Image editing method and device based on adaptive classifier-free guided implicit denoising diffusion model

Abstract

The invention discloses an image editing method and device based on an adaptive classifier-free guided implicit denoising diffusion model. The method comprises the following steps: setting different classifier-free guided parameters to carry out small-amount noise addition and conditional denoising on an image so as to quickly obtain an editing result sample; selecting the CFG corresponding to the optimal editing effect under the current precision through an optimization algorithm according to the obtained sample; the precision of the optimal CFG is improved through iterative parameter space expansion; and performing adaptive editing execution according to the optimal CFG execution to obtain a final image. According to the invention, the time for manually adjusting parameters is reduced, and the image editing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to image editing, diffusion model and conditional image generation technology. Image editing is achieved by combining a generative diffusion model, and specifically relates to an image editing method based on an adaptive implicit denoising diffusion model without classifier guidance. Background Art

[0002] Image editing is an important research direction in the field of computer vision, covering many important application fields such as image restoration, style transfer, target replacement, and semantic generation. This technology plays a key role in the production of creativity, advertising design, medical image processing and other fields, and provides strong technical support for efficient image processing and information visualization. In recent years, image editing technology based on diffusion models has made significant progress. Its core method is to initialize a noise map that conforms to the standard Gaussian distribution in the latent space, and then predict the noise through a deep neural network architecture (such as transformer or U-net), and use iterative denoising to gradually generate high-quality images. Among them, the implicit denoising diffusion model (Song, J., Denoising Diffusion Implicit Models, International Conference on Learning Representations 2021), as an efficient generation model, has become one of the mainstream methods in this field due to its advantages in generation quality and computational efficiency. By introducing a non-Markov process, DDIM significantly improves the sampling efficiency while maintaining the generation quality, providing a more flexible and efficient solution for image editing tasks. The denoising generation process of DDIM is as follows:

[0003]

[0004] in: ∈ t is random noise, is the noise predicted by the neural network, α t is a time-dependent sequence that affects noise intensity, x t Images of each step of the diffusion process.

[0005] Different from the diffusion model for image generation, current image editing based on the diffusion model usually uses the inverse diffusion process or directly adds appropriate noise to generate a noisy image. The purpose is to obtain a specific standard Gaussian noise, and then realize image editing through conditional denoising. Based on the determinism of DDIM diffusion, the DDIM denoising process can be reversed to perform noise addition with a predictive process, namely DDIM_inversion (Mokady R, Null-text inversion for editing real imagesusing guided diffusion models, Proceedings of the IEEE / CVF conference oncomputer vision and pattern recognition.2023). DDIM_inversion inverts the real image into standard Gaussian noise, which contains some of the original image features. These features can still be retained after conditional generation, thereby improving the stability of image editing. DDIM_inversion approximates the noise in DDIM diffusion Sampling is done using the following equation:

[0006]

[0007] During the diffusion process, CFG (Ho J, Classifier-Free-Guidance, arXiv preprint arXiv:2207.12598, 2022) is used to control the influence of conditions. CFG, as an unconditional classification guide, has a decisive influence on the effect of editing. The formula that affects the diffusion process is as follows:

[0008] ∈ θ (x t .c,t)=∈ θ (x t ,t)+w(∈ θ (x t ,c,t)-∈ θ (x t ,t))

[0009] Where: w is the guidance scale hyperparameter, i.e. CFG, ∈ θ (x t ,c,t) is the estimated conditional noise,∈ θ (x t ,t) is the estimated unconditional noise.

[0010] To achieve ideal image editing results, it is usually necessary to manually debug the control parameters CFG multiple times. When the CFG parameter is set to 1 (i.e., the classifier-free guidance mechanism is not enabled), the image will undergo significant and unexpected modifications; when the CFG parameter is set too high, too many original image features will be retained, resulting in failure to achieve the desired editing effect. It is worth noting that the optimal CFG parameters corresponding to different diffusion model architectures and input conditions vary significantly. Only when the CFG parameters are set appropriately can the ideal editing effect be achieved. Currently, this process mainly relies on manual experience for multiple debugging and optimization.

