DETAILED ACTION
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 .
This application claims the benefit of U.S. Provisional Application No. 63/597,483, filed November 9, 2023 (specification, ¶ 1). The effective filing date of the claimed subject matter is
treated as November 9, 2023 . The prior art relied upon below predates that date and qualifies as prior art under 35 U.S.C. 102(a)(1).
Information Disclosure Statement
The IDS(s) has/have been considered and placed in the application file.
Claim Objections
Claim 14 is/are objected to because of informalities. The limitation "perform the generate the initial latent representation by utilizing the finetuned LDM" is grammatically incomplete; "the generate" should be corrected ( e.g., "perform the generation of the initial latent representation"). Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-20 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claims 1, 12, and 16 are rejected under 35 U.S.C. 112(b) as being indefinite due to
inconsistent terminology between the claims and the specification. Each independent claim
recites sampling from "a content distribution (p content) associated with prior timesteps" and
from "a style distribution (p style) associated with subsequent timesteps". The specification
describes the same content versus style partition using opposing labels. Specification, ¶ 93 states that the model learns "the content of the image" when denoising "the later timesteps that are closer to the noise sample", and learns "the style of the image" when denoising "the earlier timesteps that are closer to the denoised image latent", whereas the claims and ¶¶ 117-118 associate the content distribution with the "earlier/prior" timesteps (closer to pure noise) and the style distribution with the "later/subsequent" timesteps. Because the specification labels the timesteps at the pure-noise end of the trajectory as both "later" (¶ 93) and "prior/earlier" (¶¶ 117-118 and the claims), the terms "prior timesteps" and "subsequent timesteps" lack a consistent frame of reference, rendering the claims indefinite. Claim 5 supplies a definite frame of reference (to the extent claim 5 is otherwise definite), anchoring the content distribution to the timesteps "closer to a pure noise distribution" and the style distribution to the timesteps "closer to the final image latent"; independent claims 1, 12, and 16 recite no comparable anchor.
Clarification and consistent usage are required. Claims 2-11 , 13-15, and 17-20 are rejected for depending from a rejected base claim. For purposes of applying the prior art below, the limitation is treated as best understood, with the content distribution associated with the timesteps closer to pure noise and the style distribution associated with the timesteps closer to the final image latent, consistent with claim 5 and ¶¶ 94, 117-118.
Claim 8 is rejected under 35 U.S.C. 112(b) as being indefinite. Claim 8 recites "the descriptive input" and "the text prompt", but parent claim 1 recites "descriptive text" and does not recite a "descriptive input" or a "text prompt". There is insufficient antecedent basis for these limitations in the claim. It is unclear whether "the descriptive input" and "the descriptive text" of claim 1 are the same element. Correction is required.
Claim 5 is rejected under 35 U.S.C. 112(b) as being indefinite. Claim 5 recites "training the denoised latents with a U-Net comprising a set of data points". It is unclear how a latent representation is "trained", as training in the art is performed on a model (e.g., the U-Net) rather than on the latents themselves. Further, "the denoised latents" lacks clear antecedent basis, as parent claim 3 recites the act of "denoise latents" rather than "denoised latents". The metes and bounds of the claim cannot be determined.
Claim 11 is rejected under 35 U.S.C. 112(b) as being indefinite. Claim 11 recites "receiving the descriptive text via the user interface", but parent claim 1 introduces "a user interface or a display" in the alternative, such that a user interface is not necessarily present. The recitation "the user interface" therefore lacks sufficient antecedent basis, and the scope of the claim is unclear where the image is output on a display rather than a user interface. Correction is required.
Claim 17 is rejected under 35 U.S.C. 112(6) as being indefinite. Claim 17 recites that the training dataset comprises "labeled pairs of content and style images". This is inconsistent with the specification, which at ¶ 94 describes the content and style data as two separate labeled datasets (a human-in-the-loop alignment dataset representing the content distribution and an experts-in-the-loop style dataset representing the style distribution) rather than matched pairs each comprising a content image and a style image. It is therefore unclear what constitutes a "pair" of content and style images, and the metes and bounds of the claim cannot be determined. Correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 2, 6-20 (with the exception of claims 3, 4, and 5) rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (a mathematical concept) without significantly more.
