pith. machine review for the scientific record. sign in

arxiv: 2604.17850 · v1 · submitted 2026-04-20 · 💻 cs.CV

Recognition: unknown

UniCSG: Unified High-Fidelity Content-Constrained Style-Driven Generation via Staged Semantic and Frequency Disentanglement

Jingwei Yang , Ruoxi Wu , Wei Shen , Meng Li , Yulong Liu , Huimin She , Lunxi Yuan
Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:12 UTC · model grok-4.3

classification 💻 cs.CV
keywords style transferdiffusion modelscontent-style disentanglementlatent spacefrequency supervisionimage generationDiTreward learning
0
0 comments X

The pith

UniCSG separates content semantics from style in latent space through staged low-frequency preprocessing and frequency-aware reconstruction to improve faithfulness in diffusion-based style transfer.

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

Style transfer in DiT-based diffusion models frequently suffers from content-style entanglement that produces leakage from the reference into the content or unstable outputs. The paper presents UniCSG as a unified framework that tackles this in both text-guided and reference-guided settings by training in two latent-space stages. The first stage combines low-frequency preprocessing with conditioning corruption to push content and style apart. The second stage applies multi-scale frequency supervision to restore detail. Pixel-space reward learning then aligns the latent training with perceptual quality after decoding.

Core claim

UniCSG is a unified framework for content-constrained style-driven generation that employs staged training consisting of a latent-space semantic disentanglement stage using low-frequency preprocessing and conditioning corruption to encourage separation, followed by a latent-space frequency-aware detail reconstruction stage with multi-scale frequency supervision, plus pixel-space reward learning to align with perceptual quality, yielding improved content faithfulness and style alignment.

What carries the argument

Staged latent-space training that first disentangles semantics via low-frequency preprocessing and conditioning corruption, then reconstructs details via multi-scale frequency supervision, augmented by pixel-space reward learning.

If this is right

  • Greater content faithfulness because original semantics remain intact while style is applied.
  • Reduced reference-content leakage through explicit separation before detail reconstruction.
  • Robust performance across both text prompts and image references as style sources.
  • Higher perceptual quality after decoding due to the reward alignment step.
  • Unified handling of multiple guidance modes without separate models.

Where Pith is reading between the lines

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

  • The frequency supervision step could be adapted to control detail levels at different scales in other generative tasks.
  • If the separation holds, the same staged approach might reduce entanglement in conditional video generation.
  • Reward learning in pixel space after latent training suggests a general way to bridge latent objectives with human preferences.

Load-bearing premise

That low-frequency preprocessing combined with conditioning corruption will produce reliable content-style separation in latent space without introducing new artifacts that the later frequency reconstruction stage cannot correct.

What would settle it

Measure content leakage on held-out reference images by comparing structural similarity or semantic consistency scores between outputs from the full UniCSG pipeline versus the same model trained without the low-frequency preprocessing and corruption steps.

Figures

Figures reproduced from arXiv: 2604.17850 by Huimin She, Jingwei Yang, Lunxi Yuan, Meng Li, Ruoxi Wu, Wei Shen, Yulong Liu.

