Pith. sign in

REVIEW 4 major objections 5 minor 5 cited by

OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data

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 A plug-and-play consistency module lets arbitrary style LoRAs restyle images without losing identity or fine details.

desk verdict A genuinely useful plug-and-play consistency module for Flux stylization, but the GPT-4o comparison is overstated and the evaluation has a circularity problem that needs fixing before publication. read the letter →

arxiv 2505.18445 v1 pith:F2KXQ4WI submitted 2025-05-24 cs.CV

Yiren Song , Cheng Liu , Mike Zheng Shou This is my paper
classification cs.CV
keywords imagestylizationconsistencypreservationDiffusionTransformerLoRAstyle-agnosticGPT-4oFluxpaireddata
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

Image stylization with diffusion models usually faces a trade-off: style LoRAs capture a look, but image-to-image pipelines that preserve structure often dilute that look, and preserving identity and fine details in complex scenes remains unreliable. The paper proposes OmniConsistency, a consistency plugin for the Flux Diffusion Transformer that claims to keep identity, composition, and fine details intact regardless of which style LoRA is applied. It is trained on 2,600 GPT-4o-generated, human-filtered image pairs spanning 22 styles, using a two-stage strategy that first trains style LoRAs and then trains the consistency module while cycling through those LoRAs. The authors report that the plugin, which adds only a few percent of inference overhead, reaches visual coherence and aesthetic quality comparable to GPT-4o on their benchmark.

What carries the argument

The central object is the Consistency LoRA Module: a low-rank adaptation applied to the QKV projections of the condition branch only, paired with a causal attention mask that makes condition tokens read-only for the text and noise branches. It is trained with the Rolling LoRA Bank, a mechanism that switches the active style LoRA and its data subset every 50 iterations during consistency training, forcing the consistency weights to be style-agnostic. Auxiliary machinery includes conditional token mapping, which aligns low-resolution condition tokens to high-resolution output positions, and feature reuse, which caches condition-branch keys and values across denoising steps to cut inference cost.

What would settle it

Take a set of portraits with known identities, stylize them with an unseen style LoRA using OmniConsistency, and compute face-identity similarity (for example, FaceNet embeddings) between input and output; if it does not beat the Flux image-to-image with ControlNet baseline, the consistency claim collapses. A second check: compare style fidelity against the corresponding text-to-image reference; if OmniConsistency's output is measurably less stylish than bare Flux text-to-image with the same LoRA, the no-style-degradation claim is false.

Watch

Extended reading notes

Core claim

OmniConsistency is a branch-isolated consistency module: a low-rank adaptation applied only to the condition branch of a Diffusion Transformer, with causal attention that lets the denoising and text branches read condition tokens but blocks condition tokens from polluting them. Because the consistency parameters never touch the style LoRA's attachment points, the module is style-agnostic and can be bolted onto arbitrary style LoRAs under Flux. The training scheme is the second half of the discovery: style LoRAs are trained first on per-style data, then the consistency LoRA is trained while style LoRAs from a bank are swapped every 50 steps, so the module learns content preservation rather than a particular aesthetic. Trained this way, the plugin reports style-fidelity and content-consistency scores that beat the Flux image-to-image, Redux, and ControlNet baselines and approach GPT-4o.

Load-bearing premise

The validity of the GPT-4o-generated, human-filtered paired dataset as ground truth: if GPT-4o silently changes identity, composition, or fine details in ways annotators miss, the consistency module learns that biased mapping and its style-agnostic generalization fails.

Editorial extensions

If this is right

  • Any community style LoRA can be combined with OmniConsistency without retraining, so the stylization quality of existing LoRA ecosystems transfers directly to image-to-image editing.
  • Because the consistency module is decoupled from style, a single trained plugin should work across both the training LoRA bank and unseen styles, removing the need for per-style consistency models.
  • Open-source image-to-image stylization could approach the consistency of proprietary APIs like GPT-4o at a fraction of the inference cost.
  • The released paired dataset and benchmark give the field a standardized way to measure stylization consistency, making future comparisons between methods more meaningful.