Claim 1.
Step 1: Claim 1 is directed to a series of steps and therefore falls within the process category.
Step 2A, Prong One: Claim 1 recites a mathematical concept. The limitation "sampling data points from a content distribution (p content) associated with prior timesteps and from a style
distribution (p style) associated with subsequent timesteps, thereby generating a final image
latent" recites sampling data points from defined probability distributions (denoted p content and p style) to generate a latent, which is a mathematical operation, and the claim sets forth the mathematical relationship by reciting the p content and p style distributions in the claim itself.
The recitations of generating an "initial latent representation" and applying a "denoising process" to produce a "refined latent representation" likewise describe mathematical operations performed on latent vectors. See MPEP 2106.04(a)(2)(1); 2024 Subject Matter Eligibility Update (AI), Example 48.
Step 2A, Prong Two: The claim does not integrate the recited mathematical concept into a
practical application. The additional elements are: "detecting input of descriptive text" (data
gathering, i.e., insignificant extra-solution activity, MPEP 2106.05(g)); "utilizing a finetuned
latent diffusion model" and "utilizing a U-Net" (mere instructions to apply the exception using a
generic model and generic components, and generally linking the exception to a technological
environment, MPEP 2106.05(f), (h)); "decoding the final image latent" (a conventional
autoencoder-decoder operation, see specification, ¶ 113); and "outputting the at least one visually aligned image on, or by, a user interface or a display" (insignificant post-solution output, MPEP 2106.05(g)). The improvement asserted in the specification, namely the Style Tailoring training of a single model from content and style distributions partitioned by a timestep cutoff (¶ 93-95), is not recited in claim 1; an improvement must be reflected in the claim to integrate a judicial exception (MPEP 2106.05(a)). Applying generic machine-learning models to a new output domain (stickers) without claiming a technical improvement to the model is not integration. Recentive Analytics, Inc. v. Fox Corp. (Fed. Cir. 2025).
Step 2B: The claim does not recite significantly more. The latent diffusion model, U-Net,
autoencoder/decoder, and display are well-understood, routine, and conventional, as the
specification itself acknowledges: the text-to-image component "may be a standard Latent
Diffusion Model, with a trainable parameter U-net architecture" (¶ 74); latent diffusion models
are described as a known product of "tremendous progress in the field of text-to-image
generation" (¶ 66); and a "pre-trained autoencoder" and a standard "diffusion process" are used (¶ 113). See MPEP 2106.07(a)(III). Outputting the generated image on a display is a generic computer function used for its ordinary purpose (MPEP 2106.05(d)). The additional elements, considered individually and as an ordered combination, add nothing significantly more than the exception itself. Claim 1 is ineligible.
Claims 12 and 16.
Claims 12 (machine) and 16 (article of manufacture) recite the same operations as claim 1 and
are ineligible for the same reasons. The recited "one or more processors" and "at least one
memory" ( claim 12), and the "non-transitory computer-readable medium" ( claim 16), are generic computing hardware that does not integrate the exception or supply an inventive concept (MPEP 2106.05(f)).
Dependent claims 2-20.
The dependent claims are addressed in the following table. None of the additional elements of
the rejected dependent claims integrates the judicial exception into a practical application or
supplies an inventive concept (MPEP 2106.05(f), (g)). Claims 3, 4, and 5 are not rejected for the
reason stated below the table.
