and avoiding the problem of detail distortion or loss due to insufficient semantic understanding.

[0077] Furthermore, the step of integrating the third feature representation with the first feature representation to obtain the fourth feature representation includes: The third feature representation and the first feature representation are input into a preset fusion weight calculation network to obtain the output fusion weights; The third feature representation is multiplied element-wise with the fusion weights, and then added element-wise with the first feature representation to obtain the fourth feature representation.

[0078] Specifically, the fusion weight calculation network can be understood as a deep learning model, such as a convolutional neural network, a fully connected network, or a combination thereof. It is designed to receive two different types of feature representations—a third feature representation and a first feature representation—as input, and dynamically calculates weights to guide feature fusion based on the contextual information of the input features. The fusion weights are a set of values ​​output by the network based on the characteristics of the input features and a pre-defined fusion strategy. These weights can be a scalar, vector, or tensor, and their dimension typically matches the dimension of the third feature representation. The fusion weights reflect the importance or strength that each element or region in the third feature representation should be assigned during the integration process. Element-wise multiplication refers to multiplying each element in the third feature representation with its corresponding fusion weight. For example, if the third feature representation is a matrix and the fusion weights are also matrices of the same dimension, then element-wise multiplication involves multiplying the elements at corresponding positions in the two matrices. By weighting the detailed information in the third feature representation according to the fusion weights, important details are enhanced while unimportant details are suppressed. In practical applications, element-wise addition refers to adding corresponding elements of the weighted third feature representation to the first feature representation. For example, if the weighted third feature representation and the first feature representation are tensors of the same dimension, then element-wise addition means adding the corresponding elements of the two tensors. By effectively combining the refined details—the weighted third feature representation—with the first feature representation, which contains the overall semantics and structure, a fourth feature representation that is both rich in detail and maintains overall consistency is generated.

[0079] This application introduces a fusion weight calculation network to achieve adaptive fusion of the third feature representation and the first feature representation. Specifically, the fusion weight calculation network can intelligently learn and generate a set of dynamic fusion weights based on the overall contextual information provided by the first feature representation and the fine detail information provided by the third feature representation. These fusion weights are not fixed but are adjusted according to the current image generation instructions and the extracted feature content, thereby ensuring that the detail information can be integrated into the overall structure in the most appropriate way under different generation scenarios. By performing element-wise multiplication of the third feature representation with these dynamically generated fusion weights, the intensity of the detail information can be finely adjusted, so that details that are more consistent with the overall semantics and more critical to image quality improvement are strengthened, while details that may introduce noise or inconsistency are appropriately suppressed. Subsequently, the weighted third feature representation is added element-wise to the first feature representation, so that the detail information can be seamlessly integrated into the overall features, avoiding the abruptness or information loss that may be caused by simple superposition, thereby effectively solving the problem of inconsistency between details and the whole that may exist in traditional integration methods.

[0080] Through the above technical solution, this application can achieve a more intelligent and refined integration of image detail information and overall features. By introducing a fusion weight calculation network, the degree of integration of detail information can be dynamically adjusted according to image content and generation requirements, avoiding the problem of excessively strong or weak detail information that may result from a fixed fusion strategy. As a result, the generated fourth feature representation can more accurately characterize the complex structure and rich details of the target image, while maintaining overall semantic consistency, significantly improving the realism, refinement, and visual quality of AIGC image content, and effectively solving the limitation of traditional integration methods where detail and overall features are difficult to harmonize.

[0081] Secondly, see Figure 2 This application also proposes an AIGC image content intelligent generation system based on artificial intelligence, the system comprising: Instruction acquisition module 210 is used to acquire image generation instructions; The first feature representation generation module 220 is used to generate a first feature representation based on an image generation instruction and an image generation model. The second feature information recognition and separation module 230 is used to identify and separate second feature information from the first feature representation whose feature value is lower than a preset half-precision floating-point stable representation threshold; The third feature representation generation module 240 is used to perform numerical amplification processing on the second feature information and encode the processed second feature information into a third feature representation used to characterize image detail information. The fourth feature representation and target image generation module 250 is used to integrate the third feature representation with the first feature representation to obtain the fourth feature representation, and generate the target image based on the fourth feature representation.

[0082] The instruction acquisition module 210 can be a user interface or an interface that receives calls from external application interfaces. Its core function is to receive and parse image generation instructions provided by the user or external system. The first feature representation generation module 220 typically includes a pre-trained image generation model, such as a diffusion model or a generative adversarial network, which generates a preliminary feature representation of the overall structure and main content of the image based on the received instructions. The second feature information recognition and separation module 230 is responsible for analyzing the first feature representation, identifying and extracting detailed features that may be lost or have insufficient precision under half-precision floating-point representation. The third feature representation generation module 240 aims to improve the numerical range of the separated detailed features so that they can be represented more stably and accurately in subsequent processing, and transforms them into high-dimensional detailed feature representations through encoding. The fourth feature representation and target image generation module 250 is responsible for fusing the enhanced detailed features with the original overall features to form a final feature representation containing rich details and a complete structure, and generating a high-quality target image based on this.

