Toning engine distillation-based AI color chasing parameterization method and system

Abstract

The invention discloses an AI color chasing parameterization method and system based on toning engine distillation. The method comprises the following steps: training a distillation network, which comprises an encoder and a decoder; the distillation network is used for learning functional logic of a toning engine, so that the distillation network can output an effect picture O'which is close to the effect directly generated by the toning engine in visual effect according to an input original picture I and a group of slide bar parameters S; constructing a parameterized network based on the encoder part of the trained distillation network; training the parameterized network so that the parameterized network can predict a group of slide bar parameters S'adapted to the toning engine according to an input original image I and a reference effect image O; and finishing image color chasing based on the trained parameterized network and a toning engine. According to the invention, the functional logic of the toning engine is simulated through the distillation network, and the parameterized network is constructed based on the encoder, so that accurate mapping from the original image and the reference image to the slide bar parameters of the toning engine is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to an AI color chasing parameterization method and system based on a color tuning engine distillation. BACKGROUND

[0002] In the current field of image style transfer, the existing technologies for realizing "specifying a reference image and converting the style of an original image into the style of the reference image" (i.e., AI color chasing) mainly include four categories: CNN (Convolutional Neural Network) schemes, GAN (Generative Adversarial Network) schemes, diffusion model schemes, and Agent schemes, etc.

[0003] The generative schemes all take "directly generating a target style image" as the core path, and have high technical maturity, and can achieve basic style transfer effects. By directly generating a target style image through end-to-end learning, although there is a breakthrough in style transfer effects, there are the following key defects: first, the generation process is black-boxed, only the final style image is output, making it difficult to optimize the stylization result through parameter fine-tuning. Second, details (such as textures and edges) and resolutions of the original image are easily lost in style conversion. When processing a Raw image, due to the lack of adaptation to Raw format native data, color discontinuity and noise amplification problems are easily encountered. Third, the generated result is incompatible with the existing color tuning engine parameter system, and cannot be integrated into a professional Raw image processing workflow.

[0004] Agent scheme: an intelligent agent system driven by a multi-modal large language model (MLLM), which improves editing efficiency, but has the following significant defects: first, it relies on large multi-modal language models, which consume a lot of computing resources and are difficult to adapt to the high computing power requirements of Raw image processing; second, the generated instructions rely on third-party toolkits, and the black-boxed tools are difficult to fine-tune.

[0005] The existing technologies have the following defects: Insufficient consistency of details and resolutions: CNN schemes, GAN schemes, and diffusion model schemes all adopt the technical path of "directly generating a result image", and image details are easily lost (such as texture blurring and edge distortion) during the generation process. In particular, due to the design problem of the loss function, GAN models have a risk of pattern collapse in high-resolution image generation, resulting in distorted output results. Even advanced diffusion models still have difficulty in ensuring regional color consistency when there is a pose difference between the reference image and the target image. These defects are more pronounced when processing high-resolution original images, and cannot meet the requirements of professional photography post-processing, film and television high-definition picture color tuning, and other scenarios that require high details.

[0006] Limitations of Raw Image Support: Generative approaches such as CNNs, GANs, and diffusion models typically take input images in standard formats like JPG and PNG. Their generation logic relies on pixel-level features, making them unable to directly process Raw format images that retain the original sensor data. Although agent-based solutions like JarvisArt can indirectly support Raw formats through integration with Lightroom, this still limits their application depth in professional image processing scenarios.

[0007] Limited adjustability: Solutions such as CNN, GAN and diffusion models only output fixed stylized result maps and do not provide interactive adjustable parameters. If users are not satisfied with the results, they need to re-execute the entire style transfer process. It is impossible to achieve precise parameter fine-tuning based on the initial color tracking results, and the operation efficiency still needs to be improved. Summary of the Invention

[0008] In view of this, the purpose of this invention is to solve the three core defects of existing AI color tracking technology: First, by adopting an innovative approach of "parameterized output + engine processing," it replaces the traditional "direct pixel generation" approach, thus eradicating the resulting image detail distortion and resolution degradation. Second, by deeply coupling the AI ​​model with a self-developed color grading engine, it directly calls the engine's native processing capabilities for RAW data, thereby unlocking high-quality RAW image style transfer and filling a technological gap in this field. Third, by outputting a set of slider parameters that can be recognized and executed by the color grading engine, it transforms the black-box AI generation into transparent and controllable parametric adjustment, providing users with flexible secondary adjustment space and greatly improving workflow efficiency and accuracy.

