Intelligent fusion method of image subject and new background

Abstract

The application discloses an image subject and new background intelligent fusion method, relates to the field of image fusion, and comprises the following steps: firstly, intelligently extracting a subject image from an original image; secondly, intelligently generating a background image to be embedded based on scene preset data, a text image dynamic optimization engine and a text image large model; and finally, fusing the background image and the subject image; the application ensures the extraction quality of the subject in the original image of a user, the generation quality of a new background image, the correctness and perfection of the depth of field, related light and shadow logic, position logic and interaction logic of the subject in the new background.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image fusion, in particular to an intelligent fusion method of image subject and new background. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and can not constitute the prior art.

[0003] Editing images has very wide applications. Among numerous editing functions, extracting the subject from an existing image and then embedding it into a new background image is particularly favored by users. However, existing editing tools have some deficiencies in performing the above process, which cannot better meet the user's quality requirements for image subject extraction and then embedding into a new image background.

[0004] These deficiencies mainly include: 1. Defects in subject extraction, incomplete subject extraction, fuzzy subject boundary, etc.; 2. The background material is not rich enough to meet the needs of users; 3. Poor quality of subject and background fusion, such as missing subject and background depth, subject and background light error, subject and background position logical error, subject and background content interaction logical error, etc.

[0005] Currently, for problem 2, although the existing artificial intelligence image generation (AIGC) technology based on text can provide more diverse pseudo-fusion backgrounds, due to the randomness of AIGC generated images, the quality reliability of directly using the generated images (as backgrounds) cannot meet the user's requirements, and it is more difficult to better fuse with the user-provided subject image. SUMMARY

[0006] The purpose of the present application is to provide an intelligent fusion method of image subject and new background to solve the above problems by ensuring the extraction quality of the subject in the user's original image, the generation quality of the new background image, and the correctness and perfection of the depth and related light logic, position logic and interaction logic of the subject in the new background.

[0007] The technical scheme of the present application is as follows:

[0008] An intelligent fusion method of image subject and new background, comprising:

[0009] Step S1: intelligently extracting a subject image from an original image;

[0010] Step S2: intelligently generating a pseudo-embedded background image based on scene preset data, text-to-image dynamic optimization engine and text-to-image large model;

[0011] Step S3: fusing the background image and the subject image.

[0012] Further, the step S1 comprises:

[0013] Step S11: generating an original image depth map based on the original image;

[0014] Step S12: submitting the original image depth map to a corresponding matting model based on content recognition;

[0015] Step S13: performing intelligent matting by the matting model to extract a subject image.

[0016] Further, the original image depth map comprises:

[0017] a foreground subject, a background, and a pending area;

[0018] The step S12 comprises:

[0019] Step S121: checking the integrity of the foreground subject content;

[0020] Step S122: submitting to a corresponding matting model according to whether the content contains a portrait after content recognition;

[0021] The matting model comprises:

[0022] a u2net general matting model and a u2net portrait matting model;

[0023] The step S122 comprises:

[0024] distinguishing whether the original image contains a portrait through a portrait recognition model;

[0025] If it contains a portrait, it is determined to be a portrait photo, and the original image depth map is submitted to the u2net portrait matting model for matting;

[0026] If it does not contain a portrait, it is determined to be a normal photo, and the original image depth map is submitted to the u2net general matting model for matting.

[0027] Further, the step S13 comprises:

[0028] Step S131: the matting model further distinguishes the pending area into a foreground subject and a background to generate a black and white mask image;

[0029] Step S132: extracting the foreground subject in the black and white mask image and restoring it to the original subject in the corresponding original image;

[0030] Step S133: setting the remaining background to a transparent color to generate a transparent background

[0031] Step S134: taking the image containing the original subject and the transparent background as a matting result image.

[0032] Further, the step S2 comprises:

[0033] Step S21: generating a text-to-image auxiliary image based on the matting result image;

[0034] Step S22: inputting the scene preset data and the text-to-image auxiliary image into a text-to-image large model to generate a background image to be embedded.

[0035] Further, the text-to-image auxiliary image comprises:

[0036] The matting white background image and the depth map of the matting white background image;

[0037] The step S21 comprises:

[0038] Converting the transparent background in the matting result image into a white background to generate a matting white background image;

[0039] Generating a depth map of the matting white background image based on the matting white background image.