Reading between the lines

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

  • If the consistency mapping learned from GPT-4o's stylization behavior generalizes beyond the 22 training styles, the same data-driven recipe could distill other proprietary model behaviors, such as instruction-following editing, into lightweight open-source plugins.
  • The read-only causal conditioning design is not Flux-specific in principle, so porting it to other Diffusion Transformer backbones could give those ecosystems the same plug-and-play consistency without redesigning control modules.
  • A testable prediction is that combining OmniConsistency with reference-based adapters like IP-Adapter, which the paper demonstrates qualitatively, should preserve identity even for unseen IP-Adapter embeddings; a user study on that setting would quantify the claim.
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. This paper introduces OmniConsistency, a plug-and-play consistency module for Flux-based image stylization. The method first trains a bank of style LoRAs, then trains a condition-branch-only consistency LoRA on paired original/stylized images using a rolling LoRA-bank loader that switches styles every 50 steps. The architecture employs low-resolution condition tokens with position interpolation, cached condition key-values for feature reuse, and a one-way attention mask that lets the main branch read condition tokens while preventing condition tokens from attending to noise or text tokens. The paired dataset (2,600 pairs across 22 styles) is generated by GPT-4o and filtered by human annotators. Experiments compare against Flux I2I/T2I pipelines with Redux or ControlNet and against GPT-4o, using DreamSim, CLIP, FID, CMMD, GPT-4o scores, and a user study. The headline claim is that OmniConsistency achieves performance comparable to GPT-4o while preserving identity, composition, and fine details for arbitrary style LoRAs.

Significance. If the style-agnostic consistency claim is substantiated, the paper offers a practically valuable plug-and-play component: users could combine it with arbitrary style LoRAs under Flux without retraining, and the decoupled two-stage training plus rolling LoRA bank is a sensible way to reduce style-content entanglement. The paper also reports efficiency overhead (4.6% memory, 5.3% time), includes several independent metrics beyond the GPT-4o score (DreamSim, CLIP, FID, CMMD), and provides a 30-participant user study; the promised public release of code and dataset would support reproducibility. However, the central comparison to GPT-4o is only partially supported, and the dataset's role as ground truth needs a quantitative audit, so the significance of the contribution hinges on the revision.

major comments (4)
  1. [§3.5, Table 1, Appendix 7.1] The evaluation of the headline claim is partially circular. The paired training data in §3.5 are produced entirely by GPT-4o, and the 'GPT-4o ↑' content-consistency column in Table 1 is assigned by GPT-4o using the prompt in Appendix 7.1.1. Since the model is trained to imitate GPT-4o's stylization, a high GPT-4o score can reflect agreement with the teacher rather than genuine content preservation. The revision should (i) report inter-annotator or human-vs-GPT-4o agreement on the scoring protocol, (ii) provide the exact scoring prompt and a calibration sample, and (iii) add a human content-consistency study focused on identity and fine details. In addition, Table 1 contains no style-consistency entries for GPT-4o and Appendix 7.2.1 states that style was not compared with GPT-4o because it does not support style LoRA injection; therefore the abstract's 'performance comparable to GPT-4o' is currently supported only for content, not for style.
  2. [§3.5 (dataset filtering)] The 5,000-to-2,600 pair filtering step is the sole quality gate for the training pairs, but the paper reports no quantitative validation of it. There are no per-rejection-category statistics, no inter-annotator agreement, and no automated check that identity, person count, small-object, or text-level distortions were actually removed. This matters because Fig. 8(b) shows that GPT-4o itself omits people in group scenes, so biased pairs that pass the filter would teach the module to reproduce the same bias and break the 'style-agnostic' claim for unseen LoRAs. The revision should add a filtering audit (e.g., rejection rates by category, identity embedding similarity before and after filtering, person-count validation) or release the pairs immediately so the community can audit them.
  3. [Table 2] Table 2 contains implausible values: the 'w/o Rolling LoRA Bank' and 'w/o Decoupled Training' rows report DreamSim of 24.98 and 21.06, while the full model reports 0.181 and all other DreamSim values in Table 1 are in the 0.18–0.32 range. These look like missing decimal points or a different metric, and as printed they make the quantitative ablation unreliable. Please correct the table and re-run any affected analysis.
  4. [§4.2 and §4.4] The style-consistency protocol compares the output with a LoRA-conditional reference generated from the same prompt and seed, which is a reasonable proxy, but it cannot be applied to GPT-4o. The qualitative statement in §4.4 that 'our method ... produces results comparable to GPT-4o' therefore rests on content and overall quality rather than on style fidelity. The authors should either define and report a verifiable GPT-4o style-consistency comparison (with the prompt-approximation caveats stated in Appendix 7.2.1) or explicitly limit the abstract and conclusion to 'comparable content consistency' and 'better style consistency than the open-source baselines'.
minor comments (5)
  1. [§3.2 vs §4.1] The style LoRA learning rate is given as 1×10⁻³ in §3.2 and 1×10⁻⁴ in §4.1; please reconcile the two values.
  2. [References] Reference [41] for CMMD is the wrong paper: the cited KDD root-cause-analysis paper is unrelated to the CMMD image metric used in Table 1; please replace it with the correct source (e.g., the CMMD paper by Sadekar et al.).
  3. [§3.3] The term 'causal attention' is used for a block-diagonal conditioning mask, which is not causal attention in the usual sequence-modeling sense; please provide the explicit attention mask and clarify the ordering, since Figure 2 alone is ambiguous.
  4. [Throughout] Several typos and incomplete sentences should be fixed: 'video gneration' in §2.1, 'enhance achieve' and 'efficien consistency control' in §3.1, 'houWe propose' at the start of §6, and 'Comparation results' in the caption of Figure 4.
  5. [§6 vs §7.3] The conclusion states that the method 'preserves identity, composition, and fine-grained details', but the failure cases in §7.3 show person-count omissions and small-region artifacts; please add a qualifier such as 'in most tested cases' to avoid overclaiming.

