Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

StyleBlend: Enhancing Style-Specific Content Creation in Text-to-Image Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read StyleBlend claims that decomposing image style into composition and texture, learning each separately, and swapping self-attention Q/KV features between two synthesis branches yields images that are both text-aligned and stylistically…

desk verdict A plausible, clearly-specified few-shot style T2I method whose main idea hinges on an SDEdit assumption that is asserted but not quantitatively verified; worth reviewing, but needs stronger evidence. read the letter →

arxiv 2502.09064 v1 pith:VNM7QN2Q submitted 2025-02-13 cs.CV

Zichong Chen , Shijin Wang , Yang Zhou This is my paper · ORCID
classification cs.CV
keywords style-specificgenerationdiffusionmodelsstyledecompositioncompositiontextureTextualInversionLoRAself-attentionfeatureinjection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's aim is to let a text-to-image diffusion model reproduce a specific artistic style from very few reference images without losing what the text prompt asked for. Existing fine-tuning style methods either capture the style but drift from the prompt or stay aligned with the prompt but wash out the style. StyleBlend treats style as two separate things: composition (semantic structure and layout) and texture (local appearance). It learns each with a different recipe and, at generation time, runs two branches that exchange attention features so the final image gets structure from one branch and appearance from the other.

What carries the argument

The mechanism is a dual-branch synthesis framework with cross-branch self-attention feature swapping in the U-Net. Query features are treated as carriers of composition (semantic structure and layout), and Key/Value features as carriers of texture; swapping $Q$ from the composition branch into the texture branch and $KV$ from the texture branch into the composition branch lets each branch contribute what it learned. The composition representation is produced by training text-encoder LoRAs on SDEdit-generated copies of the references, which are noised to 0.8T (0.9T for SDXL) and denoised with a class-name prompt so that they keep layout while varying appearance.

What would settle it

Run the composition branch alone and measure its outputs' texture similarity to the references with a style descriptor: if those outputs already match the reference texture, or if their layouts drift from the reference layout, the SDEdit step has not isolated composition from texture and the blending premise fails.

Watch

Extended reading notes

Core claim

StyleBlend's central claim is that the apparent trade-off between text alignment and style coherence in few-shot stylization is caused by conflating two distinct components of style. By learning texture with a Textual Inversion embedding plus a LoRA on the denoising network and learning composition from SDEdit-edited reference images with a LoRA on the text encoder, the method obtains two decoupled representations. During inference a dual-branch framework swaps self-attention features: the $Q$ features of the texture branch are replaced by the $Q$ features of the composition branch, while the $K$ and $V$ features of the composition branch are replaced by those of the texture branch. The final output is taken from the texture branch, giving an image that holds the prompt's semantics and layout while carrying the reference's appearance.

Load-bearing premise

The load-bearing premise is that the SDEdit recipe—noising a reference image to 0.8T (0.9T for SDXL) and denoising with its class name—yields training images that keep the reference's structure and layout while changing its texture, so that the composition LoRA learns only composition.

Editorial extensions

If this is right

  • With three reference images, StyleBlend reports higher style similarity and text alignment than the finetuning, pretrained, and training-free baselines it compares against.
  • The learned style weights transfer across Stable Diffusion v1.5, v2.1, and SDXL, so one training run serves several base models.
  • Because StyleBlend leaves the base diffusion architecture untouched, it combines with ControlNet and IP-Adapter for conditional or subject-driven stylized generation.
  • In the one-shot case the method still produces credible style transfer, but the paper reports that text alignment can be weaker than in three-shot cases.
  • Inference takes roughly twice as long as a single diffusion pass because two branches run simultaneously.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the $Q$-versus-$KV$ split may be a general control mechanism, so any latent diffusion model with self-attention could use the same swap to steer structure and appearance separately.
  • Editorial inference: a testable extension is to replace the fixed 0.8T/0.9T noise level with a per-style sweep, since styles with strong global geometry may need heavier noising to strip texture while dense-pattern styles may need lighter noising to keep layout.
  • Editorial inference: for styles where texture defines the layout, such as line art or mosaics, the two learned representations are likely to overlap, and a user may need to favor one branch's output over the other.
  • Editorial inference: because the final image is taken from the texture branch, composition serves as an auxiliary structuring signal, and blending at multiple resolutions or in both directions could improve one-shot text alignment at some cost in style fidelity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. StyleBlend proposes a fine-tuning approach for style-specific text-to-image generation with Stable Diffusion models. The method decomposes image style into composition (semantic structure and layout) and texture (local appearance). Texture style is learned by first optimizing a Textual Inversion token and then training a U-Net LoRA; composition style is learned by generating a synthetic dataset with SDEdit at high noise levels and training a text-encoder LoRA on that dataset. At inference, a dual-branch framework swaps Q features from the composition branch into the texture branch and KV features from the texture branch into the composition branch, with the final image taken from the texture branch. The paper presents qualitative comparisons against fine-tuning, training-free, and pretrained baselines, a quantitative scatter plot of CSD style similarity versus CLIP-Score text alignment on five self-collected styles, ablations of style representations and feature injection, and applications with ControlNet, IP-Adapter, and multiple Stable Diffusion versions.