Figure 1
Figure 1. Figure 1: Teaser of UniCSG on text-guided and reference-guided style transfer. Given a content image, UniCSG transforms it into user-specified styles under both text prompts and reference exemplars, showing faithful content preservation and style alignment. In the reference-guided rows, each triplet shows (left to right) the reference image, the content image, and the generated result. Abstract. Style transfer must … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of UniCSG. We use staged semantic–frequency disentanglement in latent space and reward-guided perceptual alignment in pixel space. Stage 1 (semantic disentanglement): low-frequency preprocessing plus conditioning corruption builds a robust content–style scaffold. Stage 2 (frequency-aware detail reconstruction): multi￾scale frequency supervision refines high-frequency textures atop the scaffold. Pi… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results for text-guided style transfer on CSG-Bench. OmniConsistency[31] and Ovis-U1[32] in some cases, which may contribute to the lower style-consistency scores. In the reference-guided setting, our results differ substantially from the text￾guided outputs, indicating that the model can learn style characteristics from the reference image. Meanwhile, BAGEL[5] suffers from severe reference-con… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results for reference-guided style transfer on CSG-Bench [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study results. The first row shows the Ghibli style and the second row shows the 3D Chibi style. Left: text-guided setting. Right: reference-guided setting [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results on unseen styles during training. tones, suggesting that high-frequency refinement improves stylization. Remov￾ing the reward model yields a smaller but consistent drop, consistent with its role in perceptual alignment, particularly refining background and clothing de￾tails. Gains are larger in the reference-guided setting, indicating that UniCSG is particularly effective for reference-… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results across diverse style categories. 6 Conclusion We propose UniCSG, a unified training framework for high-fidelity, content￾constrained, style-driven generation. With a staged design semantic disentangle￾ment, frequency-aware reconstruction, and pixel-space reward learning. UniCSG improves style-transfer stability and fine-detail quality while preserving content structure and substantially… view at source ↗
Figure 8
Figure 8. Figure 8: Overview of the CSG-Dataset generation pipeline. The pipeline consists of three main stages: (1) Pre-processing: data cleaning and semantic grouping of content and artistic images. (2) Generation: mutual construction between stylization and de￾stylization models creates diverse style-content combinations through iterative cycles, with text construction via image understanding. (3) Post-processing: four com… view at source ↗
Figure 9
Figure 9. Figure 9: Diversity of CSG-Dataset. The figure shows a matrix visualization where rows represent 20 style types (reference images) and columns represent 6 content types (por￾traits, landscapes, architecture, animals, objects, food). Each cell displays the gener￾ated result from our model, demonstrating the comprehensive coverage of style-content combinations in the dataset [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Diversity of CSG-Dataset (continued) [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual ablation study results. The figure shows four rows corresponding to four style types: American Cartoon, clay, Paper cutting, and Vector. C.4 Conditioning Corruption Schedule The corruption strength follows a progressive decay schedule during training. For Stage 1, we linearly decay the noise amplification factors from their initial values (γcontent = 1.5, γstyle = 2.0) to 1.0 over the first 60% of … view at source ↗
Figure 12
Figure 12. Figure 12: Representative failure cases comparing UniCSG and Nano-banana[4] across six styles in both text-guided and reference-guided settings. (a) Text distortion case: comparison of text preservation and overall stylization quality. (b) Hand gesture de￾formation case: comparison of hand structure preservation across different styles. are particularly challenging because text and hand structures require precise ge… view at source ↗
Figure 13
Figure 13. Figure 13: Extended qualitative results showcasing UniCSG’s performance across diverse content types and style categories. The figure includes various cases demonstrating the method’s effectiveness on different content domains and style categories [PITH_FULL_IMAGE:figures/full_fig_p030_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Extended qualitative results showcasing UniCSG’s performance across diverse content types and style categories(continued) [PITH_FULL_IMAGE:figures/full_fig_p031_14.png] view at source ↗
read the original abstract

Style transfer must match a target style while preserving content semantics. DiT-based diffusion models often suffer from content-style entanglement, leading to reference-content leakage and unstable generation. We present UniCSG, a unified framework for content-constrained, style-driven generation in both text-guided and reference-guided settings. UniCSG employs staged training: (i) a latent-space semantic disentanglement stage that combines low-frequency preprocessing with conditioning corruption to encourage content-style separation, and (ii) a latent-space frequency-aware detail reconstruction stage that refines details via multi-scale frequency supervision. We further incorporate pixel-space reward learning to align latent objectives with perceptual quality after decoding. Experiments demonstrate improved content faithfulness, style alignment, and robustness in both settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes UniCSG, a unified framework for content-constrained style-driven generation with DiT-based diffusion models. It uses staged latent-space training consisting of (i) semantic disentanglement via low-frequency preprocessing combined with conditioning corruption and (ii) frequency-aware detail reconstruction with multi-scale frequency supervision, plus pixel-space reward learning to align with perceptual quality. The approach targets both text-guided and reference-guided settings and claims improved content faithfulness, style alignment, and robustness.

