STYLE MATCHING USING LAYER-WISE MASKS

Notice of Pre-AIA or AIA Status The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . Response to Amendment The Amendment filed 6/18/2026 has been entered. Claims 1-20 remain pending in the application. Response to Arguments Applicant’s arguments, see pages 9-12 of Remarks/Arguments, filed 6/18/2026, with respect to how the amended claim 1, 6, 12, 17-19, the rejection of claim 1 under 35 U.S.C 102(a)(1) have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Pramanik (Masked Multi-Query Slot Attention for Unsupervised Object Discovery) and further in view of Karras (A Style-Based Generator Architecture for Generative Adversarial Networks). Claim Rejections - 35 USC § 103 In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status. The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. Claims 1-3, 5, 6, 10-13, 16-20 are rejected under 35 U.S.C. 103 as being anticipated by Rahman (Visual Concept-driven Image Generation with Text-to-Image Diffusion Model) in view of Pramanik (Masked Multi-Query Slot Attention for Unsupervised Object Discovery) and further in view of Karras (A Style-Based Generator Architecture for Generative Adversarial Networks). Regarding claim 1, Rahman teaches: A method comprising: obtaining a content prompt and a style prompt, wherein the content prompt includes an object and the style prompt includes a style element; (Rahman 1. Introduction “Photorealistic image synthesis, particularly with text-to-image diffusion models (e.g., Stable Diffusion [42]), has provided a great boost to content creation. Images of diverse and imaginative scenarios such as “A dog in space” can now be easily generated. While such approaches can faithfully generate many instances of the generic subject, … To introduce additional controllability, for example, concerning subject appearance or style … approaches have focused on controlling generation with sample images depicting concepts of interest … images of the target subject to extract the single subject-specific token across these images, e.g., using inversion techniques, and then introduce this special token in prompts in conjunction with the regular text to generate desired content” Note: Rahman teaches that multiple inputs, or prompts, are accepted and used to generate images. These prompts can contain subjects/objects of interest as well as abstract concepts like an appearance or style. This teaches style and content prompts. It is also taught that these prompts can come in the form of images or text.) encoding the content prompt and the style prompt to obtain a content embedding and a style embedding, respectively; (Rahman 2. Related Work “by employing an encoder to discern and generate the latent representation corresponding to a given image” Rahman 3. Methodology “Diffusion models encode the textual guidance for accurate image generation through cross-attention between the textual tokens” Note: Rahman teaches that its input image prompt and text prompts are encoded. As it was taught previously that these prompts can be content and style prompts this teaches the embedding of the content and style prompts.) applying a content mask to the content embedding and a style mask to the style embedding (Rahman 1 Introduction “extract the single subject-specific token across these images, e.g., using inversion technique” PNG media_image1.png 688 1448 media_image1.png Greyscale PNG media_image2.png 252 590 media_image2.png Greyscale 3. Methodology “Optimization. To ensure that the tokens learned correspond to the target concept, we introduce binary masks to constrain the loss during prompt optimization and update the tokens under the mask … In the second stage, given good initializations of the tokens and their masks, we perform an alternating update of the tokens given the masks and of the masks themselves. … The overall optimization is outlined in Fig. 2. In this example, tokens [v1], [v2], and [v3] are assigned to the two subjects and a background of the second image respectively. With a good initialization of the token embeddings, we obtain the initial masks of the three concepts by averaging the attention maps over certain timesteps. Following this, we apply LMask and Lattn to update the tokens constrained by the given masks. With the newly updated tokens, we update the model weights.” PNG media_image3.png 98 536 media_image3.png Greyscale Note: Rahman teaches that its tokens are specific embeddings which are chosen because they describe prompt aspects of particular importance, like a subject or style. A process where a mask is applied to the tokens/embeddings is taught to result in weighted embeddings. The process involves applying a mask to optimize/alter the tokens with respect to their potential error/loss, this could also be called weighting them. This results in optimized tokens, or weighted embeddings, which are now ready to be used in the image generation process.) and generating, using the image generation model, a synthetic image by processing the content