Significance. If the central decomposition is valid, StyleBlend offers a practical and computationally light recipe for few-shot style personalization that improves the style-coherence/text-alignment trade-off over several baselines. The paper deserves credit for a clearly stated mechanism (Sec. 4.2), for explicit efficiency numbers (Sec. 5.1), and for demonstrating compatibility with SD1.5, SD2.1, and SDXL as well as with plug-ins such as ControlNet and IP-Adapter. The evaluation metrics are external and independent of the training objectives, which avoids circularity. However, the paper's central claim rests on an unvalidated assumption about what SDEdit-generated images encode, and the quantitative evidence for superiority lacks error bars, significance tests, and public data. The strengths are real but the current evidence is not yet at the level that would make the claims conclusive.

major comments (4)
  1. [Sec. 4.1, Composition style representation] The load-bearing assumption is that noising a reference image to 0.8T (or 0.9T for SDXL) and denoising with a class-name prompt yields images that "share highly similar structures and layouts with reference images while introducing variations in texture style." This is asserted without quantitative verification. At these high noise levels, the denoising trajectory can be dominated by the class-name prompt and the model prior, potentially altering object shape, position, and scene layout. If so, the composition LoRA would encode content-specific layout rather than a style-level compositional signal, and the Q-feature swap in Sec. 4.2 would inject content instead of composition. The only supporting ablation (Fig. 9, third row) is qualitative and shown for a single style; Fig. 11 does not include the "without composition style" variant. Please add a quantitative layout-preservation measurement (e.g., semantic segmentation IoU, edge or depth similarity, or self-attention feature distance between SDEdit outputs and references) across several styles and prompts, and include it in the ablation.
  2. [Sec. 5.2, Fig. 8] The quantitative comparison reports CSD and CLIP-Score scatter points over five self-collected styles without error bars, confidence intervals, or significance tests. The claim of "superiority" over baselines therefore rests on point estimates computed on a small, non-released dataset. The use of the maximum CSD over reference images is also favorable to methods that match any single reference; this should be justified. Please report per-style scores, number of generated images per prompt, and appropriate variance or statistical tests, and consider releasing the dataset to allow independent verification.
  3. [Sec. 5.2, comparison protocol for StyleID] For StyleID, the content images are generated using the authors' own CSB branch, while for StyleAligned a GPT-generated style description is used and for other baselines repository defaults are used. This asymmetry may advantage StyleBlend in the qualitative comparison because the content images already carry the desired composition. Please describe how the baselines' inputs were chosen so that they receive comparable content information, or justify why the chosen protocol is fair.
  4. [Sec. 5.5, Limitations] The limitation section already acknowledges weaker text alignment in 1-shot cases and doubled inference cost. These are honest, but they directly qualify the abstract and contribution claims of "text-aligned and stylistically coherent" synthesis. Please explicitly state in the abstract or introduction that 1-shot performance is weaker on text alignment, and report the 1-shot quantitative scores separately rather than only as scatter points.
minor comments (5)
  1. [Abstract] The phrase "enabling content synthesis of both text-aligned and stylistically coherent" is missing a noun; please revise to "synthesis of images that are both text-aligned and stylistically coherent."
  2. [References] The LoRA reference [HW AZ*21] contains an erroneous space in the author field "HW AZ*21"; should be "HZ21" or the standard citation format.
  3. [Figures 9 and 10] The captions for Figures 9 and 10 contain garbled placeholder text (e.g., "/uni00000013/uni00000011/...") that appears to be a formatting artifact; these should be replaced with the intended caption text.
  4. [Sec. 5.1, Dataset] The dataset is described as "collected from the Internet" and augmented with StyleAligned and VSP outputs, but no release statement or URL is provided. Without the dataset, the quantitative comparison cannot be reproduced. Please state whether the dataset will be released and, if so, where.
  5. [Sec. 5.2, Figure 8 caption] The caption describes blue, green, red, and orange scatter points, which may be indistinguishable in grayscale printing; please use distinct markers or provide a colorblind-safe alternative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: StyleBlend's style decomposition and Q/KV feature blending are empirical hypotheses evaluated against external metrics, not derivation steps that reduce to their own inputs.