Significance. If the staged disentanglement proves reliable, the work could meaningfully advance controllable style transfer by mitigating content-style entanglement and reference leakage in modern diffusion models. The combination of frequency preprocessing, corruption-based separation, and multi-scale supervision is a coherent technical choice, and the unified treatment of text and reference guidance is a practical strength. No machine-checked proofs or parameter-free derivations are present, but the pipeline is falsifiable via standard style-transfer metrics.

major comments (1)
  1. Abstract and experimental sections: the central claim that experiments demonstrate improved content faithfulness, style alignment, and robustness rests on assertions without any reported quantitative metrics, baselines, ablation studies, or error analysis. This is load-bearing because the value of the low-frequency preprocessing plus conditioning corruption step cannot be assessed without evidence that it produces usable separation that the reconstruction stage can reliably refine.
minor comments (2)
  1. The description of the conditioning corruption mechanism would benefit from an explicit equation or pseudocode showing how corruption is applied and annealed across training stages.
  2. Figure captions and method diagrams should explicitly label the low-frequency preprocessing block and the multi-scale frequency supervision losses to improve traceability from text to visuals.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thorough review and the recommendation of major revision. We agree that the experimental claims require quantitative backing to be convincing and will substantially expand the experimental section with metrics, baselines, ablations, and analysis in the revised manuscript.

read point-by-point responses

  1. Referee: Abstract and experimental sections: the central claim that experiments demonstrate improved content faithfulness, style alignment, and robustness rests on assertions without any reported quantitative metrics, baselines, ablation studies, or error analysis. This is load-bearing because the value of the low-frequency preprocessing plus conditioning corruption step cannot be assessed without evidence that it produces usable separation that the reconstruction stage can reliably refine.

    Authors: We accept this criticism. The current manuscript relies on qualitative results and the abstract's summary statement without supporting numbers. In the revision we will add: (1) quantitative tables reporting LPIPS, SSIM, and CLIP-based content similarity for faithfulness; style alignment via Gram-matrix or VGG feature distances; and robustness metrics under reference perturbation or text variation; (2) direct comparisons against DiT fine-tuning, IP-Adapter, and recent style-transfer baselines; (3) ablations that isolate the low-frequency preprocessing and conditioning-corruption components, measuring their effect on separation quality before and after the reconstruction stage; and (4) an error-analysis subsection with representative failure cases and quantitative breakdown. These additions will allow readers to evaluate whether the staged disentanglement produces the claimed separation. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The manuscript describes an empirical staged training pipeline (latent semantic disentanglement via low-frequency preprocessing and conditioning corruption, followed by frequency-aware reconstruction and pixel-space reward learning) for content-constrained style transfer in DiT diffusion models. No equations, derivations, uniqueness theorems, or self-citations that reduce any claimed result to its own inputs by construction appear in the abstract or method summary. Performance claims rest on experimental outcomes rather than closed-form reductions or fitted parameters renamed as predictions. The separation reliability is flagged as an unverified empirical link, but this is a standard assumption open to external falsification and does not trigger any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; all technical assumptions remain implicit in the described training stages.

pith-pipeline@v0.9.0 · 5442 in / 1059 out tokens · 32941 ms · 2026-05-10T05:12:10.394868+00:00 · methodology

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

53 extracted references · 27 canonical work pages · 8 internal anchors

  1. [1]

    Cao, S., Ma, N., Li, J., Li, X., Shao, L., Zhu, K., Zhou, Y., Pu, Y., Wu, J., Wang, J., Qu, B., Wang, W., Qiao, Y., Yao, D., Liu, Y.: Artimuse: Fine-grained image aesthetics assessment with joint scoring and expert-level understanding (2025), https://arxiv.org/abs/2507.14533

    work page arXiv 2025

  2. [2]

    arXiv preprint arXiv:2503.10614 (2025)

    Chen, B., Zhao, B., Xie, H., Cai, Y., Li, Q., Mao, X.: Consislora: Enhancing content and style consistency for lora-based style transfer. arXiv preprint arXiv:2503.10614 (2025)

    work page arXiv 2025

  3. [3]

    Chen, Z., Zhu, J., Chen, X., Zhang, J., Hu, X., Zhao, H., Wang, C., Yang, J., Tai, Y.: Dip: Taming diffusion models in pixel space (2025),https://arxiv.org/abs/ 2511.18822

    work page arXiv 2025

  4. [4]

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities (2025),https://arxiv.org/abs/2507.06261