Circularity Check

1 steps flagged · score 4.0 of 10

GPT-4o is both the generator of the paired consistency training data and the judge of the headline content-consistency metric, making the 'comparable to GPT-4o' claim partially self-referential; independent metrics keep the core method from being wholly circular.

  1. self definitional [Section 3.5 (Dataset Collection), Section 4.2 / Table 1 (Evaluation Metrics), Appendix 7.1.1, Contribution 3]
    "We construct a high-quality paired dataset entirely through GPT-4o-driven generation ... we introduce a standardized evaluation protocol based on GPT-4o ... All image similarity metrics are computed using DreamSim, CLIP Image Score, and GPT-4o Score."

    The training target for 'consistency' is defined by GPT-4o-generated stylized pairs (human-filtered). The headline content-consistency metric is a GPT-4o Score assigned by the same model. Thus the model is trained to reproduce GPT-4o's stylization behavior and then judged by GPT-4o; the 'comparable to GPT-4o' result in Table 1 (4.52 vs 4.57) is partly a self-comparison rather than an independent measurement. The circularity is partial because DreamSim, CLIP-I, FID, CMMD, and a user study are also reported and favor the method, so the central claim does not reduce entirely to the GPT-4o loop.

full rationale

The paper's central training scheme (two-stage decoupled training, rolling LoRA bank, condition-branch consistency LoRA, conditional token mapping) is self-contained and does not reduce to its inputs. No uniqueness theorem or load-bearing self-citation is invoked; self-citations appear only in related-work context. The only circular element is the evaluation loop: GPT-4o both generated the paired consistency ground truth and provides the GPT-4o Score used in the headline comparison. Because the paper also reports independent metrics (DreamSim, CLIP, FID, CMMD, user study) and tests unseen style LoRAs, the method's core claims retain independent support. Score 4 reflects partial, not total, circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper relies on several hand-chosen hyperparameters and domain assumptions about the Flux backbone, the validity of GPT-4o as a consistency oracle, and the decoupling of style and content. No new physical entities are introduced; the reported modules are architectural designs, not independent postulated entities.