[0011] The fundamental reason for this reliance on manual debugging is the current lack of effective quantitative indicators to evaluate the quality of editing effects. Specifically, existing image quality assessment indicators have the following limitations: (1) Generation quality assessment indicators such as Fréchet Inception Distance (FID) are more suitable for the overall quality assessment of generation tasks, but cannot accurately determine whether the editing results meet the given conditional constraints; (2) Although evaluation indicators based on pixel differences such as mean square error (MSE) can quantify the degree of change before and after editing, they cannot effectively reflect the subjective quality of the editing effect; (3) Prediction methods based on large multimodal models (such as CLIP Similarity Score) have uncertainties in accuracy and computational efficiency, making it difficult to meet the needs of practical applications. Therefore, developing quantitative evaluation methods suitable for image editing tasks remains an important challenge in current research. Summary of the Invention

[0012] To address the above problems, the present invention proposes an image editing method and device based on an adaptive classifier-free implicit denoising diffusion model. Only a small amount of additional time is required to adaptively select CFG, and no additional training of the model is required. This avoids the tedious process of multiple manual debugging and parameter selection required in traditional methods, thereby significantly improving the efficiency and automation of image editing.

[0013] A first aspect of the present invention relates to an image editing method based on an adaptive classifier-free implicit denoising diffusion model, which specifically comprises the following steps:

[0014] (1) Generate multi-parameter samples by partial noise inversion: Perform partial noise inversion with controllable step size on the input image, generate multiple classifier-free-guidance (CFG) parameters based on adaptive arithmetic progression, and perform conditional denoising in parallel to generate edited samples;

[0015] (2) Pseudo-gradient optimization to select the best CFG: Based on the mean square error (MSE) data pair between the sample and the original image, the MSE mutation point is located by second-order differential gradient calculation to determine the initial optimal CFG parameters;

[0016] (3) Iterative parameter space expansion optimization: construct a parameter interval with the initial optimal CFG as the center, and iteratively perform steps (1) to (2) to improve parameter accuracy;

[0017] (4) Adaptive image editing execution: Full noise inversion and conditional denoising are performed based on the final optimized CFG, and the number of inversion steps is dynamically adjusted according to the output MSE to balance the editing strength and fidelity.

[0018] Wherein, step (1) specifically includes:

[0019] (1-1) CFG selection during sampling. Since pseudo-gradient calculation is required for the sampling results (see step (2) for details), and sampling requires a high computational cost, the first CFG selection is to select an adaptive arithmetic progression within the conventional range r (1 to 14) according to the change in accuracy. In the secondary optimization stage of step (2), based on the accuracy feedback of the first sampling, an improved arithmetic progression guided by accuracy is established in the corresponding parameter subinterval.

[0020] (1-2) Samples are obtained by partially inverting and adding noise. The degree of noise inversion directly affects the time required for each sampling. Unlike the traditional method of inverting the image to standard Gaussian noise, the degree of noise inversion is controlled by the number of inversion steps for each sampling to reduce computational overhead.

[0021] (1-3) Compare the obtained sample with the original image to obtain the CFG-MSE data pair.

[0022] Furthermore, the adaptive arithmetic progression formula used in step (1-1) according to the change of precision is:

[0023]

[0024] where r start , r end Represents the start and end values of the selected range, s is the precision, and the tolerance of the series is automatically adjusted by the precision. The higher the precision, the more samples are taken and the longer the time consumed.

[0025] The degree of noise inversion used in steps (1-2) is controlled by the number of inversion steps:

[0026] T skip =t / T (2)

[0027] Where T,t represents the number of diffusion steps, T represents the number of steps required to invert the image to standard Gaussian noise, and t represents the number of inversion steps required to add partial noise. The default value is 0.1T.

[0028] Wherein, step (2) specifically includes:

[0029] The pseudo gradient optimization is used to select the best CFG: based on the mean square error MSE data of the sample and the original image [x i ,y i ], locate the MSE mutation point through second-order difference calculation and determine the preliminary optimal CFG parameters;

[0030] The pseudo gradient optimization algorithm formula is as follows:

[0031] y′ i =y i+1 -y i ,y″ i =y′ i -y′ i-1 (3)

[0032] Select y″ i The maximum CFG value is used as the optimal parameter. Pseudo-gradient calculation is used instead of gradient calculation to reduce computing resources.

