Pith. sign in

REVIEW 3 major objections 5 minor 89 references

A few-step latent diffusion model with structure-semantic co-optimization can synthesize high-resolution visual anagrams that stay faithful to two different text prompts after flips or rotations, and do so in a few seconds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-15 10:45 UTC pith:VSWZWGZG

load-bearing objection Solid engineering paper that makes multi-view anagrams practical on a few-step LDM; the 'noticeably superior' semantic claim is overstated relative to the CLIP deltas, but the speed/resolution win and ablations are real. the 3 major comments →

arxiv 2606.16241 v3 pith:VSWZWGZG submitted 2026-06-15 cs.CV

Structure-Semantic Co-optimized Latent Diffusion Model for Fast Visual Anagram Synthesis

Xiang Gao , Yunpeng Jia This is my paper
classification cs.CV
keywords visual anagramslatent diffusionmulti-view denoisingstructure-semantic co-optimizationoptical illusionstext-to-imagefew-step diffusionattention-guided fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Visual anagrams are single images that read as two different scenes after a simple geometric transform such as a flip or rotation. Earlier diffusion methods could produce them, but only at low resolution, with visible artifacts, weak text fidelity, or multi-minute runtimes. This paper shows that the same multi-view parallel-denoising idea can be moved onto a fast, adversarially distilled latent model (four steps) once three light per-step optimizations are added: null-text alignment of low-frequency structure, attention-based semantic boosting, and attention-weighted noise fusion. The resulting S2CO-Anagram images are 512 by 512, score higher on both CLIP fidelity and aesthetic metrics than prior systems, and finish in roughly 2.6 seconds. The practical claim is that artists and designers can now generate usable dual-interpretation art at interactive speed without sacrificing visual harmony or conceptual clarity.

Core claim

Directly transferring multi-view parallel denoising into a few-step latent diffusion model produces structural divergence and semantic dilution; the paper claims that jointly optimizing a null-text embedding for low-frequency DCT consistency, maximizing key-concept cross-attention on the latent, and fusing noises with attention maps restores both visual harmony and dual-prompt fidelity, yielding higher-resolution anagrams that outperform prior methods while remaining far faster.

What carries the argument

The structure-semantic co-optimization (S2CO) loop run inside each of the four denoising steps: View-Cycle Low-Frequency Consistency loss on the null-text embedding, percentile-masked semantic-enhancement gradients on the latent, and attention-guided noise fusion before the next DDIM update.

Load-bearing premise

The three short per-step optimizations are enough to cancel the structural conflicts and semantic dilution that appear when multi-view denoising is moved onto a four-step latent model, without creating new failure modes that the reported CLIP, aesthetic, and user scores miss.