[0040] Further, the scene preset data comprises:

[0041] The preset sample, the preset positive prompt word, the preset negative prompt word, and the background parameter and the fusion parameter;

[0042] The preset sample is input as a reference sample of the text-to-image large model;

[0043] The preset positive prompt word is input as a positive prompt word of the text-to-image large model;

[0044] The preset negative prompt word is input as a negative prompt word of the text-to-image large model;

[0045] The background parameter is input as a parameter of the text-to-image large model;

[0046] The fusion parameter is input as a control plug-in parameter of the text-to-image large model.

[0047] Further, the step S22 comprises:

[0048] Selecting a background image preset scene from a background image preset scene library as a preset sample;

[0049] Based on the background image preset scene, dynamically matching in a corresponding positive prompt word library, supplementing the positive prompt word as a preset positive prompt word;

[0050] According to the type of the original image, dynamic matching is performed in the corresponding negative optimization effect library to supplement the negative prompt words and serve as preset negative prompt words.

[0051] The corresponding background parameters and fusion parameters are determined.

[0052] The preset sample, the preset positive prompt words, the preset negative prompt words, the background parameters, the fusion parameters, the cutout white background image and the depth map of the cutout white background image are input into the text-to-image large model to generate a background image to be embedded.

[0053] Further, the step S22 comprises:

[0054] The custom positive prompt words are determined, and the corresponding background parameters and fusion parameters are determined.

[0055] The custom positive prompt words are optimized based on the text-to-image dynamic optimization engine, and the optimized custom positive prompt words are used as preset positive prompt words.

[0056] The negative prompt words are dynamically supplemented based on the text-to-image dynamic optimization engine, and the supplemented negative prompt words are used as preset negative prompt words.

[0057] The preset positive prompt words, the preset negative prompt words, the background parameters, the fusion parameters, the cutout white background image and the depth map of the cutout white background image are input into the text-to-image large model to generate a background image to be embedded.

[0058] Further, the step S3 comprises:

[0059] The background image and the cutout result image are fused by the text-to-image large model, and the fusion result is output.

[0060] Compared with the existing technology, the beneficial effects of the present application are:

[0061] 1. An image subject and new background intelligent fusion method, comprising: step S1: intelligently extracting a subject image from an original image; step S2: intelligently generating a background image to be embedded based on scene preset data, a text-to-image dynamic optimization engine and a text-to-image large model; and step S3: fusing the background image with the subject image; which ensures the extraction quality of the subject in the original image of the user, the generation quality of the new background image and the correctness and perfection of the depth of field, related light and shadow logic, position logic and interaction logic of the subject in the new background.

[0062] 2. An image subject and new background intelligent fusion method, which differentiates the processing according to whether the original image provided by the user contains a portrait after the initial extraction and re-inputs the u2net portrait matting model for the image containing a portrait and the u2net general matting model for the image not containing a portrait, so as to better distinguish the image in the undetermined area after the initial extraction into a foreground subject and a background and generate a black and white mask image; after the process, the integrity of the subject extraction is improved, and the subject boundary blur problem is optimized.

[0063] 3. An image subject and new background intelligent fusion method, which generates a depth map of a matting white background image in the image fusion process to ensure that the generated new background is fused with the matting result image, which is manifested as transition edge fusion and whole image fusion; the new background and the matting result image have consistent color tones, the matting result image does not float on the new background, and the spatial position relationship is consistent.

[0064] 4. An image subject and new background intelligent fusion method, and an AIGC intelligent image generation algorithm based on the method can be established, which can ensure the reliability of the quality of the output background image and the quality of the subject fusion background. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 It is a flowchart of an image subject and new background intelligent fusion method;

[0066] Figure 2 It is a flowchart when the first mode is adopted in step S22;

[0067] Figure 3 It is a flowchart when the second mode is adopted in step S22;

[0068] Figure 4 It is an example of an original image;

[0069] Figure 5 It is an example of an original image depth map;

[0070] Figure 6 It is an example of a black and white mask image;

[0071] Figure 7 It is an example of a matting result image;