[0033] Wherein, step (3) specifically includes:

[0034] An iterative parameter space expansion optimization is used: a shrinking parameter interval is constructed with the initial optimal CFG as the center, and the optimization steps are iteratively performed to improve the parameter accuracy; the shrinking formula of the parameter interval is:

[0035] r=W prior +s*[-1,1] (4)

[0036] Where W prior is the optimal CFG selected, and s is the accuracy.

[0037] Wherein, step (4) specifically includes:

[0038] Adaptive image editing execution is adopted: full noise inversion and conditional denoising are performed based on the final optimized CFG, and the number of inversion steps is dynamically adjusted according to the output MSE to balance editing strength and fidelity.

[0039] The formula for dynamically adjusting the number of inversion steps is:

[0040] T skip =T skip (1+α·(τ-MSE)) (5)

[0041] τ is the preset threshold, and α is the adjustment coefficient.

[0042] The innovative features of the present invention are:

[0043] 1. Partial noise inversion sampling mechanism

[0044] By dynamically controlling the noise inversion step size, only a portion of the input image needs to be degraded to a noise state, avoiding the computational redundancy of traditional methods that require complete inversion to Gaussian noise, and significantly reducing the inversion time cost.

[0045] 2. Pseudo-gradient optimization algorithm

[0046] A fast parameter optimization method based on second-order difference is proposed. By using pseudo-gradient to simulate gradient calculation to locate the MSE mutation point, the traditional gradient optimization algorithm is replaced with O(n) time complexity, thus reducing the consumption of a large amount of computing resources.

[0047] 3. Iterative parameter space expansion strategy

[0048] By adopting the dynamic interval shrinkage formula, the parameter accuracy is improved by multiple orders of magnitude through multiple iterations, which significantly reduces the computing resources compared to directly using high precision.

[0049] 4. Adaptive Image Editing

[0050] A dynamic step adjustment formula is introduced to balance content fidelity and editing intensity in real time based on the editing results, solving the inherent problem of "over-editing" in traditional methods.

[0051] The working principle of the present invention is:

[0052] 1. In the image editing process based on the implicit diffusion model, when the influence of the CFG (text condition) increases to a certain level, other conditions being equal, it will trigger a structural mutation in the feature space, manifested as a step-by-step decrease in image quality and a sudden increase in the MSE value. Therefore, by finding the MSE mutation point, the CFG parameters corresponding to the optimal image editing effect can be determined, replacing the traditional manual trial-and-error parameter adjustment process and achieving fully automated operation.

[0053] 2. In the process of automatically finding MSE mutation points, we need to face the dual bottlenecks of sampling computing resource consumption and gradient computing resource consumption of traditional mutation point finding methods. Based on the image, adding partial noise and performing denoising under different conditional influences can also generate different image results, and the optimal conditional influence required for adding complete noise is almost the same. The partial noise inversion mechanism avoids the computational redundancy of the traditional method that requires complete inversion to Gaussian noise, significantly reduces the time cost, and completes fast sampling. In terms of gradient computing resource consumption, the preset conditional influence arithmetic progression distribution is combined with pseudo-gradient second-order difference to achieve a complexity reduction from O(n 2 ) to O(n), reducing computational overhead.

[0054] 3. In the case where sample computing resources are the main consumption, in order to further improve the CFG accuracy corresponding to the best editing effect, the iterative parameter space expansion method is used. Compared with the method of directly using higher precision to obtain the best CFG, it can reduce a lot of resource consumption. For example, if iterative parameter expansion is not used to achieve the final target accuracy, the number of samples required for one iteration is only n.

[0055] 4. To address the "over-editing" flaw in traditional methods, where no matter how the CFG is adjusted, the edit amplitude is too large. In this case, the degree of noise addition needs to be adjusted to minimize the image change. Therefore, closed-loop control based on MSE feedback is adopted to control the execution of adaptive image editing. When the MSE of the edited result falls below a preset threshold, it is determined to be a minor image change, and adaptive fidelity control is initiated.

[0056] 5. By building a model-agnostic architecture, this method supports the direct call of any mainstream diffusion pre-trained model.