What would settle it

On a fixed set of prompt pairs and transforms, remove the S2CO loop (or replace attention-guided fusion with simple averaging) and check whether 512-by-512 outputs still match or exceed the reported Amin, concealment, aesthetic, and user-study scores of the full method; a clear drop would falsify the claim that the co-optimization is what makes the few-step latent pipeline work.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Artists can generate dual-interpretation 512-by-512 images under flips and 90/180-degree rotations in a few seconds rather than minutes.
  • Few-step distilled latent models become usable for multi-view optical illusions once low-frequency structure is explicitly aligned and semantics are re-amplified.
  • CLIP-based dual-view metrics and aesthetic scores improve together, showing that speed need not trade off against either fidelity or visual quality.
  • The same three-component loop can be dropped into other multi-view or multi-prompt diffusion pipelines that suffer from structural conflict under averaging.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same low-frequency null-text alignment idea may transfer to other geometric multi-view tasks (e.g., 360-degree or jigsaw rearrangements) where pure noise averaging currently fails.
  • Because the method already relies on cross-attention maps, it could be combined with layout or identity control modules without a full retrain.
  • If the artistic-style bias of the backbone is the main reason realistic anagrams remain weak, swapping the distilled teacher for a photorealistic few-step model would be a direct next experiment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes S2CO-Anagram, which adapts multi-view parallel denoising (noise fusion across a geometric transform g) from pixel-space T2I models to the few-step adversarially distilled LDM SDXL-Turbo (T=4). To counteract structural divergence and semantic dilution that arise in the latent setting, it introduces a structure-semantic co-optimization (S2CO) framework with three components: (i) null-text structure alignment that optimizes the CFG null embedding y∅ via a View-Cycle Low-Frequency Consistency (VLC) loss on DCT low-pass components of Tweedie-predicted clean latents (Eqs. 4–6); (ii) cross-modal semantic enhancement that maximizes percentile-masked cross-attention activations of key prompt tokens w.r.t. the input latent (Eq. 8); and (iii) attention-guided noise fusion that replaces uniform averaging by spatially adaptive weighting with the extracted attention maps (Eq. 9). Algorithm 1 runs K=5 inner optimization steps per denoising step. The method claims 512×512 anagrams with superior visual harmony, conceptual fidelity and semantic saliency relative to Diffusion Illusions, SyncTweedies, Visual Anagrams, Anagram-MTL and LookingGlass, while reducing inference time to 2.6 s.

Significance. If the empirical claims hold, the work supplies a practical, high-resolution, near-real-time tool for artists and designers to create dual-interpretation optical-illusion imagery, lowering the barrier relative to score-distillation or multi-step pixel-space baselines. The technical contribution—explicitly diagnosing structural divergence and semantic dilution when moving parallel denoising into a distilled LDM and countering them with frequency-domain null-text alignment plus attention-driven objectives—is novel within the visual-anagram literature. Strengths that raise the paper’s value include systematic ablations (Tables 2–3, Fig. 5, Appendix D), a 40-participant user study, an auxiliary LLM-augmented prompt set, quantitative verification of reduced DINO self-similarity distance under NTSA, and the promise of public code. These elements make the result falsifiable and reusable even if some absolute metric gains remain modest.

major comments (3)
  1. Table 1 (and the parallel LLM-Augmented Table 9): the headline claim of “noticeably superior \ldots semantic faithfulness” rests on extremely small objective deltas versus the strongest prior method Anagram-MTL (Amin 0.2726 vs 0.2711; C 0.6918 vs 0.6913). No standard errors, confidence intervals or statistical significance tests are reported, and the deltas are smaller than typical CLIP seed-to-seed variance. Because the comparison also confounds resolution (512 vs 256) and backbone (few-step SDXL-Turbo vs multi-step pixel-space models), it remains unclear how much of the reported leap is produced by the three S2CO components successfully counteracting the structural-divergence / semantic-dilution problems that motivate the paper versus simply inheriting a stronger higher-resolution generator. Resolution-matched or backbone-controlled experiments (or at least error bars and a tempered cl
  2. Section 3.2 / Eq. (5) and the free-parameter ledger: the VLC loss is predicated on the axiom that low-frequency DCT components of the Tweedie clean latents capture the global structure that must be shared across views. While Appendix D.3 shows that NTSA reduces last-step DINO self-similarity distance, the paper never tests whether this frequency-domain proxy remains appropriate when the two concepts possess intrinsically incompatible silhouettes (the “horse vs. penguin” example given in the introduction). An ablation that replaces DCT low-pass with an alternative structure metric (e.g., edge maps or DINO features inside the loss itself) would confirm that the chosen proxy is not an ad-hoc fit to the particular prompt pairs used for tuning th=5.
  3. Section 4.1 / Fig. 4 and the evaluation protocol: qualitative panels and the user study (Fig. 7) are favorable, yet they compare outputs at mismatched native resolutions and do not control for the aesthetic prior of SDXL-Turbo itself. Because the aesthetic-score gap (AS 6.14 vs 5.65) is larger than the CLIP gaps, a controlled experiment that runs the prior multi-view algorithms on the same Turbo backbone (or that downsamples S2CO outputs to 256×256 before scoring) is needed to isolate the contribution of S2CO from the contribution of the distilled LDM.
minor comments (5)
  1. Section 3.3 and Appendix B.2: the procedure for selecting “conceptually most crucial words” (highlighted in red) is described only informally; a precise rule or automatic extraction method should be stated so that the attention maps ˜M(1), ˜M(2) are reproducible.
  2. Algorithm 1 and Section 3.5: the learning-rate schedule for λ1, λ2 inside the K-step inner loop is never specified (constant? decayed?). A one-sentence clarification would remove ambiguity for re-implementation.
  3. Figure 2 caption and the surrounding text use both “null-text structure alignment” and “NTSA”; consistent abbreviation after first use would improve readability.
  4. Appendix D.1 Table 4: the sharp drop for T>4 is correctly attributed to the Turbo training regime, yet the main text still presents T=4 as an unproblematic default; a forward reference to the ablation would prevent readers from assuming arbitrary step counts are safe.
  5. References: several arXiv preprints lack version or date stamps; a few venue names appear truncated (e.g., “Proceedings of the Computer Vision and Pattern Recognition Conference”).