full rationale

The paper's central claim is that decomposing style into composition and texture, learning each with different objectives, and swapping Q (from CSB) and KV (from TSB) features during synthesis improves both text alignment and style coherence. Walking the derivation chain: the decomposition 'Image Style = Composition Style + Texture Style' (Sec. 4.1) is an explicit modeling assumption, not a prediction derived from the method's outputs. Texture style is learned with standard Textual Inversion and LoRA objectives (Eqs. 3-4); composition style is learned with a text-encoder LoRA on SDEdit-generated images (Eq. 5). The SDEdit assumption that 0.8T/0.9T noise preserves structure while varying texture is empirically asserted and could be false, but falsity would make the method fail; it is not a circular dependency because the assumption is not defined in terms of the target result. The Q-means-structure / KV-means-appearance choice is imported from external prior works (TGBD23, CWQ*23, HVFCO24) and is independently checkable, not from the authors' own prior theorems. Evaluation uses CSD and CLIP-Score, neither of which appears in any training loss, and comparisons are against external baselines. Ablations remove components and observe degradation, which is a normal empirical test rather than a tautology. No self-citation chain is load-bearing, and no fitted parameter is renamed as a prediction. The acknowledged limitations (1-shot text-alignment weakness, failure to reproduce black-and-white color) further indicate that the claims are empirical and falsifiable rather than constructed to match the inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The ledger treats the paper's hand-set hyperparameters as free parameters because the central training recipes depend on them. The main axioms are the style decomposition, the SDEdit composition-data assumption, and the Q-versus-KV feature roles; none are proven. No new physical entities are introduced; the <T> and <C> tokens are learned parameters, not independently evidenced postulates.

free parameters (5)
  • SDEdit noise ratio t0 = 0.8T (SD v1.5/v2.1), 0.9T (SDXL)
    Chosen by hand in Sec. 5.1 to generate composition training data; controls how much texture is removed and how much layout is preserved.
  • LoRA rank = 16
    Set across all experiments (Sec. 5.1); no search reported.
  • Training steps = 500 texture, 300 composition (400/600 for SDXL)
    Chosen by early stopping after a few hundred steps in Sec. 5.1; affects overfitting and style fidelity.
  • Learning rates = 0.0001 for LoRA, 0.01 for embeddings
    Chosen in Sec. 5.1; standard values but not derived from the problem.
  • SDEdit dataset size per reference = 10 images
    Chosen in Sec. 5.1; affects the diversity and quality of the composition-style training set.
assumptions (4)
  • ad hoc to paper Style decomposes into composition and texture (Image Style = Composition Style + Texture Style).
    Stated in Sec. 4.1 as a definition based on empirical observations; no completeness or orthogonality proof is given.
  • domain assumption SDEdit at 0.8T/0.9T with class-name prompts yields training images that share structure and layout with references while varying texture.
    Implied in Sec. 4.1 composition style paragraph; if false, the composition LoRA learns content or texture, breaking the decomposition.
  • domain assumption In U-Net self-attention, Q features encode semantic structure and layout while K and V encode texture and appearance.
    Invoked in Sec. 4.2 and supported only by citations to prior empirical work, not by a proof in this paper.
  • domain assumption The pretrained Stable Diffusion series shares a compatible architecture for the same LoRA and feature-injection protocol across v1.5, v2.1, and SDXL.
    Used in Sec. 5.1 and 5.4 to justify cross-model application; architecture similarity is asserted, not formally established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StyleBlend: Enhancing Style-Specific Content Creation in Text-to-Image Diffusion Models." pith.science (2026). https://pith.science/paper/VNM7QN2Q

@misc{pith2026250209064,
  author       = {Pith},
  title        = {Pith review of: StyleBlend: Enhancing Style-Specific Content Creation in Text-to-Image Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VNM7QN2Q}},
  note         = {Machine review of arXiv:2502.09064}
}
read the original abstract

Synthesizing visually impressive images that seamlessly align both text prompts and specific artistic styles remains a significant challenge in Text-to-Image (T2I) diffusion models. This paper introduces StyleBlend, a method designed to learn and apply style representations from a limited set of reference images, enabling content synthesis of both text-aligned and stylistically coherent. Our approach uniquely decomposes style into two components, composition and texture, each learned through different strategies. We then leverage two synthesis branches, each focusing on a corresponding style component, to facilitate effective style blending through shared features without affecting content generation. StyleBlend addresses the common issues of text misalignment and weak style representation that previous methods have struggled with. Extensive qualitative and quantitative comparisons demonstrate the superiority of our approach.