[0009] According to one aspect of the present invention, an AI color tracking parameterization method based on color engine distillation is provided, the method comprising: A distillation network, including an encoder and a decoder, is trained. This distillation network is used to learn the functional logic of the color grading engine, so that the distillation network can output a visually similar image O' that is directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. Based on the encoder part of the trained distillation network, a parameterized network is constructed; the parameterized network is trained so that it can predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference effect image O. Image color matching is achieved based on the trained parameterized network and color engine.

[0010] In the aforementioned technical solution, the functional logic of the color grading engine is simulated through a distillation network, and a parametric network is constructed based on its encoder to achieve a precise mapping from the original and reference images to the slider parameters of the color grading engine. This design avoids the pixel-level reconstruction process common in generative schemes, thus fundamentally preserving the detailed features and resolution specifications of the original image. Simultaneously, the deep coupling between the parametric output and the color grading engine allows this method to directly utilize the color grading engine's native processing capabilities for RAW images, avoiding color distortion issues caused by format conversion. Furthermore, by outputting adjustable slider parameters, users can perform fine-tuning within the color grading engine based on the initial results, significantly improving operational efficiency and flexibility.

[0011] Specifically, the advantages of this approach are reflected in three aspects: First, regarding the consistency of detail and resolution, since the parametric network only predicts color parameters rather than directly generating images, the structural information of the original image, such as texture and edges, is fully preserved during the color engine's processing. This solves the problem of detail blurring or distortion caused by feature fusion bias in generative models, making it particularly suitable for high-precision image processing scenarios. Second, regarding support for Raw images, this method indirectly achieves native compatibility with the Raw format by parametrically adapting the color engine. The color engine can directly apply parameter adjustments to the original sensor data, avoiding color banding or noise amplification caused by pixel-level generation in existing technologies, thus filling the technical gap in Raw image style transfer. Finally, regarding adjustability, the output slider parameters allow users to make real-time secondary adjustments based on the prediction results without having to re-execute the complete style transfer process. This not only reduces computational resource consumption but also improves workflow integration and user experience.

[0012] In summary, this technical solution, through a combination of distillation and parameterization, demonstrates significant advantages in preserving image details, supporting Raw formats, and improving adjustability, providing a highly efficient, accurate, and integrable color matching solution for the professional image processing field.

[0013] In some embodiments, a distillation network is trained, including an encoder and a decoder; the distillation network is used to learn the functional logic of the color grading engine, enabling the distillation network to output a visually similar image O' to the one directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. Specifically: Construct a distillation network, including an encoder and a decoder; acquire a set of sample images, and for each sample image I, generate multiple sets of first slider parameters S, either randomly or based on a specific policy distribution; input each set (I, S) into a color grading engine, which processes the image and outputs the corresponding first effect image O; save each generated set of (I, S, O) data pairs as training data pairs; use each set (I, S) as input to train the distillation network, calculate the difference loss between each set (O) and the second effect image O' output by the distillation network, and train the distillation network based on the calculation results.

[0014] In the above technical solution, the distillation network training process constructs a neural network with an encoder-decoder structure and learns the internal mapping logic of a professional color grading engine in a data-driven manner. Its core advantage lies in transforming the complex, potentially black-box, color grading engine functions into a differentiable and portable neural network model.

[0015] Specifically, the advantages of this training path are reflected in three aspects. First, regarding the fidelity of knowledge transfer, by generating multiple sets of randomly or strategically distributed slider parameters for each sample image and forming training pairs with the real effect images output by the color grading engine, it is ensured that the training data can broadly cover the operational space of the color grading engine. This allows the distillation network to learn the complete behavioral patterns of the engine under various parameter combinations, rather than a single style. This guarantees the comprehensiveness and accuracy of the distillation model's simulation of the color grading engine's functions. Second, regarding the model's versatility and generalization ability, the training objective is to minimize the difference loss between the distillation network output and the color grading engine output. This drives the distillation network to capture the most essential color and tonal transformation rules of the color grading engine, rather than memorizing specific image features. This allows it to stably generate effects that conform to the logic of the color grading engine when faced with new original images and new parameters outside the training set. Finally, regarding the seamlessness of the technical path, this step successfully constructed a proxy network that is functionally equivalent to the color grading engine. Its encoder part, through the processing of inputs (I, S), already contains a high-level abstract representation of the color grading logic. This provides a seamless technical foundation for directly reusing the encoder to build a parametric network in the next step, realizing a smooth transition from the simulation of "parameters to effects" to the reverse prediction of "effects to parameters".