free parameters (6)
  • LoRA rank r = Not reported
    Rank of the consistency LoRA; capacity of the adapter, a hand-chosen hyperparameter with no ablation.
  • Style LoRA learning rate = 1e-3 (stage 1)
    Set in Section 4.1 without sensitivity analysis.
  • Consistency LoRA learning rate = 1e-4 (stage 2)
    Set in Section 4.1; no sensitivity analysis.
  • LoRA bank switching interval = 50 steps
    Every 50 steps a style LoRA and its data are loaded (Section 4.1); no ablation on this interval.
  • Number of styles and pairs = 22 styles, 2,600 pairs
    Dataset scale chosen by annotation budget; the paper does not analyze whether this covers the style space.
  • Condition image resolution = 512x512 for a 1024x1024 output
    Chosen to reduce memory; the token mapping assumes this is sufficient for spatial alignment.
assumptions (5)
  • domain assumption Flux 1.0 dev is a suitable pretrained backbone and LoRA can encode styles without degrading other capabilities.
    Used throughout, especially Sections 3.2 and 4.1.
  • ad hoc to paper Decoupling style learning from consistency learning via freezing style LoRAs and training only condition-branch LoRA yields a style-agnostic consistency module.
    Central design premise of Sections 3.2 and 3.3; not independently validated beyond this paper.
  • ad hoc to paper Causal attention where condition tokens cannot attend to noise/text tokens prevents interference with style and preserves editable generation.
    Proposed in Section 3.3; the benefit is stated but not shown with a dedicated ablation.
  • domain assumption GPT-4o-generated image pairs are reliable ground truth for structural and semantic consistency after human filtering.
    Dataset construction in Section 3.5; if the oracle is biased, the training target is biased.
  • ad hoc to paper A 512x512 condition with token position interpolation preserves pixel-level correspondence for 1024x1024 generation.
    Conditional Token Mapping in Section 3.4; no explicit measurement of alignment error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data." pith.science (2026). https://pith.science/paper/F2KXQ4WI

@misc{pith2026250518445,
  author       = {Pith},
  title        = {Pith review of: OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2KXQ4WI}},
  note         = {Machine review of arXiv:2505.18445}
}
read the original abstract

Diffusion models have advanced image stylization significantly, yet two core challenges persist: (1) maintaining consistent stylization in complex scenes, particularly identity, composition, and fine details, and (2) preventing style degradation in image-to-image pipelines with style LoRAs. GPT-4o's exceptional stylization consistency highlights the performance gap between open-source methods and proprietary models. To bridge this gap, we propose \textbf{OmniConsistency}, a universal consistency plugin leveraging large-scale Diffusion Transformers (DiTs). OmniConsistency contributes: (1) an in-context consistency learning framework trained on aligned image pairs for robust generalization; (2) a two-stage progressive learning strategy decoupling style learning from consistency preservation to mitigate style degradation; and (3) a fully plug-and-play design compatible with arbitrary style LoRAs under the Flux framework. Extensive experiments show that OmniConsistency significantly enhances visual coherence and aesthetic quality, achieving performance comparable to commercial state-of-the-art model GPT-4o.

Figures

Figures reproduced from arXiv: 2505.18445 by the authors.

Figure 1
Figure 1. Our method achieves style-consistent and structure-preserving image stylization under [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of OmniConsistency, consisting of style learning and consistency learning [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. OmniConsistency can be combined with both seen and unseen style LoRA modules to [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparation results of OmniConsistency and baseline methods. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Ablation shows that full settings ensure [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: User study: Preference rates for style and content consistency across methods. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Failure cases. was counted as one point, and the percentage score for each model was calculated based on the total number of selections. As shown in [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: More Comparation results. 7.4 More Results We present additional experimental results in this section [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: More image stylization results of OmniConsistency. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: More image stylization results of OmniConsistency. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Ghosts Beneath Textures: Texture-Relation Cues for Cross-Paradigm AI-Generated Image Detection

    cs.CV 2026-07 conditional novelty 6.5 of 10

    Semantics-irrelevant local-global texture relations, extracted after content suppression, serve as cross-paradigm forensic cues that let DTS-Det reach 99.6% accuracy on a new mixed-generation benchmark.

  2. ReMoT: Reinforcement Learning with Motion Contrast Triplets

    cs.CV 2026-02 conditional novelty 6.0 of 10

    Training a 4B vision-language model on rule-generated motion-contrast triplets with GRPO lifts spatio-temporal QA accuracy by about 17 points on the authors' own benchmark and by smaller margins on standard benchmarks.