Figures

Figures reproduced from arXiv: 2502.09064 by the authors.

Figure 1
Figure 1. Style-specific text-to-image generation. Given a limited set of style images arranged vertically on the left side, our method generates content images that align with the text semantics while also exhibiting satisfactory stylistic effects (left). Additionally, our approach can be applied to the Stable Diffusion series without any degeneracy or incompatibility (right). Abstract Synthesizing visually impressive images… view at source ↗
Figure 2
Figure 2. Motivation of StyleBlend. Given the three reference im￾ages at the top, we leverage existing methods to learn style repre￾sentations and generate results conditioned by the class-specific prompts. The widely-used techniques, such as Textual Inversion (TI) [GAA*22] and DreamBooth (DB) [RLJ*23], struggle to pro￾duce high-quality style-specific images that maintain style coher￾ence and text alignment simultaneously. We… view at source ↗
Figure 3
Figure 3. Style representation and learning in StyleBlend. We decompose the image style into two components: texture style and compo￾sition style. Texture style refers to the texture and appearance of images. We first optimize a global embedding with an identifier “<T>” to learn the subject of the given images, which can narrow the generated domain, followed by training the LoRA weights for finer-detailed style learning (left… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Dual-branch synthesis framework for inference. The framework comprises two branches: a composition style branch (CSB) and a texture style branch (TSB), which produce composi￾tion style and texture style, respectively. We employ mutual feature injection to blend these s…
Figure 5
Figure 5. Figure 5: Gallery of StyleBlend results for 1-shot (left) and 3-shot (right) cases, where style references are shown at the top. framework to separately generate composition and texture styles. Inspired by recent works [TGBD23; CWQ*23; HVFCO24] on fea￾ture manipulations in T2I m…
Figure 6
Figure 6. Figure 6: Qualitative comparison to finetuning-based methods for style-specific T2I generation in 3-shot cases. image (1-shot). The results demonstrate that our method effectively manages both style coherence and text alignment simultaneously. Comparison. To validate the effecti…
Figure 8
Figure 8. Figure 8: further presents quantitative comparisons on the met￾rics mentioned above, i.e., the CSD-based style similarity and the CLIP-Score for text alignment. All the scores are computed across the five styles presented in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison to existing 1-shot meth￾ods in style-specific T2I generation, where CSGO [XWS*24] is a model pretrained on large style dataset, and the other three (StyleAligned [HVFCO24], InstantStyle [WWB*24], and StyleID [CHH24]) are training-free methods. ag…
Figure 9
Figure 9. Figure 9: Ablation of the style representations. structure and layout of images without compromising semantic in￾tegrity, while the texture style excels in representing texture and ap￾pearance but loses nearly all semantic information. Quantitative ab￾lation results in [PITH_FU…
Figure 10
Figure 10. Figure 10: Ablation of the feature injection. 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 Style Similarity (CSD) 0.25 0.26 0.27 0.28 0.29 Text Alignment Ours Ours (CSB) Only CSB Only TSB Only Q Inject Only KV Inject [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 12
Figure 12. Figure 12 [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]
Figure 13
Figure 13. Figure 13: Creative StyleBlend results obtained with various Sta￾ble Diffusion base models. Reference image Owl Duck Cat [PITH_FULL_IMAGE:figures/full_fig_p009_13.png]
Figure 14
Figure 14. Figure 14: Limitation of StyleBlend. 5.5. Limitations and discussions While we have demonstrated the effectiveness of StyleBlend across various style examples, several limitations exist. First, while our method performs well in 3-shot cases, it occasionally struggles with text-s…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. PoseAlign: Sculpting Pose-Consistent Meshes via Text-Guided Deformation

    cs.GR 2026-07 conditional novelty 6.0 of 10

    Two-stage text-guided mesh deformation (Laplacian CLIP scaling + attention-shared SDS Jacobian sculpting) better preserves source pose while aligning to text than TextDeformer or MeshUp.