[0016] In summary, the training phase of this distillation network, through systematic data construction and precise loss function guidance, achieves high-quality distillation of the professional color grading engine's functions. This not only ensures the engine-level fidelity of the style transfer effect, but more importantly, it provides a core, reusable network module for the entire parametric color tracking solution. This is a crucial and foundational step in achieving the final technical effect.

[0017] In some embodiments, a parameterized network is constructed based on the encoder portion of the trained distillation network; this parameterized network is trained to predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference image O, specifically: Construct a parameterized network; decompose the encoder part of the trained distillation network and use it as the feature extractor of the parameterized network; based on the data pairs, use each group (I, O) as the input to train the parameterized network, calculate the difference loss between each group (S) and the second slider parameter S' output by the parameterized network, and train the parameterized network based on the calculation results.

[0018] In the above technical solution, this stage reuses the encoder of the pre-trained distillation network as the core of feature extraction, constructing a parametric network capable of accurately predicting color parameters based on the original and reference images. The core advantage of this method lies in its innovative transformation of the visual style color tracking problem into a parametric regression problem, thereby bypassing the traditional path of direct pixel generation and providing a solid technical foundation for achieving high-fidelity, adjustable style transfer that is compatible with professional workflows.

[0019] Specifically, the advantages of this training path are reflected in three aspects. First, regarding the accuracy and interpretability of the inverse mapping, this method ensures that the semantics of the features extracted from the image pair (I,O) are highly correlated with the color adjustment operation by fixing and utilizing the encoder that has already distilled the color adjustment engine logic. By training directly with the original slider parameter S as the supervision signal, the parameterized network is driven to learn an accurate inverse mapping from the target visual effect to the specific parameter combination. The output parameters themselves are the native instructions of the color adjustment engine, with clear physical meaning, fundamentally solving the problem of black box operation and non-fine-tunable results in generative models. Second, regarding model efficiency and training stability, the split encoder, as a powerful feature extractor, provides stable and discriminative feature representations. This not only significantly reduces the data requirements and convergence difficulty of training the parameterized network, but also avoids the instability and modality collapse risks that may be encountered when training an inverse model from scratch, ensuring the robustness of parameter prediction. Finally, in terms of system integration and practicality, the finally trained parametric network can directly output slider parameters S' that the color grading engine can recognize. This allows the AI ​​color tracking results to be seamlessly integrated into existing professional post-production workflows. Users can make arbitrary fine adjustments based on the predicted parameters, greatly improving work efficiency and creative flexibility.

[0020] In summary, this parameterized network successfully achieved a reliable and efficient conversion from reference images to color grading parameters during the training phase through shared encoder knowledge and regression loss supervision. This design not only ensures the quality of the predicted parameters but, more importantly, enables synergy between AI style transfer and professional color grading tools, making it the core component of this technical solution.

[0021] In some embodiments, image color matching is performed based on a trained parameterized network and a color tuning engine, specifically: The original image I provided by the user and the reference image R of the target style are input into the trained parameterized network to obtain the predicted slider parameters S'; the original image I and the slider parameters S' are input into the color grading engine to generate the final stylized result image.

[0022] In the above technical solution, the image color tracking execution stage is the final application link of the whole method. Its design reflects the closed-loop logic from perception to execution. By synergistically integrating the predictive ability of the parameterized network with the rendering ability of the professional color grading engine, efficient, reliable and high-quality automated style transfer is achieved.