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Emerging Properties in Unified Multimodal Pretraining

    Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Shi, G., Fan, H.: Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683 (2025)

    work page internal anchor Pith review arXiv 2025

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11326–11336 (June 2022)

    2022

  7. [7]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021), https://arxiv.org/abs/2010.11929 16 Yang et al

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  8. [8]

    In: Forty- first International Conference on Machine Learning (2024),https://openreview

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M¨ uller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., Podell, D., Dockhorn, T., English, Z., Rombach, R.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty- first International Conference on Machine Learning (2024),https://openreview. net/forum?id=FPnUhsQJ5B

    2024

  9. [9]

    In: Proceedings of the 37th International Conference on Neural Information Processing Systems

    Fu, S., Tamir, N.Y., Sundaram, S., Chai, L., Zhang, R., Dekel, T., Isola, P.: Dream- sim: learning new dimensions of human visual similarity using synthetic data. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. NIPS ’23, Curran Associates Inc., Red Hook, NY, USA (2023)

    2023

  10. [10]

    In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV)

    Gao, Z., Song, J., Zhang, Z., Deng, J., Patras, I.: Frequency-guided diffusion for training-free text-driven image translation. In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV). pp. 19195–19205 (October 2025)

    2025

  11. [11]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    2016

  12. [12]

    In: Proceed- ings of the 31st International Conference on Neural Information Processing Sys- tems

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Proceed- ings of the 31st International Conference on Neural Information Processing Sys- tems. p. 6629–6640. NIPS’17, Curran Associates Inc., Red Hook, NY, USA (2017)

    2017

  13. [13]

    In: ICCV (2017)

    Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV (2017)

    2017

  14. [14]

    Labs, B.F.: Flux.https://github.com/black-forest-labs/flux(2024)

    2024

  15. [15]

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., M¨ uller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: Flux.1 kontext: Flow matching for in-context image generation and editing in latent space (2025),https://a...

    work page internal anchor Pith review arXiv 2025

  16. [16]

    Lai, B., Wang, X., Rambhatla, S., Rehg, J.M., Kira, Z., Girdhar, R., Misra, I.: Toward diffusible high-dimensional latent spaces: A frequency perspective (2025), https://arxiv.org/abs/2511.22249

    work page arXiv 2025

  17. [17]

    Lee, D.J., Hur, J., Choi, J., Yu, J., Kim, J.: Frequency-aware token reduction for efficient vision transformer (2025),https://arxiv.org/abs/2511.21477

    work page arXiv 2025

  18. [18]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)

    Lei, M., Song, X., Zhu, B., Wang, H., Zhang, C.: Stylestudio: Text-driven style transfer with selective control of style elements. In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR). pp. 23443–23452 (June 2025)

    2025

  19. [19]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV)

    Li, H., Fan, Z., Wen, Z., Zhu, Z., Li, Y.: Aicomposer: Any style and content image composition via feature integration. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV). pp. 16840–16850 (October 2025)

    2025

  20. [20]

    Li, T., He, K.: Back to basics: Let denoising generative models denoise (2026), https://arxiv.org/abs/2511.13720

    work page internal anchor Pith review arXiv 2026

  21. [21]

    Liu, J., Zhang, Y., Huang, T., Xu, W., Yang, R.: Distilling cross-modal knowledge via feature disentanglement (2025),https://arxiv.org/abs/2511.19887

    work page arXiv 2025

  22. [22]

    Ma, Z., Wei, L., Wang, S., Zhang, S., Tian, Q.: Deco: Frequency-decoupled pixel diffusion for end-to-end image generation (2025),https://arxiv.org/abs/2511. 19365

    2025

  23. [23]

    Transactions on Ma- chine Learning Research (2024),https://openreview.net/forum?id=a68SUt6zFt, featured Certification

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: UniCSG 17 DINOv2: Learning robust v...