[0072] Figure 8 It is an example of the output of an AIGC large model;

[0073] Figure 9 It is an example of a matting white background image;

[0074] Figure 10 It is an example of a depth map of a matting white background image;

[0075] Figure 11 A preset sample image corresponding to the preset scene "cherry";

[0076] Figure 12 An example image output by the text-to-image model when the first mode is adopted in step S22;

[0077] Figure 13 An example image output by the text-to-image model when the second mode is adopted in step S22. DETAILED DESCRIPTION

[0078] It should be noted that the terms "first" and "second" and the like such relational terms are used merely to distinguish one entity or operation from another entity or operation without necessarily requiring or implying any such actual relationship or order between such entities or operations. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the phrase "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0079] The features and performances of the present application will be further described in detail below in conjunction with embodiments.

[0080] Embodiment One

[0081] Please refer to Figure 1 An image subject and new background intelligent fusion method, specifically comprising the following steps:

[0082] Step S1: intelligently extracting a subject image from an original image; wherein the original image is as shown in the example; Figure 4 The example;

[0083] Step S2: intelligently generating a background image to be embedded based on scene preset data, a text-to-image dynamic optimization engine, and a text-to-image model;

[0084] Step S3: fusing the background image and the subject image.

[0085] In this embodiment, specifically, the step S1 comprises:

[0086] Step S11: generating an original image depth map based on the original image;

[0087] Step S12: submitting the original image depth map to a corresponding matting model based on content recognition;

[0088] Step S13: Intelligent matting by the matting model to extract the subject image.

[0089] In this embodiment, specifically, the original image depth map includes:

[0090] foreground subject (close to white), background (close to black) and pending area (close to gray); wherein the original image depth map is as shown in the example; Figure 5

[0091] The step S12 includes:

[0092] Step S121: Check the integrity of the foreground subject content; it should be noted that this checking process can be completed by human;

[0093] Step S122: Submit to the corresponding matting model according to whether the content contains a portrait after content recognition;

[0094] The matting model includes:

[0095] u2net general matting model and u2net portrait matting model; it should be noted that the u2net general matting model and the u2net portrait matting model are open source models and have not been substantially improved, and will not be described here;

[0096] The step S122 includes:

[0097] Distinguish whether the original image contains a portrait through a portrait recognition model; it should be noted that the portrait recognition model is an open source model and has not been substantially improved, and will not be described here;

[0098] If it contains a portrait, it is determined to be a portrait photo, and the original image depth map is submitted to the u2net portrait matting model for matting;

[0099] If it does not contain a portrait, it is determined to be a normal photo, and the original image depth map is submitted to the u2net general matting model for matting.

[0100] In this embodiment, specifically, the step S13 includes:

[0101] Step S131: The matting model further distinguishes the pending area into foreground subject and background to generate a black and white mask image; wherein the foreground subject is pure white and the background is pure black; the black and white mask image is as shown in the example; Figure 6

[0102] Step S132: Extract the foreground subject in the black and white mask image and restore it to the original subject in the corresponding original image;

[0103] ​​Step S133: set the remaining background as a transparent color to generate a transparent background;

[0104] Step S134: generate a matting result image containing the original subject and the transparent background; the matting result image is as shown in Figure 7

[0105] In this embodiment, specifically, the step S2 includes:

[0106] Step S21: generate a text-to-image auxiliary image based on the matting result image;

[0107] Step S22: input the scene preset data and the text-to-image auxiliary image into a text-to-image large model to generate a background image to be embedded; it should be noted that the text-to-image large model is a current open source model and has not been substantially improved, and thus will not be described here; for example, inputting "a lipstick resting on a white table surrounded by blooming cherry blossoms" outputs an image as shown in Figure 8 ; the step S22 can provide two generation modes: 1. directly providing a background image preset scene to the user; and 2. providing a text-to-image custom generation scene to the user.