[0023] Specifically, the advantages of this stage are first reflected in the quality and reliability of the generated results. Since the predicted slider parameters S' are specifically optimized for downstream color grading engines, their instructions can be parsed and executed by the color grading engine. The color grading engine applies these parametric adjustments based on the original image I's raw data, preserving all the details, dynamic range, and resolution of the original image to the greatest extent possible. This completely avoids the information loss and uncertainty inherent in generative models due to pixel reconstruction, ensuring the visual professionalism and technical robustness of the output results. Secondly, this solution achieves seamless integration with professional image processing workflows. Its final output includes two levels: an intuitive stylized result image and a structured, readable, and editable slider parameter S'. This provides users with unprecedented operational flexibility; when users are not satisfied with the initial color grading effect, they do not need to rerun the AI ​​model but can directly fine-tune S' within the familiar color grading engine interface, greatly improving creative efficiency and the controllability of the results, fundamentally solving the rigidity problem of traditional generative models' "one-shot deal." Finally, this scheme demonstrates high efficiency in terms of computational resource utilization. The parameterized network only needs to perform one forward propagation to generate a compact parameter vector S', while all subsequent computationally intensive image rendering tasks are handled by a highly optimized dedicated color tuning engine. This collaborative strategy, compared to diffusion models that require iterative denoising or GAN models that undergo complex adversarial training, significantly reduces the overall demand for computational resources and improves processing speed while ensuring high-quality output.

[0024] In summary, this image color tracking execution stage successfully transforms the color grading knowledge learned in previous steps into stable, high-quality, and interactive visual output through a collaborative paradigm of "parametric network predicting instructions and color grading engine executing rendering." This design not only ensures professional-grade image quality in the final result but also provides a highly practical solution for the application of AI technology in professional image processing through its deep compatibility with existing tools and efficient computational path.

[0025] In some embodiments, generating a final stylized result map further includes: The color grading engine receives the user's adjustment instruction for the predicted slider parameter S', and reprocesses the original image I according to the adjusted slider parameter to generate a new result image.

[0026] In the aforementioned technical solution, this stage allows users to make real-time, fine-tuned adjustments to S' based on their own aesthetic preferences and application scenarios. This design effectively compensates for potential biases in intent understanding that purely automated models may have. Each user adjustment involves directly reprocessing the original image I within the color grading engine, without needing to re-call or run AI models such as parametric networks, resulting in fast iterative adjustments and low computational overhead. More importantly, it always operates based on the original image's raw data (especially when processing RAW images), ensuring the highest quality output image across countless adjustment iterations and completely avoiding the risk of detail accumulation loss or artifact amplification that can occur with repeated processing by generative models. Furthermore, the user's adjustment behavior itself serves as feedback to the AI's prediction results. These "parameter-adjustment" data pairs can be systematically collected and used for further fine-tuning and optimization of the parametric network in the future, enabling the entire system to continuously learn from actual human operations.

[0027] In some embodiments, the color grading engine can directly process Raw format images; The original image I and / or the reference image R processed by the method are Raw format images.

[0028] In the aforementioned technical solution, the Raw format file records the most original, uncompressed data from the image sensor, possessing the highest bit depth and dynamic range. This method directly applies AI-predicted slider parameters to this raw data through a color grading engine. All color and tonal adjustments are performed within the linear raw data space, thereby maximizing the full information potential of the Raw file and avoiding the information bottlenecks, color banding, or loss of highlight / shadow detail that inevitably occur when performing style transfer on lossy compressed images such as JPG / PNG. This solution constructs a lossless technical path from Raw to Raw. Whether it's the original image I, the reference image R, or the final processed output, all can be based on the Raw format, ensuring the highest quality data dimension throughout the entire process from acquisition and editing to output. This feature enables seamless integration of AI technology with professional photography workflows. Professional photographers' workflows are centered on the Raw format, and this method can obtain an AI-driven color-tracking starting point in their familiar Raw processing environment without converting file formats, and then perform lossless adjustments based on this.

[0029] According to another aspect of the present invention, an AI color tracking parameterization device based on color engine distillation is provided, wherein the system, based on the above method, comprises: The distillation module is used to train a distillation network, including an encoder and a decoder. The distillation network is used to learn the functional logic of the color grading engine, so that the distillation network can output a visually similar image O' to the one directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. The parameterization module is used to construct a parameterized network based on the encoder part of the trained distillation network; and to train the parameterized network so that it can predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference effect image O. The color tracking module is used to perform image color tracking based on the trained parameterized network and color engine.

[0030] In order to better utilize the above method, this application proposes an AI color tracking parameterization device based on color engine distillation. Each module corresponds to a step of the above method, and its specific principle has been described above and will not be repeated here.