Reference graph

Works this paper leans on

51 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Cross-image attention for zero-shot appearance transfer

    [AGP*24] A LALUF , Y UVAL, G ARIBI , D ANIEL , P ATASHNIK , O R, et al. “Cross-image attention for zero-shot appearance transfer”. Proc. of SIG- GRAPH. 2024, 1–12

    work page 2024

  2. [2]

    Semantic photo manipulation with a generative image prior

    [BSP*19] B AU, D AVID, S TROBELT , H ENDRIK , P EEBLES , W ILLIAM , et al. “Semantic photo manipulation with a generative image prior”. ACM Trans. on Graphics (Proc. of SIGGRAPH)38.4 (2019)

    work page 2019

  3. [3]

    DreamStyler: Paint by Style Inversion with Text-to-Image Diffusion Models

    [ALL*24] A HN, N AMHYUK , L EE, J UNSOO , L EE, C HUNGGI , et al. “DreamStyler: Paint by Style Inversion with Text-to-Image Diffusion Models”. Proc. AAAI Conf. on Artificial Intelligence. 2024, 674–681

    work page 2024

  4. [4]

    A Neural Space-Time Representation for Text-to-Image Personalization

    [ARMCO23] A LALUF , Y UVAL, R ICHARDSON , E LAD, M ETZER , G AL, and C OHEN -OR, D ANIEL . “A Neural Space-Time Representation for Text-to-Image Personalization”.ACM Trans. on Graphics42.6 (2023)

    work page 2023

  5. [5]

    Style Injection in Diffusion: A Training-free Approach for Adapting Large- scale Diffusion Models for Style Transfer

    [CHH24] C HUNG , J IWOO , H YUN, S ANGEEK , and H EO, JAE-P IL. “Style Injection in Diffusion: A Training-free Approach for Adapting Large- scale Diffusion Models for Style Transfer”.Proc. of IEEE Conf. on Com- puter Vision & Pattern Recognition. 2024, 8795–8805 3, 6,

    work page 2024

  6. [6]

    Encoder-based Domain Tuning for Fast Personalization of Text-to- Image Models

    [GAA*23] G AL, R INON , A RAR, M OAB, A TZMON , Y UVAL, et al. “Encoder-based Domain Tuning for Fast Personalization of Text-to- Image Models”. ACM Trans. on Graphics (Proc. of SIGGRAPH) 42.4 (2023)

    work page 2023

  7. [7]

    ControlStyle: Text-Driven Stylized Image Generation Using Diffusion Priors

    [CPYM23] C HEN, JINGWEN , PAN, YINGWEI , YAO, TING, and MEI, TAO. “ControlStyle: Text-Driven Stylized Image Generation Using Diffusion Priors”. ACM International Conference on Multimedia . Ottawa ON, Canada, 2023, 7540–7548

    work page 2023

  8. [8]

    MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing

    [CWQ*23] C AO, M INGDENG , WANG, X INTAO, Q I, Z HONGANG , et al. “MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing”. Proc. of Int. Conf. on Computer Vision . 2023, 22560–22570 3,

    work page 2023

Show all 51 references
  1. [9]

    Z*: Zero-shot Style Transfer via Attention Reweighting

    [DHTD24] D ENG, Y INGYING , H E, X IANGYU , TANG, FAN, and D ONG, WEIMING . “Z*: Zero-shot Style Transfer via Attention Reweighting”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition . 2024, 6934–6944

    2024

  2. [10]

    Diffusion in Style

    [EBA*23] E VERAERT , M ARTIN NICOLAS , B OCCHIO , M ARCO , A RPA, SAMI, et al. “Diffusion in Style”. Proc. of Int. Conf. on Computer Vi- sion. 2023, 2251–2261

    2023

  3. [11]

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

    [EKB*24] E SSER , P ATRICK , K ULAL , S UMITH , B LATTMANN , A N- DREAS , et al. “Scaling Rectified Flow Transformers for High-Resolution Image Synthesis”. arXiv preprint arxiv:2403.03206 (2024) 1,

    2024 arXiv

  4. [12]

    Implicit Style-Content Separation using B- LoRA

    [FVSCO24] F RENKEL , YARDEN , V INKER , YAEL, S HAMIR , A RIEL , and COHEN -O R, D ANIEL . “Implicit Style-Content Separation using B- LoRA”. Proc. of Euro. Conf. on Computer Vision. 2024