Circularity Check

0 steps flagged

No significant circularity; standard empirical ML method with independent losses, ablations, and external metrics.

full rationale

The paper proposes three hand-designed per-step optimizations (VLC loss on DCT low frequencies of Tweedie-predicted clean latents w.r.t. null-text embedding; percentile-masked maximization of key-concept cross-attention; attention-weighted noise fusion) on top of multi-view parallel denoising adapted to SDXL-Turbo. These are not derived from, nor equivalent by construction to, the evaluation metrics (CLIP Amin/C/Aavg, LAION AS, user ratings, DINO self-similarity). Hyperparameters (th=5, p=60, K=5, T=4) are selected via ablation tables that report independent quantitative gains versus the no-optimization baseline on the same backbone; SOTA comparisons use held-out prompt pairs (CIFAR10 2-views and LLM-augmented). No uniqueness theorem, fitted constant renamed as prediction, or load-bearing self-citation chain appears. Self-citations in related work are ordinary prior art on style transfer and do not underwrite the anagram claims. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on the pretrained SDXL-Turbo backbone, the multi-view parallel denoising paradigm taken from prior work, three hand-designed losses/optimizations with free hyper-parameters, and the assumption that CLIP/aesthetic/user scores adequately measure anagram quality. No new physical entities; free parameters are the usual ML knobs.

free parameters (5)
  • DCT low-pass percentile threshold th = 5
    Chosen by ablation (Table 3) and fixed at 5; controls how much structure is aligned.
  • semantic-enhancement percentile p = 60
    Chosen by ablation (Table 8) and fixed at 60; decides which attention locations receive gradient.
  • inner-loop steps K = 5
    Chosen by ablation (Table 5) and fixed at 5; trades quality for speed.
  • learning rates lambda1, lambda2 and CFG omega = 0.1 / 0.5 / 1.5
    Set by hand (0.1, 0.5, 1.5) without exhaustive search reported.
  • denoising steps T = 4
    Fixed at 4 to match SDXL-Turbo sweet spot; ablated in Table 4.
axioms (4)
  • domain assumption Multi-view parallel denoising (noise or clean-image fusion) can produce consistent anagrams when structural and semantic conflicts are mitigated.
    Inherited from Visual Anagrams / SyncTweedies / Anagram-MTL and used as the starting architecture (Section 3.1).
  • ad hoc to paper Low-frequency DCT components of predicted clean latents capture the global structure that must be shared across views.
    Introduced for the VLC loss (Eq. 5); not derived from first principles.
  • domain assumption Cross-attention maps of key prompt tokens indicate the spatial locations whose activation should be maximized for semantic fidelity.
    Standard in diffusion editing literature; used for both SE and Att-Fusion (Eqs. 7–9).
  • domain assumption CLIP image-text cosine similarity and LAION aesthetic predictor are adequate proxies for conceptual fidelity and visual quality of anagrams.
    Used for all quantitative tables; common but imperfect in generative evaluation.
invented entities (2)
  • View-Cycle Low-Frequency Consistency (VLC) loss no independent evidence
    purpose: Penalize structural divergence of multi-view predicted clean latents in DCT domain via null-text optimization.
    New loss defined in Eq. 5; no independent evidence outside this paper’s ablations.
  • Structure-Semantic Co-optimization (S2CO) framework no independent evidence
    purpose: Jointly run null-text alignment, semantic enhancement and attention-guided fusion inside each denoising step.
    Umbrella name for the three proposed modules; evaluated only inside this work.