[0083] This application, through a modular design, achieves effective execution of an intelligent AIGC image content generation method. Its specially designed modules ensure accurate processing of image generation instructions, refined extraction and enhancement of image features, and the completeness of detail in the final image. This effectively solves the problem of easy loss of detail information in traditional AIGC image generation, significantly improving the visual quality and realism of the generated images, and providing users with more refined and vivid AIGC image content.

[0084] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent generation of AIGC image content based on artificial intelligence, characterized in that, include: Obtain image generation instructions; Based on the image generation instructions, a first feature representation is generated through the image generation model; Identify and separate second feature information from the first feature representation whose feature value is lower than a preset half-precision floating-point stable representation threshold; The second feature information is numerically amplified, and the processed second feature information is encoded into a third feature representation used to characterize image detail information; The third feature representation is integrated with the first feature representation to obtain a fourth feature representation, and the target image is generated based on the fourth feature representation.

2. The AIGC image content intelligent generation method based on artificial intelligence according to claim 1, characterized in that, The step of generating the first feature representation based on the image generation instruction using the image generation model includes: The image generation instructions are converted into a digital representation by a text encoder, and a first feature representation including internal image feature representations is generated based on the digital representation by an initial image feature extraction network.

3. The AIGC image content intelligent generation method based on artificial intelligence according to claim 1, characterized in that, The step of numerically amplifying the second feature information and encoding the processed second feature information into a third feature representation for characterizing image detail information includes: Calculate the difference between each feature information in the second feature information and its local statistical value, and amplify the feature value corresponding to the feature information according to the difference and the preset amplification coefficient, so as to increase the numerical range of the second feature information to the level of the preset half-precision floating-point stable representation threshold. The processed feature information is transformed and compressed using an encoding neural network to generate the third feature representation.

4. The AIGC image content intelligent generation method based on artificial intelligence according to claim 3, characterized in that, Before performing numerical amplification processing on the second feature information, the following steps are included: Calculate the detail confidence score of each feature in the second feature information; The numerical amplification process is performed only on feature information whose detail confidence score exceeds a first preset threshold.

5. The AIGC image content intelligent generation method based on artificial intelligence according to claim 4, characterized in that, The detail confidence score is calculated based on at least one or more of the following factors: the consistency of local visual features of the feature information, the structural correlation of the feature information in the feature maps at different levels of the image generation model, and the semantic correlation between the feature information and the image generation instruction.

6. The AIGC image content intelligent generation method based on artificial intelligence according to claim 5, characterized in that, The local visual feature consistency of the feature information is calculated through the following steps: Multidimensional feature extraction is performed on the local region surrounding the feature information; Based on the extracted multidimensional features, a local structure description pattern is constructed. The local structure description pattern is compared with a preset real structure pattern library, and the degree of matching between the local structure description pattern and the real structure pattern in the preset real structure pattern library is calculated. A local visual feature consistency score is generated based on the degree of matching between the local structure description pattern and the real structure patterns in the preset real structure pattern library.

7. The AIGC image content intelligent generation method based on artificial intelligence according to claim 5, characterized in that, The structural correlation of the feature information in the feature maps of different levels of the image generation model is calculated through the following steps: Extract the structural clues present in the feature maps of different levels of the image generation model from the feature maps of different levels; The structural clues present in the feature maps at different levels are spatially aligned with the feature information, and the structural correlation score between the structural clues present in the feature maps at different levels and the feature information is calculated.

8. The AIGC image content intelligent generation method based on artificial intelligence according to claim 5, characterized in that, The semantic correlation between the feature information and the image generation instruction is calculated through the following steps: The image generation instructions are semantically encoded, textual semantic features are extracted, and the similarity score between the feature information and the textual semantic features is calculated to obtain a semantic relevance score.

9. The AIGC image content intelligent generation method based on artificial intelligence according to claim 1, characterized in that, The step of integrating the third feature representation with the first feature representation to obtain the fourth feature representation includes: The third feature representation and the first feature representation are input into a preset fusion weight calculation network to obtain the output fusion weight; The third feature representation is multiplied element-wise with the fusion weight, and then added element-wise with the first feature representation to obtain the fourth feature representation.

10. An AIGC image content intelligent generation system based on artificial intelligence, used to execute the AIGC image content intelligent generation method based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, The system includes: The instruction acquisition module is used to acquire image generation instructions; The first feature representation generation module is used to generate a first feature representation based on the image generation instruction and through an image generation model; The second feature information recognition and separation module is used to identify and separate second feature information from the first feature representation whose feature value is lower than a preset half-precision floating-point stable representation threshold; The third feature representation generation module is used to perform numerical amplification processing on the second feature information and encode the processed second feature information into a third feature representation used to characterize image detail information. The fourth feature representation and target image generation module is used to integrate the third feature representation with the first feature representation to obtain the fourth feature representation, and generate the target image based on the fourth feature representation.

Images