Claim(s)
Added Limitation
§ 101 Result
Rationale
2, 19
Finetune the LDM in a domain with high visual quality, prompt alignment, and scene diversity
Ineligible
Recites a result or goal of the exception; no additional technical mechanism
3
Train the LDM to denoise latents from at least two distributions, chosen according to timesteps
Eligible (not
rejected)
Recites the timestep-partitioned dual-distribution training mechanism
4
Denoised latents associated with decoded images
Eligible (not
rejected)
Depends from claim 3
5
Train a U-Net with the content distribution at noise-proximate timesteps and the style distribution at latent-proximate timesteps
Eligible (not
rejected)
Recites the specific Style Tailoring training mechanism
6
Output the image as a digital sticker
Ineligible
Output-type designation;
insignificant post-solution activity
7
Denoising iteratively reduces noise
Ineligible
Generic property of the exception
8
Apply an LLM to map a text embedding to a language embedding space
Ineligible
Generic instruction to apply the exception using an LLM
9
Sampling comprises utilizing a U-Net
Ineligible
Generic component used to apply the exception
10, 20
Train the LDM with a data-domain, human-in-the-loop, or experts-in-the-loop dataset
Ineligible
Specifies training-data content, not a technical mechanism
11
Receive the descriptive text via a user-interface input field
Ineligible
Extra-solution data input
13, 18
Output in multiple formats including a digital sticker format
Ineligible
Output-format designation
14
Generate the latent via an LLM mapping; the LLM trained on text or audio samples
Ineligible
Generic LLM and training-data recitation
15
Denoise via a U-Net trained on varied noise levels and image content
Ineligible
Generic component and training data recitation
17
The distributions learned from labeled pairs of content and style images
Ineligible
Recites training-data content, not a technical mechanism
Claims 3, 4, and 5 are not rejected under 35 U.S.C. 101. Claims 3 and 5 recite a specific training operation, namely jointly training a single model (claim 5, a U-Net) on two timestep-partitioned distributions, that improves the model itself, and claim 4 depends from claim 3. Whereas claim 1 recites only the use of the resulting distributions by sampling at inference, the training improvement recited in claims 3 and 5 integrates the exception into a practical application.
Claim Rejections - 35 USC § 103
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1, 2, 6-16, and 18- 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rombach et al., "High-Resolution Image Synthesis with Latent Diffusion Models," (hereinafter "ROMBACH") in view of Balaji et al., "eDiff-1: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers," (hereinafter "BALAJI") in view of Ghosh et al., US 12,124,803 B2 (hereinafter "GHOSH").
Claim 1.
ROMBACH in view of BALAJI discloses a method comprising:
detecting input of descriptive text associated with at least one of text content or audio content (ROMBACH: "we turn diffusion models into powerful and flexible generators for general conditioning inputs such as text" (Abstract); "To pre-process y from various modalities (such as language prompts) we introduce a domain specific encoder" (§ 3.3).);
generating, based on the descriptive text, an initial latent representation by utilizing a finetuned latent diffusion model (LDM) (ROMBACH: "the encoder E encodes x into a latent representation z = E(x)" (§ 3.1); latent diffusion model conditioned on text via cross-attention (§§ 3.2-3.3).);
applying a denoising process to the initial latent representation to produce a refined latent representation (ROMBACH: latent diffusion models "learn a data distribution p(x) by gradually denoising a normally distributed variable, which corresponds to learning the reverse process of a fixed Markov Chain of length T", with the backbone "realized as a time-conditional UN et"(§ 3.2).);
decoding the final image latent to obtain at least one visually aligned image corresponding to the descriptive text (ROMBACH: "the decoder D reconstructs the image from the latent, giving
x
~
= D(z) = D(E(x))" (§ 3.1).); and
outputting the at least one visually aligned image (ROMBACH: text-to-image synthesis producing output images for presentation (Abstract; § 4)).