pith-pipeline@v1.1.0-grok45 · 29505 in / 2866 out tokens · 24318 ms · 2026-07-15T10:45:33.319052+00:00 · methodology

0 comments
read the original abstract

Visual anagram is an intriguing form of art creation wherein a single image presents different conceptual interpretations under transformations such as flipping or rotation. Recent work has achieved visual anagram synthesis by leveraging pretrained text-to-image (T2I) diffusion models, yet still suffers from several key limitations including computational inefficiency, suboptimal aesthetic quality, and weak semantic fidelity and expressiveness. This work focuses on generating visual anagrams with substantially improved visual quality at minimal computational cost, thereby advancing intelligent creation of illusionary digital art. To increase image resolution while reducing time overhead, we adapt the cutting-edge parallel denoising algorithm from pixel-based T2I model to the adversarially distilled latent-based one, and accordingly propose a structure-semantic co-optimization (S2CO) framework to counteract the consequent visual degradation. As the core of our approach, S2CO framework comprises three key innovations: (\romannumeral1) null-text structure alignment optimization; (\romannumeral2) semantic enhancement optimization; (\romannumeral3) attention-guided noise fusion. Building upon these components, our method dubbed \textbf{S2CO-Anagram} is able to generate higher-resolution anagram images with noticeably superior visual harmony and semantic faithfulness than related SOTA approaches, all while achieving substantially faster inference speed. Code will be publicly available.

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 5 linked inside Pith

  1. [1]

    Adversarial diffusion distillation

    Stability AI. Adversarial diffusion distillation. https://stability.ai/research/ adversarial-diffusion-distillation, 2026

    2026

  2. [2]

    Only-style: Stylistic consistency in image generation without content leakage

    Tilemachos Aravanis, Panagiotis Filntisis, Petros Maragos, and George Retsinas. Only-style: Stylistic consistency in image generation without content leakage. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 6941–6951, 2025

    2025

  3. [3]

    Contained neural style transfer for decorated logo generation

    Gantugs Atarsaikhan, Brian Kenji Iwana, and Seiichi Uchida. Contained neural style transfer for decorated logo generation. In2018 13th IAPR International Workshop on Document Analysis Systems (DAS), pages 317–322. IEEE, 2018

    2018

  4. [4]

    Conditional image generation with score-based diffusion models.arXiv preprint arXiv:2111.13606, 2021

    Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schönlieb, and Christian Etmann. Conditional image generation with score-based diffusion models.arXiv preprint arXiv:2111.13606, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    Diffusion illusions: Hiding images in plain sight

    Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael Ryoo. Diffusion illusions: Hiding images in plain sight. InACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024

    2024

  6. [6]

    Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22560–22570, 2023

    2023

  7. [7]

    Looking- glass: Generative anamorphoses via laplacian pyramid warping

    Pascal Chang, Sergio Sancho, Jingwei Tang, Markus Gross, and Vinicius Azevedo. Looking- glass: Generative anamorphoses via laplacian pyramid warping. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 24–33, 2025