[0057] The advantages of the present invention are:

[0058] 1. By constructing an adaptive condition-guided weight optimization mechanism, we overcome the limitations of traditional methods that rely on manual experience to adjust parameters, and achieve dynamic quantitative control of the influence of text conditions. A pseudo-gradient-driven parameter search algorithm can autonomously converge to the optimal solution for editing results, significantly reducing the need for manual intervention.

[0059] 2. The partial inversion mechanism and the lightweight optimizer based on the second-order difference approximation save a lot of computing time and significantly improve the efficiency of image editing.

[0060] 3. Through iterative parameter space expansion and adaptive image editing control, the consistency of image editing results and given conditions is improved.

[0061] 4. Supports any pre-trained diffusion model (such as Stable Diffusion) without fine-tuning the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 The overall flow chart of the method of the present invention is shown in FIG1 , wherein w represents the guidance scale of the unconditional classification guide CFG, which controls the strength of the text condition. The input image x0 is unconditionally inverted and reaches x t , add the condition C and the k value corresponding to different CFGs to generate denoising diffusion, obtain the MSE after sampling, and then select the optimal CFG W according to the optimization algorithm prior , and finally get the output through conditional editing.

[0063] Figure 2 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0064] The following is a clear and complete description of the technical solution of the overall process steps of this solution.

[0065] Example 1

[0066] like Figure 1 As shown, this embodiment relates to an image editing method based on an adaptive implicit denoising diffusion model without classifier guidance, comprising the following steps:

[0067] (1) Generate multi-parameter samples by partial noise inversion: Perform partial noise inversion with controllable step size on the input image, generate multiple classifier-independent guidance parameters based on adaptive arithmetic progression, and perform conditional denoising in parallel to generate edited samples;

[0068] (2) Pseudo-gradient optimization to select the best CFG: Based on the mean square error (MSE) data pair between the sample and the original image, the MSE mutation point is located by second-order differential gradient calculation to determine the initial optimal CFG parameters;

[0069] (3) Iterative parameter space expansion optimization: construct a parameter interval with the initial optimal CFG as the center, and iteratively perform steps (1) to (2) to improve parameter accuracy;

[0070] (4) Adaptive image editing execution: Full noise inversion and conditional denoising are performed based on the final optimized CFG, and the number of inversion steps is dynamically adjusted according to the output MSE to balance the editing strength and fidelity.

[0071] Step (1) specifically includes:

[0072] Generate multi-parameter samples by partial noise inversion as described in step (1). Under normal circumstances, valid samples can also be obtained by editing the complete noise addition and conditional denoising of the image, but it takes several times longer than performing a single edit and has no advantage in time consumption compared to manual debugging. Generating multi-parameter samples by partial noise inversion can significantly reduce time consumption. The process of generating multi-parameter samples by partial noise inversion mainly includes the selection of CFG during the sampling process, obtaining samples by partial inversion and noise addition, and comparing the obtained samples with the original image to obtain CFG-MSE data pairs.

[0073] (1-1) CFG selection during the sampling process. During the image editing process, CFG is used to control the influence of the conditions. Since the sampling results need to be pseudo-gradient calculated (see step 2-1 for details), and sampling requires a high computational cost, the first CFG selection selects an adaptive arithmetic sequence that changes according to the accuracy within the conventional range r (1 to 14). In the secondary optimization stage of step 2, based on the accuracy feedback of the first sampling, an improved arithmetic sequence guided by accuracy is established in the corresponding parameter subinterval. This progressive selection strategy effectively balances computational efficiency and parameter optimization requirements. The formula for the adaptive arithmetic sequence is as follows:

[0074]

[0075] where r start , r end Represents the start and end values of the selected range, and s is the precision.