embedding and the style embedding, wherein the synthetic image depicts the object from the content prompt and the style element from the style prompt. (Rahman 3. Methodology “We define a concept as a visually illustrated instance of an object/subject or a well examplified pattern. Note that such concepts often entangle content and style.” Rahman 1. Introduction “Our approach performs token optimization while simultaneously optimizing corresponding (latent) masks such that the token embedding best represents the disentangled concept. • We show through novel combinations of subjects and interactive environments that our approach works well in practice and can generate realistic scenarios with custom subjects interacting in complex scenes and environments” Note: Rahman teaches that from its optimized tokens (weighted embeddings) a complex image can be generated according to the disentangled concepts the tokens describe from the prompt. As seen in the first provided citation and as established earlier, these concepts can be content or style, teaching the ability to generate an image using weighted content and style embeddings from content/style prompts.) While Rahman teaches the applying of a style mask and content mask to style and content embeddings, and leveraging the content and style embeddings to generate a synthetic image with the content and style of the respective prompts it does not teach applying a mask to the embeddings to obtain multiple weighted embeddings from a single input embedding. This idea, to obtain multiple weighted embeddings from a single input embedding by applying a mask is taught by Pramanik which teaches: applying a mask to the embedding to obtain a first weighted embedding and a second weighted embedding (Pramanik 3A “We devise our masking strategy with inspiration from the MAEs literature [13]. Let us consider an image x with spatial resolution H × W pixels reshaped into N = H·W P2 non overlapping patches and with P the patch size in pixels. Each patch token extracted from ViT-S/16 pre-trained with DINO is represented by a vector p ∈ RDfeats, with dimensionality Dfeats. We compute the mean (1 Dfeats Dfeats i=1 pi) for each patch token to obtain a rough estimate of the saliency of the patch w.r.t. to the entire image. We then sort the patches in increasing order and filter out the top m% of the patches with the highest means. We mask these patches by replacing the patch tokens with a zero vector ⃗0. Implementation-wise, we use the arg sort operator from standard libraries to obtain the indices in increasing order. We then filter out the last m% of the indices to mask them.” 3D “We follow a process similar to previous works [19], [42] to normalize the patch tokens across the slots using the softmax operation. Finally, we use these weights to produce the reconstructed patch tokens as a weighted sum of all the decoded slot representations. We use a mean squared error (L2) loss to optimize our network.” PNG media_image4.png 696 1030 media_image4.png Greyscale Note: Pramanik teaches that from an input image x multiple “patch tokens” are extracted with ViT-s/16, it is known in the art that the patch tokens obtained from a vision transformer (ViT) are embeddings. Pramanik teaches the patches are masked so that they can be provided to the model to be normalized with the softmax operation to tune the weights of the patch tokens, and to then use the updated weights. Thus, Pramanik teaches making multiple weighted embeddings from an input embedding by applying a mask.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Pramanik where a model that generates a synthetic image from a style and content embedding that have respective style and content masks applied to them makes a first and second weighted embedding from applying the mask to the embeddings to get first and second weighted style and content embeddings. There are several reasons that would motivate one to do so, the described models of Rahman and the present invention seek to generate a synthetic image with the content and style of their input images; to do so the model must have a good understanding of its input content. Different parts of images contain different pieces of data, to help the model understand all present information without focusing too much on one area a mask could be used to make multiple embeddings with different weights to allow the model to learn more from a single image. While Rahman has been established to teach applying a content and style mask to content and style embeddings to obtain a first and second content and style embedding it is not detailed that different layers of the model will receive the different first and second weighted embeddings. A model that generates a synthetic image from a first layer and second layer that receive first and second embeddings respectively is taught by Karras which teaches wherein the first weighted content embedding corresponds to a first layer of an image generation model, and wherein the second weighted content embedding corresponds