    2025

  8. [8]

    Artistic style transfer with internal-external learning and contrastive learning.Advances in Neural Information Processing Systems, 34:26561–26573, 2021

    Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning.Advances in Neural Information Processing Systems, 34:26561–26573, 2021

    2021

  9. [9]

    Optical illusion shape texturing using repeated asymmetric patterns.The Visual Computer, 30(6):809–819, 2014

    Ming-Te Chi, Chih-Yuan Yao, Eugene Zhang, and Tong-Yee Lee. Optical illusion shape texturing using repeated asymmetric patterns.The Visual Computer, 30(6):809–819, 2014

    2014

  10. [10]

    Att- adapter: A robust and precise domain-specific multi-attributes t2i diffusion adapter via condi- tional variational autoencoder

    Wonwoong Cho, Yan-Ying Chen, Matthew Klenk, David I Inouye, and Yanxia Zhang. Att- adapter: A robust and precise domain-specific multi-attributes t2i diffusion adapter via condi- tional variational autoencoder. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15626–15635, 2025

    2025

  11. [11]

    Camouflage images.ACM Trans

    Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images.ACM Trans. Graph., 29(4):51–1, 2010

    2010

  12. [12]

    Improving diffusion models for inverse problems using manifold constraints.Advances in Neural Information Processing Systems, 35:25683–25696, 2022

    Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints.Advances in Neural Information Processing Systems, 35:25683–25696, 2022

    2022

  13. [13]

    Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer

    Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8795–8805, 2024

    2024

  14. [14]

    Latentpaint: Image inpainting in latent space with diffusion models

    Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4334–4343, 2024

    2024

  15. [15]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  16. [16]

    A variational approach to geometric-optical illusions modeling.Proceedings of Fechner Day, 27(1):41–46, 2011

    Werner Ehm. A variational approach to geometric-optical illusions modeling.Proceedings of Fechner Day, 27(1):41–46, 2011

    2011

  17. [17]

    Stam: Zero-shot style transfer using diffusion model via attention modulation

    Masud An Nur Islam Fahim, Nazmus Saqib, and Jani Boutellier. Stam: Zero-shot style transfer using diffusion model via attention modulation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 6333–6343, 2025. 30

    2025

  18. [18]

    Generative diffusion prior for unified image restoration and enhancement

    Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9935–9946, 2023

    2023

  19. [19]

    Motion without movement.ACM Siggraph Computer Graphics, 25(4):27–30, 1991

    William T Freeman, Edward H Adelson, and David J Heeger. Motion without movement.ACM Siggraph Computer Graphics, 25(4):27–30, 1991

    1991

  20. [20]

    Style transfer in text: Exploration and evaluation

    Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. Style transfer in text: Exploration and evaluation. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  21. [21]

    Xiang Gao and Yunpeng Jia. Fbsdiff++: Improved frequency band substitution of diffusion features for efficient and highly controllable text-driven image-to-image translation.arXiv preprint arXiv:2601.19115, 2026

    arXiv 2026

  22. [22]

    Fbsdiff: Plug-and-play frequency band substitution of diffusion features for highly controllable text-driven image translation

    Xiang Gao and Jiaying Liu. Fbsdiff: Plug-and-play frequency band substitution of diffusion features for highly controllable text-driven image translation. InProceedings of the 32nd ACM International Conference on Multimedia, pages 4101–4109, 2024

    2024

  23. [23]

    Sragan: Saliency regularized and attended generative adversarial network for chinese ink-wash painting style transfer.Pattern Recognition, 162:111344, 2025

    Xiang Gao and Yuqi Zhang. Sragan: Saliency regularized and attended generative adversarial network for chinese ink-wash painting style transfer.Pattern Recognition, 162:111344, 2025

    2025

  24. [24]

    Rpd-gan: Learning to draw realistic paintings with generative adversarial network.IEEE Transactions on Image Processing, 29:8706–8720, 2020