  3. WorldWander: Bridging Egocentric and Exocentric Worlds in Video Generation

    cs.CV 2025-11 conditional novelty 5.0 of 10

    A bidirectional egocentric-to-exocentric video translation framework trained with in-context attention on a new synthetic+real dataset, with evaluation flaws around reference leakage and missing direct baselines.

  4. FeRA: Frequency-Energy Constrained Routing for Effective Diffusion Adaptation Fine-Tuning

    cs.CV 2025-11 conditional novelty 5.0 of 10

    A frequency-energy router that blends LoRA experts according to the latent's bandwise energy improves diffusion fine-tuning quality and style consistency across multiple backbones.

  5. RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers

    cs.CV 2025-06 conditional novelty 5.0 of 10

    A decoupled-attention adapter transfers image-pair edits to new photos in diffusion transformers, trained with a new 218-task visual editing dataset.

Reference graph

Works this paper leans on

48 extracted references · 6 canonical work pages · cited by 5 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023

    arXiv 2023

  3. [3]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392–18402, 2023

    2023

  4. [4]

    Consislora: Enhancing content and style consistency for lora-based style transfer

    Bolin Chen, Baoquan Zhao, Haoran Xie, Yi Cai, Qing Li, and Xudong Mao. Consislora: Enhancing content and style consistency for lora-based style transfer. arXiv preprint arXiv:2503.10614, 2025

    arXiv 2025

  5. [5]

    Transanimate: Taming layer diffusion to generate rgba video

    Xuewei Chen, Zhimin Chen, and Yiren Song. Transanimate: Taming layer diffusion to generate rgba video. arXiv preprint arXiv:2503.17934, 2025

    arXiv 2025

  6. [6]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

    2024

  7. [7]

    Implicit style-content separation using b-lora

    Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In European Conference on Computer Vision, pages 181–198. Springer, 2024

    2024

  8. [8]

    Dreamsim: Learning new dimensions of human visual similarity using synthetic data

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023

    arXiv 2023

Show all 48 references
  1. [9]

    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

  2. [10]

    Any2anytryon: Leveraging adaptive position embeddings for versatile virtual clothing tasks

    Hailong Guo, Bohan Zeng, Yiren Song, Wentao Zhang, Chuang Zhang, and Jiaming Liu. Any2anytryon: Leveraging adaptive position embeddings for versatile virtual clothing tasks. arXiv preprint arXiv:2501.15891, 2025

    2025 arXiv

  3. [11]

    Animated- iff: Animate your personalized text-to-image diffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animated- iff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023

    2023 arXiv

  4. [12]

    Prompt-to- prompt image editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to- prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022

    2022 arXiv

  5. [13]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021

    2021 arXiv

  6. [14]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  7. [15]

    Denoising diffusion probabilistic models

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

    2020

  8. [16]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. ICLR, 2022

    2022

  9. [17]

    PhotoDoodle: Learning artistic image editing from few-shot pairwise data

    Shijie Huang, Yiren Song, Yuxuan Zhang, Hailong Guo, Xueyin Wang, Mike Zheng Shou, and Ji- aming Liu. PhotoDoodle: Learning artistic image editing from few-shot pairwise data. arXiv preprint arXiv:2502.14397, 2025

    2025 arXiv

  10. [18]

    Multi-concept customization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931–1941, 2023

    1931

  11. [19]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 10

    2024

  12. [20]

    Flux-redux

    Black Forest Labs. Flux-redux. https://huggingface.co/black-forest-labs/FLUX. 1-Redux-dev, 2024

    2024

  13. [21]

    Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024

    2024 arXiv

  14. [22]

    Liblibai, 2025

    LiblibAI. Liblibai, 2025

    2025

  15. [23]

    Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021

    2021 arXiv

  16. [24]

    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. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 4...

    2024

  17. [25]

    Multi-modal attention for speech emotion recognition

    Zexu Pan, Zhaojie Luo, Jichen Yang, and Haizhou Li. Multi-modal attention for speech emotion recognition. arXiv preprint arXiv:2009.04107, 2020

    2009 arXiv

  18. [26]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023

    2023

  19. [27]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  20. [28]

    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. In ICML, 2021

    2021

  21. [29]

    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. In CVPR, 2022

    2022

  22. [30]

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

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 225...