[0108] In this embodiment, specifically, the text-to-image auxiliary image includes:

[0109] The matting white background image and the depth map of the matting white background image;

[0110] The step S21 includes:

[0111] Converting the transparent background in the matting result image into a white background to generate a matting white background image; specifically, the matting white background image is as shown in Figure 9

[0112] Generating a depth map of the matting white background image based on the matting white background image; it should be noted that the depth map of the matting white background image functions to ensure that the generated new background is fused with the matting result image, which is manifested as transition edge fusion and whole image fusion; for example, the projection of the matting result image is displayed on the new background, the color tone of the new background and the matting result image is consistent, the matting result image is not suspended on the new background, and the spatial position relationship is consistent; specifically, the depth map of the matting white background image is as shown in Figure 10

[0113] In this embodiment, specifically, the scene preset data includes:

[0114] Preset sample, preset positive prompt word, preset negative prompt word, and background parameter and fusion parameter;​​​

[0115] The preset sample image serves as the reference sample input for the large-scale text image model;

[0116] The preset positive prompt words serve as the positive prompt word input for the text-to-image large model;

[0117] The preset negative prompt words serve as the negative prompt word input for the text-to-image large model;

[0118] The background parameters are used as input parameters for the large-scale model of the text image;

[0119] The fusion parameters are used as input parameters for the control plugin of the Wenshengtu large model.

[0120] In this embodiment, for details, please refer to the relevant documentation. Figure 2 When step S22 adopts the first mode, step S22 includes:

[0121] Select a preset background image from the preset background image scene library as the preset sample image; for example, select the preset scene "cherry" from the preset background image scene library as the preset sample image. Figure 11 As shown;

[0122] Based on a preset scene in the background image, the system dynamically matches and supplements positive prompt words from the corresponding positive prompt word dictionary, which are then used as preset positive prompt words. For example, for the preset scene "cherry", the system dynamically supplements positive prompt words by matching the corresponding positive prompt word dictionary according to four categories: "photography type", "composition", "image quality" and "image lighting effect", as follows:

[0123] “Resting on a white table surrounded by blooming cherry blossoms,clean background,trending on instagram,trending on tumblr,HDR,4k,8k”;

[0124] According to the type of the original image, dynamic matching is performed in the corresponding negative optimization effect library to supplement the negative prompt words and serve as preset negative prompt words; that is, according to the type of the original image, “portrait photo” or “ordinary photo”, the corresponding negative optimization effect library is dynamically matched to dynamically supplement the full negative prompt words; it should be noted that when the original image type is “portrait photo”, the preset negative prompt words include 5 categories of words: general, face, muscle and bone, limbs; when the original image type is “ordinary photo”, the preset negative prompt words include 3 categories of words: general, shape, proportion; the corresponding negative library is matched to supplement the negative prompt words as the negative prompt input of the text-to-image large model; for example, the negative prompt words of the preset scene “cherry” are as follows:

[0125] “EasyNegative, bad anatomy, extra legs, extra arms, extra fingers, poorly drawn hands, poorly drawn feet, disfigured, out of frame, tiling, bad art, deformed, mutated, nsfw, (worst quality, low quality: 1.4), (realistic, lip, nose, tooth, rouge, lipstick, eyeshadow: 1.0), (dusty sunbeams: 1.0), (abs, muscular, rib: 1.0), (depth of field, bokeh, blurry: 1.4), (greyscale, monochrome: 1.0), text, title, logo, signature, mutated hands, missing fingers, extra fingers, missing arms, extra arms”;

[0126] Determine the corresponding background parameters and fusion parameters;

[0127] It should be noted that the background parameters are input as parameters of the text-to-image large model, and the function is to ensure the generation quality of the generated quasi-embedded background image; the background parameters include:

[0128] Sampling method, Sampling steps, CFG Scale, Denoising strength;

[0129] Wherein:

[0130] Sampling method represents the sampling used for generating the background image, ensuring the accuracy of the generated background image;

[0131] Sampling steps represent the frequency of optimization for generating the background image, maintaining the output of the generated background image with better optimization results;

[0132] CFG Scale represents the associative space for generating the background image, ensuring that the generated background image has certain automatic associative content;

[0133] Denoising strength represents the emphasis of generating the background image, ensuring consistency between the generated background image and the description of the positive prompt word;

[0134] Fusion parameters are control plug-in parameters for the text-to-image large model, which ensure the fusion degree of the background image and the cutout result image. The fusion parameters include:

[0135] Mask blur, Mask content, Masked pixels;

[0136] Among them:

[0137] Mask blur represents the blur value of the edge of the fusion of the generated background image and the cutout result image, ensuring the tightness of the edge fusion.