    Xiang Gao, Yingjie Tian, and Zhiquan Qi. Rpd-gan: Learning to draw realistic paintings with generative adversarial network.IEEE Transactions on Image Processing, 29:8706–8720, 2020

    2020

  25. [25]

    Learning to incorporate texture saliency adaptive attention to image cartoonization

    Xiang Gao, Yuqi Zhang, and Yingjie Tian. Learning to incorporate texture saliency adaptive attention to image cartoonization. InInternational Conference on Machine Learning, pages 7183–7207. PMLR, 2022

    2022

  26. [26]

    Frequency-controlled diffusion model for versatile text-guided image-to-image translation

    Xiang Gao, Zhengbo Xu, Junhan Zhao, and Jiaying Liu. Frequency-controlled diffusion model for versatile text-guided image-to-image translation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1824–1832, 2024

    2024

  27. [27]

    Ptdiffusion: Free lunch for generating optical illusion hidden pictures with phase-transferred diffusion model

    Xiang Gao, Shuai Yang, and Jiaying Liu. Ptdiffusion: Free lunch for generating optical illusion hidden pictures with phase-transferred diffusion model. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 18240–18249, 2025

    2025

  28. [28]

    A neural algorithm of artistic style.Nature Communica- tions, 2015

    LA Gatys, AS Ecker, and M Bethge. A neural algorithm of artistic style.Nature Communica- tions, 2015

    2015

  29. [29]

    Factorized diffusion: Perceptual illusions by noise decomposition

    Daniel Geng, Inbum Park, and Andrew Owens. Factorized diffusion: Perceptual illusions by noise decomposition. InEuropean Conference on Computer Vision, pages 366–384. Springer, 2024

    2024

  30. [30]

    Visual anagrams: Generating multi-view optical illusions with diffusion models

    Daniel Geng, Inbum Park, and Andrew Owens. Visual anagrams: Generating multi-view optical illusions with diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24154–24163, 2024

    2024

  31. [31]

    On the synthesis of visual illusions using deep generative models.Journal of Vision, 22(8):2–2, 2022

    Alex Gomez-Villa, Adrián Martín, Javier Vazquez-Corral, Marcelo Bertalmío, and Jesús Malo. On the synthesis of visual illusions using deep generative models.Journal of Vision, 22(8):2–2, 2022

    2022

  32. [32]

    Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    2020

  33. [33]

    Ganmouflage: 3d object nondetection with texture fields

    Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4702–4712, 2023

    2023

  34. [34]

    Chipgan: A generative adversarial network for chinese ink wash painting style transfer

    Bin He, Feng Gao, Daiqian Ma, Boxin Shi, and Ling-Yu Duan. Chipgan: A generative adversarial network for chinese ink wash painting style transfer. InProceedings of the 26th ACM international conference on Multimedia, pages 1172–1180, 2018. 31

    2018

  35. [35]

    Style aligned image generation via shared attention

    Amir Hertz, Andrey V oynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024

    2024

  36. [36]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications

    2021

  37. [37]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  38. [38]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. InProceedings of the IEEE international conference on computer vision, pages 1501–1510, 2017

    2017

  39. [39]

    Image-to-image translation with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1125–1134, 2017

    2017

  40. [40]

    Infiniteyou: Flexible photo recrafting while preserving your identity

    Liming Jiang, Qing Yan, Yumin Jia, Zichuan Liu, Hao Kang, and Xin Lu. Infiniteyou: Flexible photo recrafting while preserving your identity. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10898–10907, 2025

    2025

  41. [41]

    Deep learning for fashion style generation.IEEE Transactions on Neural Networks and Learning Systems, 33(9):4538–4550, 2021

    Shuhui Jiang, Jun Li, and Yun Fu. Deep learning for fashion style generation.IEEE Transactions on Neural Networks and Learning Systems, 33(9):4538–4550, 2021