ROMBACH does not teach sampling data points from a content distribution associated with prior timesteps and from a style distribution associated with subsequent timesteps and on, or by, a user interface or a display. However, BALAJI teaches the limitation of sampling at prior (early, closer-to-noise) timesteps, from a content distribution, and at subsequent (later) timesteps, from a style distribution (BALAJI: "Early in sampling, generation strongly relies on the text prompt to generate text-aligned content, while later, the text conditioning is almost entirely ignored, and the task changes to producing outputs of high visual fidelity" (Abstract); "we instead train an ensemble of expert denoisers that are specialized for denoising in different intervals of the generative process" (Fig. 2). This teaches sampling, at prior (early, closer-to-noise) timesteps, from a content distribution that generates text-aligned content, and at subsequent (later) timesteps, from a style distribution that supplies the high-visual-fidelity appearance of the image.). Also, GHOSH teaches a user interface or a display (col. 2/claim 1: “displaying the image on the display of the portable device”).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the latent diffusion model of ROMBACH to sample from a content distribution at the prior timesteps and a style distribution at the subsequent timesteps as taught by BALAJI and to display the image on a display as taught by GHOSH. The motivation would have been to capture the distinct content-forming and detail-forming behaviors that occur at different stages of the denoising trajectory, thereby improving prompt alignment while preserving visual fidelity, as BALAJI teaches that sharing a single denoiser across all timesteps ("may not be ideal to best capture these distinctly different modes of the generation process" (Abstract)), yielding a predictable improvement in generation quality. It further would have been obvious to output the resulting image on a user interface or display as taught by GHOSH, in order to present the generated image to the user, a predictable use of a generated image.
Claim 12.
ROMBACH in view of BALAJI and GHOSH discloses an apparatus comprising: one or more processors; and at least one memory storing instructions, that when executed by the one or more processors, cause the apparatus to … (ROMBACH: "training the most powerful DMs often takes hundreds of GPU days" (§ 1).). The remaining limitation correspond to the method of claim 1 and rejected on the same grounds as claim 1 above, mutatis mutandis.
Claim 16.
ROMBACH in view of BALAJI and GHOSH discloses a non-transitory computer-readable medium comprising instructions that, when executed, cause … (ROMBACH: “https://github.com/CompVis/latent-diffusion” is the link for the code found under the title on the first page; GHOSH: “a memory storing instructions” (claim 13).). The remaining limitation correspond to the method of claim 1 and rejected on the same grounds as claim 1 above, mutatis mutandis.
Claims 2 and 19.
ROMBACH in view of BALAJI and GHOSH discloses finetuning the LDM in a data domain with high visual quality, prompt alignment and scene diversity (BALAJI: "results in improved text alignment while maintaining the same inference computation cost and preserving high visual quality" (Abstract); training models specialized to the generation domain (§ 4).).
Claims 6, 13, and 18.
ROMBACH in view of BALAJI teaches outputting the at least one visually aligned image (claims 1, 12, and 16) but does not expressly teach outputting the image as a digital sticker (claim 6), or in multiple formats comprising a digital sticker format based on an application (claims 13 and 18).
GHOSH teaches generating an image as a digital sticker for a messaging application using a text-to-image latent diffusion model (GHOSH: "generating a sticker from the expanded user input using a machine-learning image generation model" (col. 2, lines 1-5); the model is "a text-to-image machine-learning model that converts input text to an image" and is "a Latent Diffusion Model (LDM)" (col. 7, lines 40-45; FIG. 3); a sticker is one of "graphical elements that can be applied like a label to a photo or video, or included in or over a messaging conversation" (col. 1).).
It would have been obvious to one of ordinary skill in the art before the effective filing date to output the generated image of ROMBACH and BALAJI as a digital sticker, including in a digital sticker format based on a messaging application as recited in claims 13 and 18, as taught by GHOSH. The motivation would have been to adapt the generated image for use in a messaging or content application, a predictable use of a generated image yielding the expected result of a usable digital sticker.
Claim 7.
ROMBACH in view of BALAJI and GHOSH discloses wherein the denoising process iteratively reduces noise from the initial latent representation (ROMBACH: "gradually denoising" as "an equally weighted sequence of denoising autoencoders" over t = 1 ... T (§ 3.2).).
Claim 8.
As best understood in view of the 35 U.S.C. 112(b) rejection above, ROMBACH in view of
BALAJI and GHOSH discloses wherein the descriptive input includes the text prompt, and further comprising applying a large language model to generate a mapping between a text embedding derived from the descriptive input and a language embedding space; and providing the mapping to the finetuned latent diffusion model (BALAJI: "the T5 text encoder [56], which is trained for the language modeling task" provides text embeddings to the model (§ 1); ROMBACH: a "domain specific encoder" projects the language prompt to an embedding mapped to the U-Net via cross-attention (§ 3.3).).