[0138] Mask content represents the fusion degree of the overall effect of the fusion of the generated background image and the cutout result image, ensuring the consistency of the color tone and light, etc.

[0139] Masked pixels represent the transition pixel value of the fusion growth factor of the generated background image and the cutout result image, which will generate the projection of the fusion, etc.

[0140] The preset sample, the preset positive prompt word, the preset negative prompt word, the background parameter, the fusion parameter, the cutout white background image, and the depth map of the cutout white background image are input into the text-to-image large model to generate a background image to be embedded.

[0141] In this embodiment, specifically, the step S3 includes:

[0142] The background image and the cutout result image are fused by the text-to-image large model, and the fusion result is output. In this embodiment, specifically, the output result of the preset scene "cherry" is as shown in Figure 12 .

[0143] In this embodiment, specifically, please participate Figure 3 When the step S22 adopts the second mode, the step S22 includes:

[0144] custom forward prompt words, and determine corresponding background parameters and fusion parameters;

[0145] based on the text-to-image dynamic optimization engine, the custom forward prompt words are optimized, and the optimized custom forward prompt words are used as preset forward prompt words;

[0146] based on the text-to-image dynamic optimization engine, the negative prompt words are dynamically supplemented, and the supplemented negative prompt words are used as preset negative prompt words;

[0147] The preset forward prompt words, the preset negative prompt words, the background parameters, the fusion parameters, the cutout white background image, and the depth map of the cutout white background image are input into the text-to-image large model to generate a background image to be embedded.

[0148] It should be noted that the text-to-image dynamic optimization engine is used to optimize the corresponding parameters input into the text-to-image large model to ensure high-quality output results.

[0149] Specifically, the dynamic optimization logic for optimizing the custom forward prompt words is: by constructing a word library with positive optimization effect for text-to-image, the corresponding word library is matched according to four categories of “photography type”, “composition”, “image quality”, and “image light effect” to dynamically complete the forward prompt words.

[0150] For example, the user inputs “On the stone, Behind is the sunset and the coastline, large area of sky”;

[0151] The text-to-image dynamic optimization engine outputs “On the stone, Behind is the sunset and the coastline, large area of sky, product shot, center composition, best quality, studio lighting”;

[0152] Specifically, the dynamic supplement of the negative prompt words includes:

[0153] According to the type of the original image, the negative optimization effect word library corresponding thereto is dynamically matched to supplement the negative prompt words and serve as preset negative prompt words; that is, according to the type of the original image, “portrait photo” or “ordinary photo”, the negative optimization effect word library corresponding thereto is dynamically matched to dynamically supplement the full negative prompt words; it should be noted that when the original image type is “portrait photo”, the preset negative prompt words include 5 categories of words: general, face, muscle and bone, limbs; when the original image type is “ordinary photo”, the preset negative prompt words include 3 categories of words: general, shape, proportion; the negative prompt words supplemented by matching the corresponding negative word library serve as the negative prompt word input of the text-to-image large model;

[0154] For example: the user's matting result image is a “portrait photo”, and the text-to-image dynamic optimization engine outputs the negative prompt words “EasyNegative, bad art, deformed, mutated, nsfw, (worst quality, low quality: 1.4), (depth of field, bokeh, blurry: 1.4), (realistic, lip, nose, tooth, rouge, lipstick, eyeshadow: 1.0), (dusty sunbeams: 1.0), (abs, muscular, rib: 1.0), mutated hands, missing fingers, extra fingers, missing arms, extra arms, text, title, logo, signature”.

[0155] In this embodiment, specifically, the step S3 comprises:

[0156] The background image is fused with the matting result image by the text-to-image large model, and a fusion result is output; in this embodiment, specifically, the preset scene “cherry” outputs a result as shown in Figure 13 .

[0157] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application.

[0158] This Background section is intended to provide a general overview of the context of the application, the work of the current named inventors, to the extent the work is associated with the present application, and the work of others in the field of the application to the extent the work is associated with the present application, and to the extent the work described in this section is associated with the present application, neither expressly nor impliedly, is admitted to be prior art to the present application.