    2024

  5. [13]

    An Image is Worth One Word: Personalizing Text-to-Image Generation us- ing Textual Inversion

    [GAA*22] G AL, RINON , ALALUF , YUVAL, ATZMON , YUVAL, et al. “An Image is Worth One Word: Personalizing Text-to-Image Generation us- ing Textual Inversion”. Proc. of Int. Conf. on Learning Representations. 2022 2–4,

    2022

  6. [14]

    Image Style Transfer Using Convolutional Neural Net- works

    [GEB16] G ATYS, L EON A., E CKER , A LEXANDER S., and B ETHGE , MATTHIAS . “Image Style Transfer Using Convolutional Neural Net- works”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2016, 2414–2423

    2016

  7. [15]

    ViCo: Plug-and-play Visual Condition for Personalized Text-to-image Generation

    [HHZW24] H AO, S HAOZHE , H AN, K AI, Z HAO, S HIHAO , and W ONG, KWAN-Y EE K. “ViCo: Plug-and-play Visual Condition for Personalized Text-to-image Generation”. arXiv preprint arxiv:2306.00971 (2024)

    2024 arXiv

  8. [16]

    Denoising Diffusion Probabilistic Models

    [HJA20] H O, JONATHAN , JAIN, AJAY, and ABBEEL , PIETER . “Denoising Diffusion Probabilistic Models”. Advances in Neural Information Pro- cessing Systems. 2020

    2020

  9. [17]

    DreamTuner: Single Image is Enough for Subject-Driven Generation

    © 2025 The Authors. Computer Graphics Forum published by Eurographics and John Wiley & Sons Ltd. 10 of 10 Chen et al. / StyleBlend: Enhancing Style-Specific Content Creation in Text-to-Image Diffusion Models [HLD*23] H UA, M IAO, L IU, J IAWEI , D ING, F EI, et al. “DreamTuner...

    2023 arXiv

  10. [18]

    Freestyle: Free lunch for text-guided style transfer using diffusion mod- els

    [HLZ*24] H E, F EIHONG , L I, G ANG, Z HANG , M ENGYUAN , et al. “Freestyle: Free lunch for text-guided style transfer using diffusion mod- els”. arXiv preprint arXiv:2401.15636 (2024)

    2024 arXiv

  11. [19]

    Style Aligned Image Generation via Shared Attention

    [HVFCO24] H ERTZ, A MIR, V OYNOV , A NDREY , F RUCHTER , S HLOMI , and COHEN -OR, DANIEL . “Style Aligned Image Generation via Shared Attention”. Proc. of IEEE Conf. on Computer Vision & Pattern Recogni- tion. 2024, 4775–4785 3, 5–7. [HW AZ*21] HU, EDWARD J, WALLIS , PHILLIP ,...

    2024

  12. [20]

    Vi- sual Style Prompting with Swapping Self-Attention

    [JKC*24] J EONG , J AESEOK , K IM, J UNHO , C HOI, Y UNJEY , et al. “Vi- sual Style Prompting with Swapping Self-Attention”. arXiv preprint arXiv:2402.12974 (2024) 3,

    2024 arXiv

  13. [21]

    Customizing Text-to-Image Models with a Single Image Pair

    [JWK*24] J ONES , MAXWELL , WANG, SHENG -YU, KUMARI , NUPUR , et al. “Customizing Text-to-Image Models with a Single Image Pair”.arXiv preprint arxiv:2405.01536 (2024)

    2024 arXiv

  14. [22]

    StyleShot: A Snapshot on Any Style

    [JYY*24] J UNYAO, G AO, Y ANCHEN , L IU, Y ANAN , S UN, et al. “StyleShot: A Snapshot on Any Style”.arXiv preprint arxiv:2407.01414 (2024)

    2024 arXiv

  15. [23]

    Multi-Concept Customization of Text-to-Image Diffusion

    [KZZ*23] K UMARI , N UPUR , Z HANG , B INGLIANG , Z HANG , R ICHARD , et al. “Multi-Concept Customization of Text-to-Image Diffusion”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2023 2, 3,

    2023

  16. [24]

    SDEdit: Guided Image Synthesis and Editing with Stochastic Differen- tial Equations

    [MHS*22] M ENG, C HENLIN , H E, Y UTONG , S ONG, Y ANG, et al. “SDEdit: Guided Image Synthesis and Editing with Stochastic Differen- tial Equations”. Proc. of Int. Conf. on Learning Representations. 2022 2–