[0031] According to another aspect of the present invention, an AI color tracking parameterization device based on color engine distillation is provided, comprising: At least one processor and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0032] In the above technical solution, to better operate and process the method, the method is stored in memory, and the processor executes the stored method. It should be noted that the principle and effect of each step have been described above and will not be elaborated upon here.

[0033] According to another aspect of the present invention, a computer-readable storage medium is provided storing a computer program, characterized in that the computer program implements the above-described method when executed by a processor.

[0034] In the above technical solution, to better operate and use the method, the method is stored in a computer-readable storage medium and implemented using a processor. It should be noted that the principle and effect of each step have been described above and will not be elaborated upon here. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating an embodiment of an AI color tracking parameterization method based on color engine distillation according to the present invention. Figure 2 This is a schematic diagram of an embodiment of an AI color tracking parameterization device based on color engine distillation according to the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] This invention provides an AI color tracking parameterization method and system based on color grading engine distillation, which can solve three core defects of existing AI color tracking technologies: First, by adopting an innovative path of "parameterized output + engine processing," it replaces the traditional "direct pixel generation" path, fundamentally solving the problems of image detail distortion and resolution reduction caused by this approach; Second, by deeply coupling the AI ​​model with a self-developed color grading engine, it directly calls the engine's native processing capabilities for RAW data, thereby unlocking high-quality RAW image style transfer and filling a technological gap in this field; Third, by outputting a set of slider parameters that can be recognized and executed by the color grading engine, it transforms the black-box AI generation into transparent and controllable parameterized adjustment, providing users with flexible secondary adjustment space and greatly improving the efficiency and accuracy of the workflow.

[0039] Example 1 This method is applicable to the field of image style transfer, specifically in scenarios such as post-processing of photography, color grading of film and television footage, and unifying the style of graphic design. It receives the original image and a reference image of a specified style from the user, and generates a set of slider parameters adapted to a self-developed image color grading engine using AI algorithms. Users can input these parameters into the self-developed image color grading engine to adjust the original image to match the style of the reference image. It also supports secondary adjustments based on these parameters and is compatible with Raw format images, ensuring consistency in detail and resolution between the resulting image and the original image.

[0040] Please see Figure 1 An AI color tracking parameterization method based on color engine distillation, the method comprising: S1. Train a distillation network, including an encoder and a decoder; the distillation network is used to learn the functional logic of the color grading engine, so that the distillation network can output a visually similar effect O' to the effect directly generated by the color grading engine based on the input original image I and a set of slider parameters S. In this embodiment, a distillation network, including an encoder and a decoder, is trained. This distillation network is used to learn the functional logic of the color grading engine, enabling it to output a visually similar effect image O' generated directly by the color grading engine based on the input original image I and a set of slider parameters S. Specifically: a distillation network, including an encoder and a decoder, is constructed; a sample image set is obtained, and for each sample image I, multiple sets of first slider parameters S, either randomly distributed or based on a specific strategy, are generated; each set (I, S) is input to the color grading engine, which processes it and outputs the corresponding first effect image O; each generated set (I, S, O) data pair is saved as a training data pair; each set (I, S) is used as the input to train the distillation network, and the difference loss between each set (O) and the second effect image O' output by the distillation network is calculated; the distillation network is trained based on the calculation results. For example, the first stage: color grading engine distillation This stage involves learning the functional logic of the self-developed color grading engine through distillation networks. Specific steps include: (1) Construct a distillation network: including an encoder and a decoder, where the encoder is responsible for extracting image features and the decoder is responsible for generating the color-corrected image based on the features and slider values; (2) Data preparation: Obtain a sample image set and generate training data pairs of "original image + slider value → effect image" through the color adjustment engine (i.e., the input is the original image I and the slider parameter S, and the output is the effect image O). (3) Distillation training: Input the original image I and the slider parameter S into the distillation network and output the predicted image O'; take the image O generated by the color engine as the target, calculate the difference loss (such as MSE loss) between O and O', and optimize the distillation network parameters through backpropagation; (4) Convergence determination: When the loss value is lower than the preset threshold (i.e., the network output is close enough to the color grading engine output), the distillation is determined to be complete. At this time, the distillation network has "understood" the functional logic of the color grading engine (i.e., the mapping relationship between the slider parameters and the image style changes).