Claims

1. A method for intelligently fusing an image subject with a new background, characterized in that, include: Step S1: Intelligently extract the subject image from the original image; Step S2: Based on the scene preset data, the text image dynamic optimization engine, and the text image large model, intelligently generate the background image to be embedded; Step S3: Blend the background image with the subject image; Step S1 includes: Step S11: Generate the original image depth map based on the original image; Step S12: Based on content recognition, submit the original image depth map to the corresponding matting model; Step S13: The image is intelligently cut out using the image cutout model to extract the main image; Step S2 includes: Step S21: Generate a text-based auxiliary image based on the cutout result image; Step S22: Input the scene preset data and the text image auxiliary image into the text image large model to generate the background image to be embedded; The original image depth map includes: Foreground subject, background, and area to be determined; Step S13 includes: Step S131: The image matting model further divides the area to be determined into foreground subject and background, generating a black and white mask image; Step S132: Extract the foreground subject from the black and white mask image and restore it to the original subject in the corresponding original image; Step S133: Set the remaining background to transparent to generate a transparent background; Step S134: Use the image containing the original subject and the transparent background as the cutout result image; The textural auxiliary image includes: Image with white background cutout and depth map of image with white background cutout; Step S21 includes: Convert the transparent background in the cutout result image to a white background to generate a cutout image with a white background. Based on the cut-out white background image, generate a depth map of the cut-out white background image; The preset data for the scenario includes: Preset sample images, preset positive prompts, preset negative prompts, as well as background parameters and blending parameters; The preset sample image serves as the reference sample input for the large-scale text image model; The preset positive prompt words serve as the positive prompt word input for the text-to-image large model; The preset negative prompt words serve as the negative prompt word input for the text-to-image large model; The background parameters are used as input parameters for the large-scale model of the text image; The fusion parameters are used as input parameters for the control plugin of the Wenshengtu large model.

2. The intelligent fusion method for an image subject and a new background according to claim 1, characterized in that, Step S12 includes: Step S121: Check the completeness of the foreground content; Step S122: After content recognition, submit the image to the corresponding cutout model based on whether it contains a human image; The image matting model includes: u2net general image matting model and u2net portrait image matting model; Step S122 includes: Using a facial recognition model, it can distinguish whether the original image contains a human image; If it contains a human figure, it is identified as a human figure photo, and the original image depth map is submitted to the u2net human figure matting model for matting. If the image does not contain a human figure, it is considered a regular photo, and the original image depth map is submitted to the u2net general image matting model for matting.

3. The intelligent fusion method for an image subject and a new background according to claim 2, characterized in that, Step S22 includes: Select a background image preset scene from the background image preset scene library as the preset sample image; Based on the background image and a preset scene, dynamic matching is performed in the corresponding positive prompt word library to supplement positive prompt words and use them as preset positive prompt words; Based on the type of the original image, dynamic matching is performed in the corresponding negative optimization effect word library to supplement negative prompt words and use them as preset negative prompt words; Determine the corresponding background parameters and fusion parameters; Input the preset sample image, preset positive prompt words, preset negative prompt words, background parameters, fusion parameters, cut-out white background image, and depth map of the cut-out white background image into the Wenshengtu large model to generate the background image to be embedded.

4. The intelligent fusion method for an image subject and a new background according to claim 1, characterized in that, Step S22 includes: Customize positive prompts and determine the corresponding background and blending parameters; Based on the text-based image dynamic optimization engine, the custom positive prompt words are optimized, and the optimized custom positive prompt words are used as preset positive prompt words; Based on the text-based image dynamic optimization engine, negative prompt words are dynamically supplemented, and the supplemented negative prompt words are used as preset negative prompt words; Input the preset positive prompts, preset negative prompts, background parameters, fusion parameters, cutout white background image, and depth map of the cutout white background image into the Wenshengtu large model to generate the background image to be embedded.

5. The intelligent fusion method for an image subject and a new background according to claim 3 or 4, characterized in that, Step S3 includes: The Wenshengtu large model merges the background image with the cutout result image and outputs the merged result.

Images