[0076] (1-2) Obtain samples by partially inverting and adding noise. In the field of diffusion model image generation, denoising is usually performed by initializing random standard Gaussian noise. Although the same initial noise contains some spatial features, different images can be generated when different conditions are added or different sampling is used. In the generation process, the incomplete denoised image, that is, the intermediate step of the denoising process, can also generate different image results by adding different conditions for denoising. Therefore, in the editing process, there is no need to invert the image to the initial standard Gaussian noise. It is only necessary to invert the image and add some noise, and then perform the conditional denoising prediction process, which can reduce the diffusion steps and greatly improve the editing efficiency. The degree of noise inversion directly affects the time of each sampling. Different from the traditional method of inverting the image to standard Gaussian noise, the degree of noise inversion is controlled according to the number of inversion steps for each sampling to reduce the computational overhead. The inversion step uses formula (2) and does not apply CFG control. The denoising step uses formula (1) and uses different CFGs. Use T skip Represents the degree of reversal, and the calculation formula is as follows:

[0077] T skip =t / T (2)

[0078] T,t represents the number of diffusion steps, T represents the number of steps required to invert the image to standard Gaussian noise, and t represents the number of inversion steps required to add partial noise.

[0079] (1-3) Compare the sample and the original image to obtain the CFG-MSE data pair. Although MSE cannot be used directly to judge the quality of an image, it is possible to find the mutation point through the change of MSE and determine the optimal CFG parameters. The calculation formula of MSE is as follows:

[0080]

[0081] Where: m*n is the total number of pixels in the image, I(i,j) and I′(i,j) represent the pixel values of the two images at each point

[0082] Step (2) specifically includes:

[0083] The best CFG is selected by pseudo gradient optimization as described in step (2). When CFG gradually improves, some features will mutate, resulting in the phenomenon of MSE mutation, and at the same time, the image quality will be reduced. Therefore, the best CFG parameters can be determined by finding the MSE mutation point. In order to reduce the consumption of computing resources, the current best CFG is selected by pseudo gradient optimization algorithm. Specifically, the CFG-MSE data obtained in step (1-3) [x i ,y i ]Use the pseudo-gradient optimization algorithm to obtain the optimal CFG. The algorithm formula is as follows:

[0084] y′ i =y i+1 -y i ,y″ i =y′ i -y′ i-1 (3)

[0085] Then select y″ i The largest CFG value is taken as the initial optimal parameter.

[0086] Unlike conventional gradient calculations, this algorithm can achieve a similar effect as obtaining gradients without performing gradient calculations, with a time complexity of only O(n). This is because calculating the second-order gradient also consumes certain additional resources, and the purpose of obtaining the gradient is only for comparison. Therefore, there is no need to obtain the true gradient. In the case of an arithmetic progression distribution, the pseudo-gradient obtained by subtraction operation can be used for comparison to achieve the same effect as gradient comparison.

[0087] Step (3) specifically includes:

[0088] Iterative parameter space expansion optimization as described in step (3). After obtaining the optimal CFG in step 2, perform interval expansion and repeat steps (1) to (2) to obtain the optimal CFG with higher accuracy. Compared with the method of directly using higher accuracy to obtain the optimal CFG, the iterative method can significantly improve the accuracy while reducing resource consumption.

[0089] Interval selection for quadratic optimization. To improve computational efficiency and parameter optimization accuracy, quadratic optimization is used to reduce sampling resources. The interval selection formula is as follows:

[0090] r=W prior +s*[-1,1] (4)

[0091] Where W prior is the optimal CFG selected by step 2, and s is the precision.

[0092] Step (4) specifically includes:

[0093] Adaptive image editing is performed as described in step (4). Adaptive image editing includes performing complete noise inversion and conditional denoising based on the final optimized CFG, and dynamically adjusting the number of inversion steps according to the output MSE to balance the editing strength and fidelity. By default, the best CFG is used to perform complete image inversion and noise addition to Gaussian space, and conditional diffusion denoising is used to complete the editing to obtain the final image with the best editing effect. Using the default method may result in excessive image modification for certain tasks, so T can be dynamically adjusted. skip To balance the editing range and fidelity. Dynamically adjust T skip The formula is as follows:

[0094] T skip =T skip (1+α·(τ-MSE)) (5)

[0095] τ is the preset threshold, and α is the adjustment coefficient.

[0096] Example 2

[0097] like Figure 2 This embodiment relates to an image editing device based on an adaptive classifier-free guided implicit denoising diffusion model, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the image editing method based on the adaptive classifier-free guided implicit denoising diffusion model of embodiment 1.

[0098] Example 3

[0099] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the image editing method based on the adaptive classifier-free implicit denoising diffusion model of the present invention is implemented.