    2023

  23. [31]

    Ziplora: Any subject in any style by effectively merging loras

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422–438. Springer, 2024

    2024

  24. [32]

    Layertracer: Cognitive-aligned layered svg synthesis via diffusion transformer

    Yiren Song, Danze Chen, and Mike Zheng Shou. Layertracer: Cognitive-aligned layered svg synthesis via diffusion transformer. arXiv preprint arXiv:2502.01105, 2025

    2025 arXiv

  25. [33]

    Processpainter: Learn painting process from sequence data

    Yiren Song, Shijie Huang, Chen Yao, Xiaojun Ye, Hai Ci, Jiaming Liu, Yuxuan Zhang, and Mike Zheng Shou. Processpainter: Learn painting process from sequence data. arXiv preprint arXiv:2406.06062, 2024

    2024 arXiv

  26. [34]

    Makeanything: Harnessing diffusion transformers for multi-domain procedural sequence generation

    Yiren Song, Cheng Liu, and Mike Zheng Shou. Makeanything: Harnessing diffusion transformers for multi-domain procedural sequence generation. arXiv preprint arXiv:2502.01572, 2025

    2025 arXiv

  27. [35]

    Ominicontrol: Minimal and universal control for diffusion transformer

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer. arXiv preprint arXiv:2411.15098, 2024

    2024 arXiv

  28. [36]

    Ominicontrol2: Efficient conditioning for diffusion transformers

    Zhenxiong Tan, Qiaochu Xue, Xingyi Yang, Songhua Liu, and Xinchao Wang. Ominicontrol2: Efficient conditioning for diffusion transformers. arXiv preprint arXiv:2503.08280, 2025

    2025 arXiv

  29. [37]

    Grid: Visual layout generation

    Cong Wan, Xiangyang Luo, Zijian Cai, Yiren Song, Yunlong Zhao, Yifan Bai, Yuhang He, and Yihong Gong. Grid: Visual layout generation. arXiv preprint arXiv:2412.10718, 2024

    2024 arXiv

  30. [38]

    Instantstyle: Free lunch towards style-preserving in text-to-image generation

    Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024

    2024 arXiv

  31. [39]

    Styleadapter: A unified stylized image generation model

    Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A unified stylized image generation model. arXiv preprint arXiv:2309.01770, 2023

    2023 arXiv

  32. [40]

    Stylealign: Analysis and applications of aligned stylegan models

    Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 11

    2021 arXiv

  33. [41]

    Cmmd: Cross-metric multi-dimensional root cause analysis

    Shifu Yan, Caihua Shan, Wenyi Yang, Bixiong Xu, Dongsheng Li, Lili Qiu, Jie Tong, and Qi Zhang. Cmmd: Cross-metric multi-dimensional root cause analysis. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4310–4320, 2022

    2022

  34. [42]

    Editworld: Simulating world dynamics for instruction-following image editing

    Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. Editworld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024

    2024 arXiv

  35. [43]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023

    2023 arXiv

  36. [44]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023

    2023

  37. [45]

    Ssr-encoder: Encoding selective subject representation for subject-driven generation

    Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2024

  38. [46]

    Stable-makeup: When real-world makeup transfer meets diffusion model

    Yuxuan Zhang, Lifu Wei, Qing Zhang, Yiren Song, Jiaming Liu, Huaxia Li, Xu Tang, Yao Hu, and Haibo Zhao. Stable-makeup: When real-world makeup transfer meets diffusion model. arXiv preprint arXiv:2403.07764, 2024

    2024 arXiv

  39. [47]

    Easycontrol: Adding efficient and flexible control for diffusion transformer

    Yuxuan Zhang, Yirui Yuan, Yiren Song, Haofan Wang, and Jiaming Liu. Easycontrol: Adding efficient and flexible control for diffusion transformer. arXiv preprint arXiv:2503.07027, 2025

    2025 arXiv

  40. [48]

    style_consistency

    Yuxuan Zhang, Qing Zhang, Yiren Song, and Jiaming Liu. Stable-hair: Real-world hair transfer via diffusion model. arXiv preprint arXiv:2407.14078, 2024. 12 7 Appendix 7.1 Implementation Details of the GPT-4o Evaluation In the GPT-4o evaluation process, we establish specific me...

    2024 arXiv

Pith tools

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