to a second layer of the image generation model;(Karras 2 “Given a latent code z in the input latent space Z, a non-linear mapping network f:Z→W first produces w∈W (Figure 1b, left). For simplicity, we set the dimensionality of spaces to 512, and the mapping f is implemented using an 8-layer MLP, a decision we will analyze in Section 4.1. Learned affine transformations then specialize w to styles y = (ys, yb) that control adaptive instance normalization (AdaIN) [26], [16], [20], [15] operations after each convolution layer of the synthesis network g. The AdaIN operation is defined as AdaIN(xi,y)=ys,ixi−μ(xi)σ(xi)+yb,i where each feature map xi is normalized separately, and then scaled and biased using the corresponding scalar components from style y. Thus the dimensionality of y is twice the number of feature maps on that layer.” PNG media_image5.png 616 426 media_image5.png Greyscale Note: Karras teaches that an embedding w will be obtained from an image encoded in latent space with latent code z. Once embeddings w are made they are given to different layers as seen above in Fig. 1 where the synthesis network g receives w. It does not receive the embedding w directly however, Karras 2 teaches that y, referred to as ‘styles’, is specialized after each convolution layer to be specifically input to the adaptive instance normalization (AdaIN) function. As w is put through an affine transformation to become y, y is implicitly a weighted embedding as the embedding w must have its weights changed to be transformed into the new space. Furthermore, the AdaIN function in each individual layer uses xi, xi in this context refers to the feature map of given layer. Thus, the embeddings used by each layer are weighted embeddings obtained from the same original embedding that are further transformed by the AdIN function, meaning each layer receives its own different weighted embedding.) and generating, using the image generation model, a synthetic image by processing the first weighted embedding using the first layer, and by processing the second weighted embedding using the second layer, (Karras 2 ‘Finally, we provide our generator with a direct means to generate stochastic detail by introducing explicit noise inputs. These are single-channel images consisting of un correlated Gaussian noise, and we feed a dedicated noise image to each layer of the synthesis network. The noise image is broadcasted to all feature maps using learned per feature scaling factors and then added to the output of the corresponding convolution, as illustrated in Figure 1b. The implications of adding the noise inputs are discussed in Sections 3.2 and 3.3.” PNG media_image6.png 692 500 media_image6.png Greyscale Note: Karras teaches that the feature maps of the layers are what will produce the final generated image, as shown previously these feature maps xi are determined based on the different weighted embeddings y, or syle, that each layer recieves. As the weighted embeddings directly inform and define the feature map that will produce the final generated synthetic image, Karras teaches leveraging a first and second weighted embedding of a first and second layer to generate a synthetic image.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Karras where providing different layers of an image generation model with differently weighted embeddings to produce a synthetic image: obtains its weighted embeddings from content and style embeddings to provide to each layer respectively resulting in differently weighted content and style embeddings, and the multiple weighted content and style embeddings are obtained by applying a mask to the embeddings. There are several reasons that would motivate one to do so, one benefit of Karras’s approach is that it allows the isolation of certain features within an embedding to be brought out more by individual layers, as different layers will have different weighted embeddings. Karras leverages this to better impart the style of its input images onto its synthetic images. This is the same goal of Rahman and the present invention, except an input content prompt is also used. If one wished for their model to be able to accurately impart specific features from input prompts to generated images the described method of Karras where different layers receive different weighted embeddings could be leveraged. Regarding claim 2, Rahman teaches: The method of claim 1, wherein: the content prompt comprises a text prompt and the style prompt comprises an image prompt. (Rahman 1. Introduction “Photorealistic image synthesis, particularly with text-to-image diffusion models (e.g., Stable Diffusion [42]), has provided a great boost to content creation. Images of diverse and imaginative scenarios such as “A dog in space” can now be easily generated. While such approaches can faithfully generate many instances of the generic subject. To introduce additional controllability, for example, concerning subject appearance or style … approaches have focused on controlling generation with