    2024

  24. [24]

    Semantics lead the way: Harmonizing semantic and texture modeling with asynchronous latent diffusion.arXiv preprint arXiv:2512.04926, 2025

    Pan, Y., Feng, R., Dai, Q., Wang, Y., Lin, W., Guo, M., Luo, C., Zheng, N.: Seman- tics lead the way: Harmonizing semantic and texture modeling with asynchronous latent diffusion. arXiv preprint arXiv:2512.04926 (2025)

    work page arXiv 2025

  25. [25]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4195– 4205 (October 2023)

    2023

  26. [26]

    In: Kim, B., Yue, Y., Chaudhuri, S., Fragkiadaki, K., Khan, M., Sun, Y

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M¨ uller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution im- age synthesis. In: Kim, B., Yue, Y., Chaudhuri, S., Fragkiadaki, K., Khan, M., Sun, Y. (eds.) International Conference on Learning Representations. vol. 2024, pp. 1862–1874 (2024),https://pr...

    2024

  27. [27]

    In: Meila, M., Zhang, T

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceed- ings of Machine Learning Res...

    2021

  28. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Shang, C., Wang, Z., Wang, H., Meng, X.: Scsa: A plug-and-play semantic continuous-sparse attention for arbitrary semantic style transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13051–13060 (June 2025)

    2025

  29. [29]

    Measuring style similarity in diffusion models

    Somepalli, G., Gupta, A., Gupta, K., Palta, S., Goldblum, M., Geiping, J., Shri- vastava, A., Goldstein, T.: Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292 (2024)

    work page arXiv 2024

  30. [30]

    Song, X., Wang, L., Wang, W., Li, Z., Sun, J., Zheng, D., Chen, J., Li, Q., Sun, Z.: 3sgen: Unified subject, style, and structure-driven image generation with adaptive task-specific memory (2025),https://arxiv.org/abs/2512.19271

    work page arXiv 2025

  31. [31]

    Song, Y., Liu, C., Shou, M.Z.: Omniconsistency: Learning style-agnostic consis- tency from paired stylization data (2025),https://api.semanticscholar.org/ CorpusID:278905729

    2025

  32. [32]

    Ovis-u1 technical report

    Wang, G.H., Zhao, S., Zhang, X., Cao, L., Zhan, P., Duan, L., Lu, S., Fu, M., Zhao, J., Li, Y., Chen, Q.G.: Ovis-u1 technical report. arXiv preprint arXiv:2506.23044 (2025)

    work page arXiv 2025

  33. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Wang, X., Xu, W., Zhang, Q., Zheng, W.S.: Domain generalizable portrait style transfer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15802–15811 (October 2025)

    2025

  34. [34]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR)

    Wang, Y., Liu, R., Lin, J., Liu, F., Yi, Z., Wang, Y., Ma, R.: Omnistyle: Filtering high quality style transfer data at scale. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR). pp. 7847–7856 (June 2025)

    2025

  35. [35]

    ArXivabs/2509.05970(2025),https://api.semanticscholar

    Wang, Y., Yi, Z., Zhang, Y., Zheng, P., Xie, X., Lin, J., Wang, Y., Ma, R.: Omnistyle2: Scalable and high quality artistic style transfer data generation via destylization. ArXivabs/2509.05970(2025),https://api.semanticscholar. org/CorpusID:286975228

    work page arXiv 2025

  36. [36]

    International Journal 18 Yang et al

    Wang, Z., Wang, X., Xie, L., Qi, Z., Shan, Y., Wang, W., Luo, P.: StyleAdapter: A Unified Stylized Image Generation Model. International Journal 18 Yang et al. of Computer Vision133(4), 1894–1911 (Apr 2025).https://doi.org/10.1007/ s11263-024-02253-x,https://doi.org/10.1007/s11263-024-02253-x

    work page doi:10.1007/s11263-024-02253-x 1911

  37. [37]

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., L...

    work page internal anchor Pith review arXiv 2025

  38. [38]

    Wu, C., Zheng, P., Yan, R., Xiao, S., Luo, X., Wang, Y., Li, W., Jiang, X., Liu, Y., Zhou, J., Liu, Z., Xia, Z., Li, C., Deng, H., Wang, J., Luo, K., Zhang, B., Lian, D., Wang, X., Wang, Z., Huang, T., Liu, Z.: Omnigen2: Exploration to advanced multimodal generation (2025),https://arxiv.org/abs/2506.18871

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Wu, S., Huang, M., Cheng, Y., Wu, W., Tian, J., Luo, Y., Ding, F., He, Q.: Uso: Unified style and subject-driven generation via disentangled and reward learning (2025)