[0041] S2. Based on the encoder part of the trained distillation network, construct a parameterized network; train the parameterized network so that it can predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference effect image O. In this embodiment, a parameterized network is constructed based on the encoder portion of the trained distillation network; this parameterized network is trained to predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference image O. Specifically: Construct a parameterized network; decompose the encoder part of the trained distillation network and use it as the feature extractor of the parameterized network; based on the data pairs, use each group (I, O) as the input to train the parameterized network, calculate the difference loss between each group (S) and the second slider parameter S' output by the parameterized network, and train the parameterized network based on the calculation results.

[0042] For example, the second stage: parameterized network training This stage leverages the pre-training results of the distillation network to achieve the inverse mapping of "original image + result image → slider parameters". Specific steps are as follows: (1) Network splitting: The encoder part of the distillation network trained in the first stage is split into independent modules and used as feature extractors of the parameterized network; (2) Constructing a parametric network: Based on the split encoder, a new parameter prediction head (output layer) is added to form a complete parametric network (the input is the original image I and the effect image O, and the output is the slider parameter S'). (3) Data preparation: Using the same sample image set as in the first stage, construct training data pairs of "original image I + effect image O → slider parameter S"; (4) Parametric training: Input the original image I and the effect image O into the parametric network, extract the feature differences between the two through the encoder, and output the predicted slider parameter S' by the parameter prediction head; take the real slider parameter S as the target, calculate the difference loss between S and S' (such as L1 loss), and optimize the network parameters; (5) Accelerated convergence: Since the encoder has learned the functional logic of the color tuning engine through the distillation stage, the parameterized network can quickly converge to a high-precision state.

[0043] S3. Based on the trained parameterized network and color engine, image color matching is completed.

[0044] In this embodiment, the original image I provided by the user and the reference image R of the target style are input into the trained parameterized network to obtain the predicted slider parameter S'; the original image I and the slider parameter S' are input into the color grading engine to generate the final stylized result image.

[0045] In this embodiment, the final stylized result map is generated, followed by: The color grading engine receives the user's adjustment instruction for the predicted slider parameter S', and reprocesses the original image I according to the adjusted slider parameter to generate a new result image.

[0046] In this embodiment, the color grading engine can directly process Raw format images; the original image I and / or the reference image R processed by the method are Raw format images.

[0047] In this embodiment, the main goal of optimizing the sampling strategy for the slider parameter S is to ensure that the generated parameters not only conform to the operation definition and usage specifications of the slider itself, but also effectively cover the complete operating space of the color grading engine. Operations that do not conform to the specifications often lead to problems such as color banding and brightness inversion in the image, resulting in image quality degradation. Therefore, the core of the sampling strategy lies in generating effective data pairs that can avoid such image quality degradation. Specific methods include, but are not limited to: avoiding undesirable parameter values ​​from the algorithm principle level, or extracting parameters based on high-quality sample images actually processed by designers. Furthermore, if the parameter S in the color grading engine is a discrete variable or has complex dependencies, the sampling strategy can be adjusted accordingly. For example, the distribution of sampling points can be reasonably designed in the discrete parameter space, or the sampling process can be guided by modeling the dependencies between parameters, thereby maintaining the adaptability and robustness of the model while ensuring the quality of the generated data.

[0048] In this embodiment, addressing the perceptual consistency problem of high-level visual features (such as color distribution and contrast) in color grading tasks, relying solely on pixel loss may be insufficient to fully capture the visual similarity perceived by the human eye. Although pixel loss can numerically measure the difference between two images—generally, the smaller the value, the closer the images are at the pixel level—the quality of color grading depends more on the relative color relationships within the images. Only when the mutual constraints and correlations between colors are maintained can a harmonious and natural visual result be produced. In this regard, the distillation object used in this invention—the color grading engine itself—naturally possesses the ability to maintain such relative color relationships within its internal mechanism. Therefore, during the distillation process, even if pixel-level loss cannot fully bear the high-level semantic consistency, by learning and imitating the color processing behavior of the engine itself, the model can still maintain perceptual color harmony to a certain extent, thereby alleviating the perceptual inconsistency problem that may be caused by pure pixel matching.