sample images depicting concepts of interest … images of the target subject to extract the single subject-specific token across these images, e.g., using inversion techniques, and then introduce this special token in prompts in conjunction with the regular text to generate desired content” Rahman 3. Methodology “Diffusion models encode the textual guidance for accurate image generation through cross-attention between the textual tokens” PNG media_image1.png 688 1448 media_image1.png Greyscale Rahman 3. Methodology “In the second stage, given good initializations of the tokens and their masks, we perform an alternating update of the tokens given the masks and of the masks themselves. Using Eq. (3) and Eq. (4), we first optimize the prompt given the masks for each concept at the iteration … The overall optimization is outlined in Fig. 2. In this example, tokens [v1], [v2], and [v3] are assigned to the two subjects and a background of the second image respectively.” Note: Rahman states that the token optimization process with masks optimizes the prompt, citing Fig. 2 with tokens v1-v3 as an example. This teaches image prompts as the tokens v1-v3 which originate from a prompt that is also optimized originate from an image. In the previous Rahman citation, it is also directly stated that text prompts can be leveraged. In the first provided citation Rahman teaches that its image prompts and text prompts can be leveraged together, in the example Rahman gives the text prompt is said to describe content like “a dog in space” and the image prompt allows for things like “subject appearance or style” to be conveyed. This teaches style prompts from an image prompt and content prompts from a text prompt.) Regarding claim 3, The method of claim 1, wherein obtaining the content prompt and the style prompt comprises: obtaining a text prompt; and dividing the text prompt to obtain the content prompt and the style prompt. ( PNG media_image7.png 216 418 media_image7.png Greyscale PNG media_image8.png 204 908 media_image8.png Greyscale Rahman 3. Methodology “We define a concept as a visually illustrated instance of an object/subject or a well examplified pattern. Note that such concepts often entangle content and style. Diffusion models encode the textual guidance for accurate image generation through cross-attention between the textual tokens and the corresponding region in the image … In a prompt y with k tokens of interest, diffusion loss is applied to the corresponding pixels which in turn are obtained from the binary mask of each token.” Note: In the above images Rahman gives the example of two prompt images being used with a text prompt. The text prompt references content from the image prompt, in this case a character and a stylistic background/setting, where the character is in the setting. To generate this image Rahman must divide the text prompt, this is done when the whole text prompt is split into “textual tokens”. Specific text parts of the prompt, like ones that describe v1 and v3, are separated out and viewed as tokens in order to associate them with the proper part of the image prompt they are describing.) Regarding claim 5, Rahman teaches: The method of claim 1, wherein: the style embedding is retrieved from a style embedding library. (Rahman Fig. 2, 3. Methodology “We first obtain initial concept token embeddings by either initializing them from CLIP or by performing inversion without masking to optimize the tokens for a fixed number of iterations. This provides us with good initial token embeddings.” Note: Rahman details how embeddings can be initialized via CLIP embeddings, leveraging the fact that the model is pretrained and has a large knowledge base at its disposal. No direct description of an “embedding library” is detailed in the specifications, we interpret it to mean a knowledge base or collection of embeddings which we are able to pull from. As obtaining initial embeddings from CLIP meets this definition, Rahman teaches the use of a style embedding library.) Regarding claim 6, Rahman teaches: The method of claim 1, further comprising: obtaining a plurality of content masks and a plurality of style masks; model, wherein the content mask is one of the plurality of content masks, and wherein the style mask is one of the plurality of style masks, and applying the plurality of content masks to the content embedding and the plurality of style masks to the style embedding respectively ( PNG media_image1.png 688 1448 media_image1.png Greyscale PNG media_image9.png 596 1422 media_image9.png Greyscale Note: This claimed method of splitting the prompt’s embeddings into individual embeddings so that they can be handled as individual layers where multiple masks per embedding/layer are made to determine how they should be weighted is taught. In Fig. 2 above the multiple masks for each different layer (split up by which token v[1], v[2], or v[3] is being weighted) are made teaching a plurality of masks for a plurality of image layers. While these layers initially take in the