[0100] The above embodiments are intended only to illustrate the present invention. Each step is subject to change. Any equivalent transformations and improvements based on the technical solution of the present invention should not be excluded from the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An image editing method based on an adaptive classifier-free implicit denoising diffusion model, characterized in that: The following steps are involved: (1) Generate multi-parameter samples by partial noise inversion: Perform partial noise inversion with controllable step size on the input image, generate multiple classifier-independent guided CFG parameters based on adaptive arithmetic progression, and perform conditional denoising in parallel to generate edited samples; (2) Pseudo-gradient optimization to select the best CFG: Based on the mean square error (MSE) data pair between the sample and the original image, the MSE mutation point is located by second-order differential gradient calculation to determine the initial optimal CFG parameters; (3) Iterative parameter space expansion optimization: construct a parameter interval with the initial optimal CFG as the center, and iteratively perform steps (1) to (2) to improve parameter accuracy; (4) Adaptive image editing execution: Full noise inversion and conditional denoising are performed based on the final optimized CFG, and the number of inversion steps is dynamically adjusted according to the output MSE to balance the editing strength and fidelity.

2. The image editing method based on the adaptive classifier-free implicit denoising diffusion model according to claim 1, wherein: Step (1) specifically includes: (1-1) CFG selection during sampling; the first CFG selection is to select an adaptive arithmetic progression within the conventional range r (1 to 14) according to the change of accuracy; (1-2) Obtain samples by partially inverting and adding noise; the degree of noise inversion is controlled according to the number of inversion steps for each sampling; (1-3) Compare the obtained sample with the original image to obtain the CFG-MSE data pair.

3. The image editing method based on the adaptive classifier-free implicit denoising diffusion model according to claim 2, wherein: The adaptive arithmetic progression formula used in step (1-1) according to the change of precision is: where r start , r end Represents the start and end values of the selected range, s is the precision, and the tolerance of the series is automatically adjusted by the precision. The higher the precision, the more samples are sampled and the longer the time consumed. The degree of noise inversion used in steps (1-2) is controlled by the number of inversion steps: T skip =t / T (2) Where T,t represents the number of diffusion steps, T represents the number of steps required to invert the image to standard Gaussian noise, and t represents the number of inversion steps required to add partial noise. The default value is 0.1T.

4. The image editing method based on the adaptive classifier-free implicit denoising diffusion model according to claim 1, characterized in that: Step (2) specifically includes: The best CFG is selected by pseudo gradient optimization: based on the mean square error MSE data of the sample and the original image [x i ,y i ], locate the MSE mutation point through second-order difference calculation and determine the preliminary optimal CFG parameters; The pseudo gradient optimization algorithm formula is as follows: and' i =and i+1 -and i ,and" i =y′ i -and' i-1 (3) Select y″ i The maximum CFG value is used as the optimal parameter; pseudo gradient calculation is used instead of gradient calculation to reduce the consumption of computing resources.

5. The image editing method based on the adaptive classifier-free implicit denoising diffusion model according to claim 1, wherein: Step (3) specifically includes: Iterative parameter space expansion optimization is used: a shrinking parameter interval is constructed with the initial optimal CFG as the center, and the optimization steps are iteratively performed to improve the parameter accuracy; the shrinking formula of the parameter interval is: r=W prior +s*[-1,1] (4) Where W prior is the optimal CFG selected, and s is the accuracy.

6. The image editing method based on the adaptive classifier-free implicit denoising diffusion model according to claim 1, wherein: Step (4) specifically includes: Adaptive image editing is performed: full noise inversion and conditional denoising are performed based on the final optimized CFG, and the number of inversion steps is dynamically adjusted according to the output MSE to balance editing strength and fidelity; The formula for dynamically adjusting the number of inversion steps is: T skip =T skip (1+α·(τ-MSE)) (5) τ is the preset threshold, and α is the adjustment coefficient.

7. An image editing device based on an adaptive classifier-free implicit denoising diffusion model, characterized in that: The method comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the method is used to implement the image editing method based on the adaptive classifier-free implicit denoising diffusion model according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by a processor, the image editing method based on the adaptive classifier-free implicit denoising diffusion model according to any one of claims 1 to 6 is implemented.

Images