    2021

  42. [42]

    Scenimefy: Learning to craft anime scene via semi-supervised image-to-image translation

    Yuxin Jiang, Liming Jiang, Shuai Yang, and Chen Change Loy. Scenimefy: Learning to craft anime scene via semi-supervised image-to-image translation. InProceedings of the IEEE/CVF international conference on computer vision, pages 7357–7367, 2023

    2023

  43. [43]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. InEuropean conference on computer vision, pages 694–711. Springer, 2016

    2016

  44. [44]

    Synctweedies: A general generative framework based on synchronized diffusions.arXiv preprint arXiv:2403.14370, 2024

    Jaihoon Kim, Juil Koo, Kyeongmin Yeo, and Minhyuk Sung. Synctweedies: A general generative framework based on synchronized diffusions.arXiv preprint arXiv:2403.14370, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    Style transfer by relaxed optimal transport and self-similarity

    Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051–10060, 2019

    2019

  46. [46]

    Learning multiple layers of features from tiny images.(2009), 2009

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.(2009), 2009

    2009

  47. [47]

    Universal style transfer via feature transforms.Advances in neural information processing systems, 30, 2017

    Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms.Advances in neural information processing systems, 30, 2017

    2017

  48. [48]

    Control color: Multimodal diffusion-based interactive image colorization: Z

    Zhexin Liang, Zhaochen Li, Shangchen Zhou, Chongyi Li, and Chen Change Loy. Control color: Multimodal diffusion-based interactive image colorization: Z. liang et al.International Journal of Computer Vision, 133(11):7897–7923, 2025

    2025

  49. [49]

    Diffusion-based aesthetic qr code generation via scanning-robust perceptual guidance

    Jia-Wei Liao, Winston Wang, Tzu-Sian Wang, Li-Xuan Peng, Cheng-Fu Chou, and Jun-Cheng Chen. Diffusion-based aesthetic qr code generation via scanning-robust perceptual guidance. arXiv preprint arXiv:2403.15878, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Diffqrcoder: Diffusion-based aesthetic qr code generation with scanning robustness guided iterative refinement

    Jia-Wei Liao, Winston Wang, Tzu-Sian Wang, Li-Xuan Peng, Ju-Hsuan Weng, Cheng-Fu Chou, and Jun-Cheng Chen. Diffqrcoder: Diffusion-based aesthetic qr code generation with scanning robustness guided iterative refinement. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5916–5925. IEEE, 2025

    2025

  51. [51]

    Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer

    Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5141–5150, 2021. 32

    2021

  52. [52]

    Non-confusing generation of customized concepts in diffusion models

    Wang Lin, Jingyuan Chen, Jiaxin Shi, Yichen Zhu, Chen Liang, Junzhong Miao, Tao Jin, Zhou Zhao, Fei Wu, Shuicheng Yan, et al. Non-confusing generation of customized concepts in diffusion models. InProceedings of the 41st International Conference on Machine Learning, pages 29935–29948, 2024

    2024

  53. [53]

    Structure matters: Tackling the semantic discrepancy in diffusion models for image inpainting

    Haipeng Liu, Yang Wang, Biao Qian, Meng Wang, and Yong Rui. Structure matters: Tackling the semantic discrepancy in diffusion models for image inpainting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8038–8047, 2024

    2024

  54. [54]

    Arttypo: Multi- level controlled artistic typography with iterative feedback

    Kaiyue Liu, Lei Wu, Mingzhe Yu, Xiaole Liu, Yajie Xu, and Xiangxu Meng. Arttypo: Multi- level controlled artistic typography with iterative feedback. In2025 IEEE International Confer- ence on Multimedia and Expo (ICME), pages 1–6. IEEE, 2025

    2025

  55. [55]

    Adaattn: Revisit attention mechanism in arbitrary neural style transfer

    Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649–6658, 2021

    2021

  56. [56]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in neural information processing systems, 35:5775–5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in neural information processing systems, 35:5775–5787, 2022