unweighted embeddings in Fig. 3 we can see that the layer’s process of determining the weights with a plurality of masks is iterative, and the layer will accept the optimized token/weighted embedding again to further refine its mask making, allowing it to better optimize the tokens. As seen from Fig. 3 many different versions of the given masks will be made over hundreds of time steps during an optimization process where the the mask is optimized. As these are the content and style masks Rahman teaches a plurality of both.) While Rahman teaches that masks can be applied to the embeddings it makes it does not teach that by applying the masks the embeddings become weighted. This is taught by Pramanik which teaches obtaining a plurality of masks; model, wherein the mask is one of the plurality masks, and applying the plurality of masks to the embedding obtain a plurality of weighted embeddings (Pramanik 3A “We devise our masking strategy with inspiration from the MAEs literature [13] Let us consider an image x with spatial resolution H × W pixels reshaped into N = H·W P2 non overlapping patches and with P the patch size in pixels. Each patch token extracted from ViT-S/16 pre-trained with DINO is represented by a vector p ∈ RDfeats, with dimensionality Dfeats. We compute the mean ( 1 Dfeats Dfeats i=1 pi) for each patch token to obtain a rough estimate of the saliency of the patch w.r.t. to the entire image. We then sort the patches in increasing order and filter out the top m% of the patches with the highest means. We mask these patches by replacing the patch tokens with a zero vector ⃗0. Implementation-wise, we use the arg sort operator from standard libraries to obtain the indices in increasing order. We then filter out the last m% of the indices to mask them.” 3D “We follow a process similar to previous works [19], [42] to normalize the patch tokens across the slots using the softmax operation. Finally, we use these weights to produce the reconstructed patch tokens as a weighted sum of all the decoded slot representations. We use a mean squared error (L2) loss to optimize our network.” PNG media_image4.png 696 1030 media_image4.png Greyscale Note: Pramanik teaches that from an input image x multiple “patch tokens” are extracted with ViT-s/16, it is known in the art that the patch tokens obtained from a vision transformer (ViT) are embeddings and the final patches used are determined by masking. The masking process determines the final patches to be used. Pramanik teaches the masked patches patches are normalized with the softmax operation to tune the weights of the patch tokens, and to then use the updated weights. Thus, Pramanik teaches the applying of a mask to a plurality of embeddings to obtain weighted embeddings, which it uses to provide to the model.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Pramanik where a model that has a plurality of content and style embeddings and applies a plurality of masks to the embeddings does so to obtain weighted embeddings. There are several reasons that would motivate one to do so, the described models of Rahman and the present invention seek to generate a synthetic image with the content and style of their input images; to do so the model must have a good understanding of its input content. Different parts of images contain different pieces of data, to help the model understand all present information without focusing too much on one area a mask could be used to make multiple embeddings with different weights to allow the model to learn more from its images. While Rahman teaches a plurality of masks the layers it assigns masks to are not layers of the image generation model. A plurality of layers of a generation model that receive a plurality of weighted embeddings is taught by Karras which teaches wherein the first layer and the second layer belong to the plurality of layers; obtain a plurality of weighted content embeddings wherein each of the plurality of layers of the image generation model takes a corresponding weighted embedding as input (Karras 2 “Given a latent code z in the input latent space Z, a non-linear mapping network f:Z→W first produces w∈W (Figure 1b, left). For simplicity, we set the dimensionality of spaces to 512, and the mapping f is implemented using an 8-layer MLP, a decision we will analyze in Section 4.1. Learned affine transformations then specialize w to styles y = (ys, yb) that control adaptive instance normalization (AdaIN) [26], [16], [20], [15] operations after each convolution layer of the synthesis network g. The AdaIN operation is defined as AdaIN(xi,y)=ys,ixi−μ(xi)σ(xi)+yb,i where each feature map xi is normalized separately, and then scaled and biased using the corresponding scalar components from style y. Thus the dimensionality of y is twice the number of feature maps on that layer.” PNG media_image5.png 616 426 media_image5.png Greyscale Note: Karras teaches that an embedding w will be obtained from an image encoded in latent space with latent code z. Once embeddings w are