[0049] In this embodiment, when the reference image R used in the color matching process differs significantly from the original image O in the training data in terms of style, content, or image quality, the slider parameters S' predicted by the parameterized network may indeed be inaccurate, leading to a deviation from the expected color matching effect. Fundamentally, the color matching engine itself has a specific range of expressive capabilities. An extreme case is when the input image I is black and white, while the target image O is color; the color mapping relationship between the two exceeds the range that the engine can achieve with a set of parameters S. Therefore, whether as training data or in actual color matching tasks, a fundamental premise is that the color mapping relationship from I to O must be within the expressive capabilities of the engine. If the difference between the reference image and the source image causes this premise to no longer be met, then the predicted slider parameters will naturally struggle to achieve the desired color matching effect.

[0050] In this embodiment, encoder sharing may lead to a bias in feature representation towards the image reconstruction task, thus affecting the prediction performance of slider parameters. However, the features extracted by the encoder are largely universal. Since these features can support the reconstruction task from input image I to target image O, they contain sufficient information to fully express the I-to-O transformation relationship. Therefore, these features should theoretically also be applicable to parameter prediction tasks. Practical results have verified this universality. Even with some task bias, only slight refinement of the features is usually needed to effectively adapt them to the prediction requirements of slider parameters without redesigning or training the entire encoder structure.

[0051] In this embodiment, different combinations of slider parameters may produce similar visual effects, i.e., the phenomenon of "non-unique solution". Simultaneously, when the reference image R differs significantly from the source image I, it also poses a challenge to parameter prediction. However, during the training process of the parameter prediction network, the method of this invention consistently emphasizes that the predicted slider parameters must be within the effective range acceptable to the color grading engine. This constraint comes not only from the design of the model structure itself but also from the explicit constraint on the parameter range in the loss function. From the construction stage of the training data, we have already screened the validity of the true parameters S used in all samples, ensuring that they are within a reasonable and feasible range. The network learns based on this, and its prediction results naturally tend to fall within a similar effective distribution. Even in rare cases where the model outputs parameters outside the normal range, the color grading engine itself has an inherent fault tolerance and saturation handling mechanism. Specifically, if a parameter slightly exceeds its design boundary, the engine will automatically truncate it to the closest effective value, thereby avoiding illegal operations or image anomalies. This "soft constraint" characteristic makes the entire system more stable and robust in practical applications.

[0052] Based on the above embodiments, the present invention has the following advantages: 1. Existing technologies (such as CNN and GAN solutions) often suffer from detail loss and resolution mismatch when directly generating stylized images due to feature fusion deviations or generator limitations. This invention addresses these issues by using an "AI color-tracking to generate effect images + parametric adaptation color tuning engine" approach. Instead of directly reconstructing image pixels, it leverages the original image's details and resolution by adjusting parameters to achieve style transfer. This fundamentally preserves the original image's detailed features (such as texture and edge contours) and resolution specifications, solving the core problems of image detail distortion and resolution inconsistency in existing technologies. It is particularly suitable for scenarios with extremely high image precision requirements, such as printing and film post-production.

[0053] 2. Current CNN and GAN solutions do not support raw image processing, and existing generative techniques are difficult to be compatible with their data formats. This invention, through deep coupling of parameterized output and a self-developed color grading engine, utilizes the color grading engine's native processing capabilities for raw images to directly apply the parameterized results to the original raw image data. This avoids the color banding and detail compression problems caused by format incompatibility in existing technologies, fills the technical gap in the field of raw image style transfer, and provides key technical support for professional post-processing photography.

[0054] 3. Existing technologies only output the final stylized image, preventing users from fine-tuning based on the original conversion logic, resulting in extremely low operational flexibility. This invention outputs slider parameters (such as exposure, saturation, and tone separation) adapted to a self-developed color grading engine. Users can directly adjust these parameters within the color grading engine without needing to re-initiate a style conversion request, significantly improving operational efficiency.

[0055] Example 2 Please see Figure 2 An AI color tracking parameterization device based on color engine distillation, based on the method described in one embodiment, the system comprising: The distillation module is used to train a distillation network, including an encoder and a decoder. The distillation network is used to learn the functional logic of the color grading engine, so that the distillation network can output a visually similar image O' to the one directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. The parameterization module is used to construct a parameterized network based on the encoder part of the trained distillation network; and to train the parameterized network so that it can predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference effect image O. The color tracking module is used to perform image color tracking based on the trained parameterized network and color engine.