    2025

  40. [40]

    Xia, B., Peng, B., Zhang, Y., Huang, J., Liu, J., Li, J., Tan, H., Wu, S., Wang, C., Wang, Y., Wu, X., Yu, B., Jia, J.: Dreamomni2: Multimodal instruction-based editing and generation (2025),https://arxiv.org/abs/2510.06679

    work page arXiv 2025

  41. [41]

    Xiang, X., Tian, X., Zhang, G., Chen, Y., Zhang, S., Wang, X., Tao, X., Fan, Q.: Denoising vision transformer autoencoder with spectral self-regularization (2025), https://arxiv.org/abs/2511.12633

    work page arXiv 2025

  42. [42]

    In: The Thirty- ninth Annual Conference on Neural Information Processing Systems (2025),https: //openreview.net/forum?id=n4nmiIq3qj

    Xing, P., Wang, H., Sun, Y., wangqixun, Baixu, Ai, H., Huang, J.Y., Li, Z.: CSGO: Content-style composition in text-to-image generation. In: The Thirty- ninth Annual Conference on Neural Information Processing Systems (2025),https: //openreview.net/forum?id=n4nmiIq3qj

    2025

  43. [43]

    Towards scalable pre-training of visual tokenizers for generation

    Yao, J., Song, Y., Zhou, Y., Wang, X.: Towards scalable pre-training of visual tokenizers for generation. arXiv preprint arXiv:2512.13687 (2025)

    work page arXiv 2025

  44. [44]

    Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)

    2023

  45. [45]

    Yin, B., Hu, X., Zhou, X., Jiang, P.T., Liao, Y., Zhu, J., Zhang, J., Tai, Y., Wang, C., Yan, S.: Fera: Frequency-energy constrained routing for effective diffusion adap- tation fine-tuning (2025),https://arxiv.org/abs/2511.17979

    work page arXiv 2025

  46. [46]

    Yu, Y., Xiong, W., Nie, W., Sheng, Y., Liu, S., Luo, J.: Pixeldit: Pixel diffusion transformers for image generation (2025),https://arxiv.org/abs/2511.20645

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  47. [47]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3836–3847 (October 2023)

    2023

  48. [48]

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric (2018),https://arxiv.org/ abs/1801.03924

    work page Pith review arXiv 2018

  49. [49]

    In: Fumero, M., Domine, C., L¨ ahner, Z., Crisostomi, D., Moschella, L., Stachen- feld, K

    Zhang, S.: Fast imagic: Solving overfitting in text-guided image editing via disen- tangled UNet with forgetting mechanism and unified vision-language optimization. In: Fumero, M., Domine, C., L¨ ahner, Z., Crisostomi, D., Moschella, L., Stachen- feld, K. (eds.) Proceedings of UniReps: the Second Edition of the Workshop on Unifying Representations in Neur...

    2024

  50. [50]

    13770 UniCSG 19

    Zhang, S., Chen, Z., Chen, L., Wu, Y.: Cdst: Color disentangled style transfer for universal style reference customization (2025),https://arxiv.org/abs/2506. 13770 UniCSG 19

    2025

  51. [51]

    Zhang, S., Huang, H., Zhang, C., Li, X.: Qwenstyle: Content-preserving style trans- fer with qwen-image-edit (2026),https://arxiv.org/abs/2601.06202

    work page arXiv 2026

  52. [52]

    Zhang, S., Xiao, S., Huang, W.: Forgedit: Text guided image editing via learning and forgetting (2024),https://openreview.net/forum?id=nh4vQ1tGCt

    2024

  53. [53]

    U- stydit: Ultra-high quality artistic style transfer using diffu- sion transformers.arXiv preprint arXiv:2503.08157, 2025

    Zhang, Z., Ma, A., Cao, K., Wang, J., Liu, S., Ma, Y., Cheng, B., Leng, D., Yin, Y.: U-stydit: Ultra-high quality artistic style transfer using diffusion transformers (2025),https://arxiv.org/abs/2503.08157 20 Yang et al. A Details of CSG-Dataset A.1 Dataset Generation Pipeline To meet the unique data requirements of high-fidelity style transfer, we desig...

    work page arXiv 2025