made they are given to different layers as seen above in Fig. 1 where the synthesis network g receives w. It does not receive the embedding w directly however, Karras 2 teaches that y, referred to as ‘styles’, is specialized after each convolution layer to be specifically input to the adaptive instance normalization (AdaIN) function. As w is put through an affine transformation to become y, y is implicitly a weighted embedding as the embedding w must have its weights changed to be transformed into the new space. Furthermore, the AdaIN function in each individual layer uses xi, xi in this context refers to the feature map of given layer. Thus, the embeddings used by each layer are weighted embeddings obtained from the same original embedding that are further transformed by the AdIN function, meaning each layer receives its own different weighted embedding. As seen from Fig. 1 there exists a first and second layer that are of a plurality of layers as the descrption teaches there are 18 layers of the synthesis network g. Karras Fig. 5 and 2, cited in the rejection of claim 1 above, teach that the described model is an image generation model that will produce an image with the style of its input prompt images.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Karras where providing different layers of an image generation model with differently weighted embeddings to produce a synthetic image: obtains its weighted embeddings from content and style embeddings to provide to each layer respectively resulting in differently weighted content and style embeddings, and the multiple weighted content and style embeddings are obtained by applying a mask to the embeddings. There are several reasons that would motivate one to do so, one benefit of Karras’s approach is that it allows the isolation of certain features within an embedding to be brought out more by individual layers, as different layers will have different weighted embeddings. Karras leverages this to better impart the style of its input images onto its synthetic images. This is the same goal of Rahman and the present invention, except an input content prompt is also used. If one wished for their model to be able to accurately impart specific features from input prompts to generated images the described method of Karras where different layers receive different weighted embeddings could be leveraged. Regarding claim 10, 16, Rahman teaches: The method of claim 1, further comprising: identifying a diffusion timestep, wherein the content mask and the style mask are based on the diffusion timestep. (Rahman 3. Methodology “These attention maps are averaged across all layers of the U-Net to get the response for the regions attended by the target tokens. Note that the latent diffusion objective and hence the attention map are a function of the timestep t. Therefore, the quality of the mask would in general differ depending on one or a range of timesteps … we obtain the initial masks of the three concepts by averaging the attention maps over certain timesteps.” Note: Rahman teaches that its diffusion is a function of its timesteps, and that the timesteps determine the makeup of the masks.) Regarding claim 11, Rahman teaches: The method of claim 1, wherein: the content embedding and the style embedding are provided to a cross-attention layer of the image generation model. (Rahman 3. Methodology “Mask extraction. Following [50], we leverage the cross-attention maps to generate the mask of the target concept. Given a conditioning prompt y ∈ RN×d where N is the number of tokens in the prompt and d is the embedding dimension, a cross-attention layer with key (K), query (Q) and value (V) is given by, A=softmax (QKT /√d )*V.” Note: Rahman teaches that tokens, or individual embeddings, are input to a cross attention portion of the model to a cross attention layer.) Regarding claim 12, Rahman teaches: Claim 12 is rejected with the rationale provided for claims 1 and 6 as their content encompasses the entirety of claim 12’s scope, as claim 12 is mostly a verbatim copy of claims 1 and 6. Regarding claim 13, Rahman teaches: The method of claim 12, wherein: the image generation model comprises a diffusion model. (Rahman 3. Methodology citations provided previously teach the use of a diffusion model.) Regarding claim 17, Rahman teaches: An apparatus comprising: at least one processor; at least one memory storing instructions executable by the at least one processor; (Rahman Fig. 3 and 3. Methodology show and describe the machine learning models and stored parameters being used and ran for testing, teaching implicitly the use of some processor and memory to obtain performance results.) As the entirety of claim 17 is encompassed by claims 1 and 6, aside from the apparatus preamble of the claim which has been addressed above, the remaining body portion of claim 17 is rejected under the same rationale as claims 1 and 6 as no new content is introduced. Regarding claim 18, Rahman teaches: The apparatus of claim 17, further comprising: a text encoder