    2022

  17. [25]

    DiffuseKronA: A Parameter Efficient Fine-tuning Method for Per- sonalized Diffusion Model

    [MSM*23] M ARJIT , SHYAM, SINGH , HARSHIT , MATHUR , NITYANAND , et al. “DiffuseKronA: A Parameter Efficient Fine-tuning Method for Per- sonalized Diffusion Model”. arXiv preprint arxiv:2402.17412 (2023)

    2023 arXiv

  18. [26]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    [PEL*24] P ODELL , D USTIN , E NGLISH , Z ION, L ACEY , K YLE, et al. “SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis”. Proc. of Int. Conf. on Learning Representations . 2024 1–3, 5,

    2024

  19. [27]

    At- tnDreamBooth: Towards Text-Aligned Personalized Text-to-Image Gen- eration

    [PYZ*24] P ANG, L IANYU , Y IN, J IAN, Z HAO, B AOQUAN , et al. “At- tnDreamBooth: Towards Text-Aligned Personalized Text-to-Image Gen- eration”. arXiv preprint arXiv:2406.05000 (2024)

    2024 arXiv

  20. [28]

    Encoding in Style: a StyleGAN Encoder for Image-to-Image Trans- lation

    [RAP*21] R ICHARDSON , E LAD, A LALUF , Y UVAL, PATASHNIK , O R, et al. “Encoding in Style: a StyleGAN Encoder for Image-to-Image Trans- lation”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2021

    2021

  21. [29]

    High-Resolution Image Synthesis With Latent Diffusion Models

    [RBL*22] R OMBACH , R OBIN , B LATTMANN , A NDREAS , L ORENZ , D O- MINIK , et al. “High-Resolution Image Synthesis With Latent Diffusion Models”. Proc. of IEEE Conf. on Computer Vision & Pattern Recogni- tion. 2022, 10684–10695 1–3, 5,

    2022

  22. [30]

    RB-Modulation: Training- Free Personalization of Diffusion Models using Stochastic Optimal Con- trol

    [RCR*24] R OUT, L, CHEN, Y, RUIZ, N, et al. “RB-Modulation: Training- Free Personalization of Diffusion Models using Stochastic Optimal Con- trol”. arXiv preprint arxiv:2405.17401 (2024)

    2024 arXiv

  23. [32]

    DreamBooth: Fine tuning text-to-image diffusion models for subject- driven generation

    [RLJ*23] R UIZ, N ATANIEL , L I, Y UANZHEN , J AMPANI , V ARUN , et al. “DreamBooth: Fine tuning text-to-image diffusion models for subject- driven generation”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2023, 22500–22510 2–4,

    2023

  24. [33]

    HyperDreamBooth: HyperNetworks for Fast Personalization of Text- to-Image Models

    [RLJ*24] R UIZ, N ATANIEL , L I, Y UANZHEN , J AMPANI , V ARUN , et al. “HyperDreamBooth: HyperNetworks for Fast Personalization of Text- to-Image Models”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2024, 6527–6536

    2024

  25. [34]

    Zero-Shot Text-to-Image Generation

    [RPG*21] R AMESH , ADITYA, PAVLOV, MIKHAIL , GOH, GABRIEL , et al. “Zero-Shot Text-to-Image Generation”. Proc. of IEEE Int. Conf. on Ma- chine Learning. 2021, 8821–8831

    2021

  26. [35]

    Deep unsupervised learning using nonequilibrium thermodynamics

    [SDWMG15] S OHL -DICKSTEIN , J ASCHA , W EISS , E RIC, M AH- ESWARANATHAN , N IRU, and G ANGULI , S URYA. “Deep unsupervised learning using nonequilibrium thermodynamics”. Proc. of IEEE Int. Conf. on Machine Learning. PMLR. 2015, 2256–2265

    2015

  27. [36]

    Measuring Style Similarity in Diffusion Models

    [SGG*24] S OMEPALLI , G OWTHAMI , G UPTA, A NUBHAV , G UPTA, K A- MAL, et al. “Measuring Style Similarity in Diffusion Models”. arXiv preprint arxiv:2404.01292 (2024)

    2024 arXiv

  28. [37]

    Curran Associates, Inc., 2023, 66860–66889

    2023

  29. [38]

    ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs

    [SRC*23] S HAH, V IRAJ , R UIZ, N ATANIEL , C OLE, F ORRESTER , et al. “ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs”. arXiv preprint arxiv:2311.13600 (2023)