Claim 9.
ROMBACH in view of BALAJI discloses wherein the sampling data points comprises utilizing a U-Net to process the refined latent representation (ROMBACH: the model backbone is "realized as a time-conditional UNet" (§ 3.2); BALAJI: the per-interval expert denoisers are applied to denoise the latent during sampling (§ 4).).
Claims 10 and 20.
ROMBACH in view of BALAJI discloses training the LDM with at least one of a data domain alignment dataset, a human-in-the-loop alignment dataset, or an expert-in-the-loop style dataset (Under the broadest reasonable interpretation, the recited "at least one of' is satisfied by a data domain alignment dataset, i.e., the large image-text training corpus on which the LDM is trained – ROMBACH: training on image data to model the data distribution p(x) (§ 3.2); BALAJI: training on the benchmark dataset (§ 4).).
Claim 11.
ROMBACH in view of BALAJI and GHOSH discloses receiving the descriptive text via the user interface, wherein the user interface comprises an input field to receive the text content. ROMBACH and BALAJI generate an image from a text prompt but do not expressly describe
the user interface by which the descriptive text is received; however, GHOSH teaches receiving
the descriptive text via a user interface comprising an input field (GHOSH: "the input text is
user input into a conversation starting or taking place in a message session in the messaging client 104" running on the user device, i.e., received via the user interface of the messaging client, and "the text may be input into a search bar for stickers" (col. 8, lines 60-67), the search bar being an input field of the messaging client's user interface that receives the text content.).
It would have been obvious to one of ordinary skill in the art before the effective filing date to receive the descriptive text of ROMBACH and BALAJI via a user interface comprising an input field as taught by GHOSH, in order to provide a practical means for a user to enter the text prompt to the text-to-image system, a predictable use of a user interface.
Claim 14.
ROMBACH in view of BALAJI discloses perform[ing] the generate the initial latent representation by utilizing the finetuned LDM by applying a large language model to map the descriptive text to an embedding space, wherein the large language model is trained on a dataset comprising text samples or audio samples (BALAJI: "the T5 text encoder [56], which is
trained for the language modeling task" provides text embeddings to the model (§ 1);
ROMBACH: a "domain specific encoder" projects the language prompt to an embedding
mapped to the U-Net via cross-attention (§ 3.3).).
Claim 15.
ROMBACH in view of BALAJI discloses perform the denoising process by applying a U-Net architecture to iteratively reduce noise from the initial latent representation, wherein the U-Net architecture is trained on a dataset with varied noise levels and image content (ROMBACH: the "time-conditional UNet" is trained over t= 1 ... T, i.e., across varied noise levels (§ 3.2); BALAJI: denoisers trained across the full range of noise levels before specialization (§ 4).).
Claims 3-5 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over ROMBACH in view of BALAJI as applied to claim 1 above, and further in view of Cho et al., US 2025/0078349 A1 (hereinafter "CHO").
Claim 3.
ROMBACH in view of BALAJI disclose the method of claim 1. However, they do not teach “training the LDM to denoise latents, associated with at least two distributions at a same time, chosen according to timesteps.”
CHO discloses training the LDM to denoise latents, associated with at least two distributions at a same time, chosen according to timesteps (CHO: "At training, one or more embodiments jointly learn a content latent space, a style latent space together with training a diffusion model (e.g., U-Net)" (¶ 29), whereby a single diffusion model is trained on a content distribution and a style distribution at the same time, and "the image generation model is trained to learn low-frequency layout information in earlier steps and high-frequency details in the later steps of the reverse diffusion process" (¶ 28), thereby associating the content and style distributions with the timesteps. BALAJI: "we instead train an ensemble of expert denoisers that are specialized for denoising in different intervals of the generative process" (Fig. 2), each timestep interval associated with a corresponding distribution (§ 4).). Under the broadest reasonable interpretation, and consistent with ¶ 94 ("train data to denoise latents from two distributions ... conditioned on timesteps"), the limitation "at a same time" is taught by CHO's single diffusion model jointly trained on the content and style latent spaces, with the two distributions "chosen according to timesteps" per CHO's training of the model to learn layout (content) at the earlier reverse-diffusion steps and detail (style) at the later steps (¶¶ 28-29).