configured to generate the first content embedding and the second content embedding, wherein the text encoder comprises a text adapter trained based on the image generation model. (Note: An “adapter” used in the encoding process which is based on the image generation model is vague, we refer to the specification’s Figures 5A and 5B PNG media_image10.png 712 1024 media_image10.png Greyscale These clarify that to produce content/style embedding a text/image encoder first encodes the input prompts, and the adapter then handles the encoder outputs to produce its final embedding. The style/content embeddings have already been taught to be analogous to Rahman’s tokens which are individual embeddings that can describe content/styles. Rahman has been shown to teach multiple content embeddings, providing an example where in Fig. 2 of 2 content embeddings, two characters, being identified. Rahman in 3. Methodology clearly shows how through mask creation, latent diffusion, and other steps all directly stated to be based on the image generation model, U-Net, a final usable token, or content/style embedding, is produced after the initial encoding is created.) Regarding claim 19, Rahman teaches: The apparatus of claim 17, further comprising: an image encoder configured to generate the first style embedding and the second style embedding, wherein the image encoder comprises an image adapter trained based on the image generation model. (Rahman 2, and Fig. 2, cited previously, clearly teach that an input image have a style embedding made for it via an image encoder. In the example provided by Fig. 2 only one style embedding is used, however as Rahman teaches that a plurality of tokens can be used for multiple inputs, which may depict either content or style, Rahman teaches the ability to generate a first and second style embedding.) Regarding claim 20, The apparatus of claim 17, further comprising: a mask component configured to generate a plurality of content masks and a plurality of style masks corresponding to a plurality of layers of the image generation model. (The mask components and masks corresponding to a plurality of layers is taught in Rahman Fig.2, and Fig. 3, in claim 6. Rahman 3. Methodology, cited previously, teaches that the generation of masks corresponds to a plurality of layers of the image generation model, U-Net) Claim 7, 5, are rejected under 35 U.S.C. 103 as being unpatentable over Rahman (Visual Concept-driven Image Generation with Text-to-Image Diffusion Model) in view of Pramanik (Masked Multi-Query Slot Attention for Unsupervised Object Discovery), further in view of Karras (A Style-Based Generator Architecture for Generative Adversarial Networks), and further in view of Ye (IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models). Regarding claim 7, 15 Rahman teaches: The method of claim 1, wherein the content mask and the style mask Rahman does not however teach the an “influence parameter” which can determine how much the given style or other embedding will impact the final image generation. This is taught in Ye, which teaches obtaining an embedding influence parameter, wherein the image generation is based on the embedding influence parameter. (Ye 3.2.2 decoupled cross-attention “The image features are integrated into the pretrained UNet model by the adapted modules with decoupled cross-attention. In the original SD model, the text features from the CLIP text encoder are plugged into the UNet model by feeding into the cross-attention layers. Given the query features Z and the text features ct, the output of cross-attention Z′ can be defined by the following equation: Z′ = Attention(Q,K,V) … we simply zero out the CLIP image embedding if the image condition is dropped. As the text cross-attention and image cross-attention are detached, we can also adjust the weight of the image condition in the inference stage: Znew = Attention(Q,K,V) +λ · Attention(Q,K′,V′) where λ is weight factor, and the model becomes the original text-to-image diffusion model if λ = 0.” Note: Ye teaches image conditions used in generation that are associated with specific embeddings, like Rahmans tokens. These image conditions/embeddings that are used to generate the image have a weight factor λ that determines how much the condition influences the image generation, the same concept as an influence parameter.