[0056] In order to better utilize the method described in one of the embodiments, this application proposes an AI color tracking parameterization device based on color engine distillation. Each module corresponds to each step of the above method, and its specific principle has been described above and will not be repeated here.

[0057] Example 3 An AI-based color tracking parameterization device based on color engine distillation includes: At least one processor and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described in one of the embodiments.

[0058] In the above technical solution, in order to better operate and process the method described in one of the embodiments, the method is stored in a memory, and the stored method is executed by a processor. It should be noted that the principle and effect of each step have been described above and will not be elaborated further here.

[0059] Example 4 A computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the method described in one of the embodiments.

[0060] In the above technical solution, to better operate and use the method described in one of the embodiments, the method is stored in a computer-readable storage medium and implemented using a processor. It should be noted that the principle and effect of each step have been described above and will not be elaborated further here.

[0061] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An AI color tracking parameterization method based on color engine distillation, characterized in that, The method includes: A distillation network, including an encoder and a decoder, is trained. This distillation network is used to learn the functional logic of the color grading engine, so that the distillation network can output a visually similar image O' that is directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. Based on the encoder part of the trained distillation network, a parameterized network is constructed; the parameterized network is trained so that it can predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference effect image O. Image color matching is achieved based on the trained parameterized network and color engine.

2. The AI ​​color tracking parameterization method based on color engine distillation as described in claim 1, characterized in that, A distillation network, including an encoder and a decoder, is trained. This distillation network learns the functional logic of the color grading engine, enabling it to output a visually similar image O' to the one directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. Specifically: Construct a distillation network, including an encoder and a decoder; acquire a set of sample images, and for each sample image I, generate multiple sets of first slider parameters S, either random or based on a specific policy distribution; Each group (I, S) is input into the color grading engine, which processes it and outputs the corresponding first effect image O. Each generated (I, S, O) data pair is saved as a training data pair; each (I, S) is used as the input to train the distillation network; the difference loss is calculated between each (O) and the second effect map O' output by the distillation network; and the distillation network is trained based on the calculation results.

3. The AI ​​color tracking parameterization method based on color engine distillation as described in claim 2, characterized in that, Based on the encoder portion of the trained distillation network, a parameterized network is constructed; this parameterized network is trained to predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference image O. Specifically: Construct a parameterized network; decompose the encoder part of the trained distillation network and use it as the feature extractor of the parameterized network; based on the data pairs, use each group (I, O) as the input to train the parameterized network, calculate the difference loss between each group (S) and the second slider parameter S' output by the parameterized network, and train the parameterized network based on the calculation results.

4. The AI ​​color tracking parameterization method based on color engine distillation as described in claim 1, characterized in that, Image color matching is performed based on the trained parameterized network and color engine, specifically: The original image I provided by the user and the reference image R of the target style are input into the trained parameterized network to obtain the predicted slider parameters S'; the original image I and the slider parameters S' are input into the color grading engine to generate the final stylized result image.

5. The AI ​​color tracking parameterization method based on color engine distillation as described in claim 4, characterized in that, Generate the final stylized result image, which then includes: The color grading engine receives the user's adjustment instruction for the predicted slider parameter S', and reprocesses the original image I according to the adjusted slider parameter to generate a new result image.

6. The AI ​​color tracking parameterization method based on color engine distillation as described in any one of claims 1 to 5, characterized in that, The color grading engine can directly process Raw format images; the original image I and / or the reference image R processed by the method are Raw format images.

7. An AI color tracking parameterization device based on color engine distillation, characterized in that, Based on the method according to any one of claims 1-6, the system comprises: The distillation module is used to train a distillation network, including an encoder and a decoder. The distillation network is used to learn the functional logic of the color grading engine, so that the distillation network can output a visually similar image O' that is directly generated by the color grading engine, based on the input original image I and a set of slider parameters S. The parameterization module is used to construct a parameterized network based on the encoder part of the trained distillation network; and to train the parameterized network so that it can predict a set of slider parameters S' adapted to the color grading engine based on the input original image I and the reference effect image O. The color tracking module is used to perform image color tracking based on the trained parameterized network and color engine.

8. An AI-based color tracking parameterization device based on color engine distillation, characterized in that, include: At least one processor and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images