It would have been obvious to one of ordinary skill in the art before the effective filing date to train the latent diffusion model of ROMBACH and BALAJI to jointly denoise latents from a content distribution and a style distribution at the same time, chosen according to timesteps, as taught by CHO, in order to increase the controllability and editability of the diffusion model and to leverage the inductive bias by which layout is learned at the earlier steps and detail at the later steps (¶¶ 28-29), yielding the predictable result of a single model providing separable content and style control.
Claim 4.
ROMBACH in view of BALAJI and CHO discloses wherein the denoised latents are associated with decoded images (ROMBACH: "samples from p(z) can be decoded to image space with a single pass through D" (§ 3.2).).
Claim 5.
ROMBACH in view of BALAJI and CHO discloses, as best understood in view of the 35 U.S.C.
112(6) rejection above, training the denoised latents with a U-Net comprising a set of data points sampled from the content distribution associated with timesteps closer to a pure noise distribution, and data points sampled from the style distribution associated with timesteps closer to the final image latent (BALAJI: "At the early sampling stage, when the input data to the denoising network is closer to the random noise, the diffusion model mainly relies on the
text prompt" (§ 1); at lower noise levels "the attention maps exhibit patterns better
correlated with the image content" (Fig. 3). CHO: "computing a content weight based on a
diffusion timestep, wherein the image is generated based on the spatial layout mask
according to the content weight" (claim 5) and a corresponding style weight (claim 6),
whereby "structural information from the content input is injected into the initial steps (or
first few steps) of the denoising process" (closer to pure noise) and "the style information
from the style input is injected into the later steps of the denoising process" (closer to the
final image latent) (¶ 28).).
Claim 17.
Claim 17 recites that the content distribution (p content) and the style distribution are learned
from a training dataset comprising "labeled pairs of content and style images". As best understood in view of the 35 U.S.C. 112(b) rejection of claim 17 set forth above (the
specification at ¶ 94 describing the content and style data as two separate labeled datasets rather than matched content-and-style image pairs), this limitation is treated as requiring that the content distribution and the style distribution be learned from labeled content images and labeled style images. ROMBACH in view of BALAJI discloses the content and style distributions
(claim 16). ROMBACH and BALAJI do not expressly teach learning those distributions from
content and style images; however, CHO teaches learning a content distribution and a style
distribution from a content image and a style image (CHO: "the content input comprises a
content image and the style input comprises a style image" (claim 2; ¶ 32); the apparatus
"encodes the content image, via a content encoder" and "encodes the style image, via a style
encoder" (¶ 39), and "At training, one or more embodiments jointly learn a content latent
space, a style latent space together with training a diffusion model (e.g., U-Net)" (¶ 29),
whereby the content and style distributions are learned from the content and style images.).
It would have been obvious to one of ordinary skill in the art before the effective filing date to learn the content distribution and the style distribution of ROMBACH and BALAJI from labeled content and style images as taught by CHO, in order to obtain controllable, reference-based control over the content and style of the generated image, which CHO teaches as the benefit of its content-image and style-image conditioning.
Conclusion
The prior art made of record but not relied, yet considered pertinent to the applicant’s disclosure: US 2024/0378858 A1, directed to training a generative model to generate stylized content, pertinent to the claimed finetuning of a latent diffusion model toward a target style; US 2025/0095256 A1, directed to image generation using style images, pertinent to the claimed learning of a style distribution from style images (claim 17); and US 2025/0117973 A1, directed to style-based image generation, pertinent to the claimed content-versus-style control of generated images.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Ross Varndell whose telephone number is (571)270-1922. The examiner can normally be reached M-F, 9-5 EST.
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, O’Neal Mistry can be reached at (313)446-4912. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see https://ppair-my.uspto.gov/pair/PrivatePair. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Ross Varndell/Primary Examiner, Art Unit 2674
Prosecution Insights