    2022

  57. [57]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023

    2023

  58. [58]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 4296–4304, 2024

    2024

  59. [59]

    Swiftedit: Lightning fast text-guided image editing via one-step diffusion

    Trong-Tung Nguyen, Quang Nguyen, Khoi Nguyen, Anh Tran, and Cuong Pham. Swiftedit: Lightning fast text-guided image editing via one-step diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21492–21501, 2025

    2025

  60. [60]

    Hybrid images.ACM Transactions on Graphics (TOG), 25(3):527–532, 2006

    Aude Oliva, Antonio Torralba, and Philippe G Schyns. Hybrid images.ACM Transactions on Graphics (TOG), 25(3):527–532, 2006

    2006

  61. [61]

    Camou- flaging an object from many viewpoints

    Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camou- flaging an object from many viewpoints. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2782–2789, 2014

    2014

  62. [62]

    Arbitrary style transfer with style-attentional networks

    Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. Inproceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019

    2019

  63. [63]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Pith/arXiv arXiv 2022

  64. [64]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  65. [65]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  66. [66]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 conference proceedings, pages 1–10, 2022

    2022

  67. [67]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision, pages 87–103. Springer, 2024. 33

    2024

  68. [68]

    Improved aesthetic predictor

    Christoph Schuhmann. Improved aesthetic predictor. GitHub repository. URL https:// github.com/christophschuhmann/improved-aesthetic-predictor

  69. [69]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations

  70. [70]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  71. [71]

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

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1921–1930, 2023

    1921

  72. [72]

    A computational model for color assimila- tion illusions and color constancy

    Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. A computational model for color assimila- tion illusions and color constancy. InProceedings of the Asian Conference on Computer Vision, pages 630–647, 2024

    2024

  73. [73]

    Designdiffusion: High-quality text-to-design image generation with diffusion models

    Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. Designdiffusion: High-quality text-to-design image generation with diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20906–20915, 2025

    2025

  74. [74]

    Diffir: Efficient diffusion model for image restoration

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. InProceedings of the IEEE/CVF international conference on computer vision, pages 13095–13105, 2023

    2023

  75. [75]

    Dg-font: Deformable generative networks for unsupervised font generation

    Yangchen Xie, Xinyuan Chen, Li Sun, and Yue Lu. Dg-font: Deformable generative networks for unsupervised font generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5130–5140, 2021

    2021

  76. [76]

    Diffusion- based visual anagram as multi-task learning

    Zhiyuan Xu, Yinhe Chen, Huan-ang Gao, Weiyan Zhao, Guiyu Zhang, and Hao Zhao. Diffusion- based visual anagram as multi-task learning. In2025 IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV), pages 919–928. IEEE, 2025

    2025

  77. [77]

    Controllable artistic text style transfer via shape-matching gan

    Shuai Yang, Zhangyang Wang, Zhaowen Wang, Ning Xu, Jiaying Liu, and Zongming Guo. Controllable artistic text style transfer via shape-matching gan. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4442–4451, 2019

    2019

  78. [78]

    Apdrawinggan: Generating artistic portrait drawings from face photos with hierarchical gans

    Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Apdrawinggan: Generating artistic portrait drawings from face photos with hierarchical gans. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10743–10752, 2019

    2019

  79. [79]

    Freedom: Training- free energy-guided conditional diffusion model

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training- free energy-guided conditional diffusion model. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174–23184, 2023

    2023

  80. [80]

    Diffusing colors: Image colorization with text guided diffusion

    Nir Zabari, Aharon Azulay, Alexey Gorkor, Tavi Halperin, and Ohad Fried. Diffusing colors: Image colorization with text guided diffusion. InSIGGRAPH Asia 2023 Conference Papers, pages 1–11, 2023

    2023

Showing first 80 references.