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Ye where the content and style masks used to weigh/optimize embeddings are based on an influence parameter. As the masks in Rahman already consider loss, attention, and a variety of other factors which determine how the content/style embedding will impact the image generation a higher degree of control could be gained over content/style embedding influence by leveraging an influence parameter. Claims 4 is rejected under 35 U.S.C. 103 as being unpatentable over Rahman (Visual Concept-driven Image Generation with Text-to-Image Diffusion Model) in view of Pramanik (Masked Multi-Query Slot Attention for Unsupervised Object Discovery), further in view of Karras (A Style-Based Generator Architecture for Generative Adversarial Networks) and further in view of Ramesh (US 11922550 B1). Regarding claim 4, Rahman teaches: The method of claim 1, wherein: the style embedding Rahman does not however teach the use of a diffusion prior model to generate embeddings. Doing so is taught in Ramesh which teaches that the embedding is generated using a diffusion prior model. (Ramesh Col. 11 Line 61 “first sub-model 138 may include a diffusion prior model, which may include a transformer. The output of first sub-model may be the corresponding image embedding 140. For example, the diffusion prior or the autoregressive prior may generate the corresponding image embedding 140, as discussed herein.” Note: Here Ramesh clearly teaches that a diffusion prior model is used to generate embeddings.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Ramesh where the style embeddings are generated using a diffusion prior model. There are several reasons that would motivate one to do so, as diffusion prior models are generally known to produce higher quality, more structured embeddings Rahman could leverage the model to further improve the quality of its embeddings. Claims 8, 9, 14, are rejected under 35 U.S.C. 103 as being unpatentable over Rahman (Visual Concept-driven Image Generation with Text-to-Image Diffusion Model) in view of Pramanik (Masked Multi-Query Slot Attention for Unsupervised Object Discovery), further in view of Karras (A Style-Based Generator Architecture for Generative Adversarial Networks), and further in view of Avrahami (Blended Latent Diffusion). Regarding claim 8, Rahman teaches: The method of claim 1, wherein: the content mask and the style mask Rahman does not however teach that the masks have a joint condition between them associating them together, this is taught in Avrahami, which teaches masks that satisfy a joint condition. (Avrahami 4.1 Blended Latent Diffusion “We therefore downsample the input mask𝑚 to these spatial dimensions to obtain the latent space binary mask 𝑚latent, which will be used to perform the blending. Now, we noise the initial latent 𝑧init to the desired noise level (in a single step) and manipulate the denoising diffusion process in the following way: at each step, we first perform a latent denoising step, conditioned directly on the guiding text prompt 𝑑, to obtain a less noisy foreground latent denoted as 𝑧fg, while also noising the original latent 𝑧init to the current noise level to obtain a noisy background latent 𝑧bg. The two latents are then blended using the resized mask, i.e. 𝑧fg ⊙ 𝑚latent + 𝑧bg ⊙ (1 − 𝑚latent),” Note: Avrahami specifically teaches two masks are blended with two latents zfg and zbg respectively. These masks are shown to satisfy a joint condition, as one of them is 𝑚latent and the other is described as (1 − 𝑚latent), making one them inverse masks over the same space/image.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Avrahami where content and style masks satisfy a joint condition. There are several reasons that would motivate one to do so, Avrahami details how two components can be separated, or disentangled, from each other with the joint condition they share. Rahman already seeks to disentangle its styles and content; this process could be further enhanced if a joint probability/condition between their masks is used. Regarding claim 9, 14, Rahman teaches: The method of claim 8, wherein: the content mask and the style mask Rahman does not however teach that its masks have jointly related values. This is taught in Avrahami which teaches that its masks comprise scalar values that sum to one. (Avrahami 4.1 Blended Latent Diffusion cited directly teaches this. We know that 𝑚latent is a binary mask meaning it is composed of scalar values that are either 0 or 1, because 𝑚latent is composed of scalar values we can obtain our other mask which is (1 − 𝑚latent). As the second mask is 1 – 𝑚latent we know that all of their values will sum to one.) It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Rahman with Avrahami where style and content masks comprise scalar values that sum to one. There are several reasons that would motivate on to do so, previously the benefits of having a joint condition between the two masks is established. In order to have a joint condition/probability between the masks the values which compose them must be related as well. Having them be scalar values which sum to one, similar to how two values in a joint condition will have their probabilities sum to one, is an easy way to ensure this. Conclusion Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALAN GREGORY HAKALA whose telephone number is (571)272-7863. The examiner can normally be reached 8:00am-5:00pm. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, King Poon can be reached at (571) 270-0728. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /ALAN GREGORY HAKALA/ Examiner, Art Unit 2617 /KING Y POON/Supervisory Patent Examiner, Art Unit 2617