    2023

  30. [39]

    Plug-and-play diffusion features for text-driven image- to-image translation

    [TGBD23] T UMANYAN , N AREK , G EYER , M ICHAL , B AGON , S HAI, and DEKEL , TALI. “Plug-and-play diffusion features for text-driven image- to-image translation”. Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2023, 1921–1930 3,

    2023

  31. [40]

    P+: Extended Textual Conditioning in Text-to-Image Generation

    [VCCOA24] V OYNOV , A NDREY , C HU, Q INGHAO , C OHEN -O R, DANIEL , and ABERMAN , KFIR. “P+: Extended Textual Conditioning in Text-to-Image Generation”. arXiv preprint arxiv:2303.09522 (2024)

    2024 arXiv

  32. [41]

    Curran Associates, Inc., 2017

    2017

  33. [42]

    In- stantStyle: Free Lunch towards Style-Preserving in Text-to-Image Gen- eration

    [WWB*24] W ANG, H AOFAN, W ANG, Q IXUN , B AI, X U, et al. “In- stantStyle: Free Lunch towards Style-Preserving in Text-to-Image Gen- eration”. arXiv preprint arXiv:2404.02733 (2024) 3, 6,

    2024 arXiv

  34. [43]

    StyleAdapter: A Unified Stylized Image Generation Model without Test-Time Fine-Tuning

    [WWX*24] W ANG, Z HOUXIA , WANG, X INTAO, X IE, L IANGBIN , et al. “StyleAdapter: A Unified Stylized Image Generation Model without Test-Time Fine-Tuning”.arXiv preprint arxiv:2309.01770 (2024)

    2024 arXiv

  35. [44]

    StyleDiffu- sion: Controllable Disentangled Style Transfer via Diffusion Models

    [WZX23] W ANG, Z HIZHONG , Z HAO, L EI, and X ING, W EI. “StyleDiffu- sion: Controllable Disentangled Style Transfer via Diffusion Models”. Proc. of Int. Conf. on Computer Vision. 2023, 7677–7689

    2023

  36. [45]

    Holistically-Nested Edge Detection

    [XT15] X IE, S AINING and T U, Z HUOWEN . “Holistically-Nested Edge Detection”. Proceedings of the IEEE International Conference on Com- puter Vision (ICCV). 2015

    2015

  37. [46]

    CSGO: Content-Style Composition in Text-to-Image Generation

    [XWS*24] X ING, P ENG, W ANG, H AOFAN, S UN, Y ANPENG , et al. “CSGO: Content-Style Composition in Text-to-Image Generation”. arXiv preprint arxiv:2408.16766 (2024) 3, 6,

    2024 arXiv

  38. [47]

    IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Mod- els

    [YZL*23] Y E, H U, Z HANG , J UN, L IU, S IBO, et al. “IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Mod- els”. arXiv preprint arxiv:2308.06721 (2023) 2, 3,

    2023 arXiv

  39. [48]

    ProSpect: Prompt Spectrum for Attribute-Aware Personalization of Dif- fusion Models

    [ZDT*23] Z HANG , Y UXIN , D ONG, W EIMING , T ANG, F AN, et al. “ProSpect: Prompt Spectrum for Attribute-Aware Personalization of Dif- fusion Models”. ACM Trans. on Graphics 42.6 (2023), 244:1–244:14

    2023

  40. [49]

    Inversion-Based Style Transfer With Diffusion Models

    [ZHT*23] Z HANG , Y UXIN , H UANG , N ISHA , T ANG, F AN, et al. “Inversion-Based Style Transfer With Diffusion Models”.Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2023, 10146–10156

    2023

  41. [50]

    Adding Conditional Control to Text-to-Image Diffusion Models

    [ZRA23] Z HANG , L VMIN , R AO, A NYI, and A GRAWALA , M ANEESH . “Adding Conditional Control to Text-to-Image Diffusion Models”.Proc. of IEEE Conf. on Computer Vision & Pattern Recognition. 2023, 3836– 3847

    2023

  42. [51]

    Computer Graphics Forum published by Eurographics and John Wiley & Sons Ltd

    © 2025 The Authors. Computer Graphics Forum published by Eurographics and John Wiley & Sons Ltd

    2025

  43. [139]

    PMLR, 2021, 8748–8763 3,

    Proceedings of Ma- chine Learning Research. PMLR, 2021, 8748–8763 3,

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.