Pith. sign in

REVIEW 3 major objections 5 minor 76 references

QR-LoRA: Efficient and Disentangled Fine-tuning via QR Decomposition for Customized Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read QR-LoRA claims that fixing Q and R from a QR decomposition of the pretrained weights and training only a residual matrix ΔR halves LoRA's trainable parameters and keeps merged content and style adaptations from interfering.

desk verdict A coherent but incremental PEFT variant that likely reduces to SVD-initialized LoRA with a frozen factor; the disentanglement evidence is suggestive, not conclusive, yet the paper deserves a serious referee. read the letter →

arxiv 2507.04599 v2 pith:2CFTOCFZ submitted 2025-07-07 cs.CV

Jiahui Yang , Yongjia Ma , Donglin Di , Hao Li , Wei Chen , Yan Xie , Jianxun Cui , Xun Yang
show 1 more author
Wangmeng Zuo
This is my paper
classification cs.CV
keywords QRdecompositionLoRAparameter-efficientfine-tuningcontent-styledisentanglementtext-to-imagegenerationdiffusionmodelsmodelmergingorthogonaladaptation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes QR-LoRA, a fine-tuning method for text-to-image diffusion models that aims to solve the entanglement problem that arises when separately trained LoRA adapters for content and style are merged. Its central claim is that if the update is parameterized through a QR decomposition of the pretrained weight matrix, with Q and R frozen, the only trained component—a residual matrix ΔR—can carry task-specific information without interfering with other tasks. The paper reports that this design halves trainable parameters relative to conventional LoRA and that the maximum cosine similarity between ΔR matrices across injection layers stays below 0.2, which it reads as evidence of feature disentanglement. It further reports improved content preservation and style transfer scores, and higher user ratings, across SDXL, SD3, and FLUX.1-dev. The paper's stated limitation is confined to a one-time initialization cost of about one to five minutes per backbone, with no effect on inference efficiency.

What carries the argument

The load-bearing object is the fixed orthogonal decomposition W_core^T = Q R, where Q is an orthogonal basis obtained by reduced QR decomposition of the SVD-based core's right-factor term, and R is the upper-triangular remainder (in the paper, R = R_s T). The trainable residual ΔR sits in that frozen basis, so every update is a linear combination of the same orthogonal feature directions. That structure is what the paper credits for both the parameter reduction (only the r×m entries of ΔR are trained, versus r(m+n) for standard LoRA) and for the near-zero cross-task similarity of ΔR matrices; the element-wise merge rule is the direct payoff.

What would settle it

Measure the projection residual: train a standard full-rank or high-rank LoRA to convergence on a single-image style task, call the update ΔW*, and compute the best relative error ||ΔW* − QX||_F / ||ΔW*||_F over all X. If that error stays large even as the rank grows to the practical limit, the frozen-Q assumption is the bottleneck; repeating the test on a radical style far outside the training distribution should expose the ceiling quickly.

Watch

Extended reading notes

Core claim

The discovery the paper argues for is that LoRA-style adaptation can be made inherently disentangled by construction rather than by post-hoc merging. Starting from the SVD of the pretrained weight W, the paper keeps a rank-r core W_core and writes W_core^T = Q R, with Q orthogonal and R upper triangular. Both Q and R are frozen, and fine-tuning is confined to a task-specific ΔR. Because Q is shared and orthogonal, the argument goes, different tasks' ΔR matrices occupy independent directions; the paper measures this as maximum cosine similarity below 0.2 across all layers. Merging is then plain element-wise addition ΔR_cs = λ_c ΔR_c + λ_s ΔR_s, with no per-layer coefficient optimization. The paper claims this yields better content-style disentanglement than ZipLoRA, B-LoRA, and StyleAligned on SDXL, and beats naive direct merging on SD3 and FLUX.

Load-bearing premise

The method assumes that the changes needed to teach a model a new content or style lie entirely inside the fixed set of feature directions defined by Q, so freezing Q and R loses nothing; if a task needs weight changes in a direction outside that set, QR-LoRA cannot learn them.

Editorial extensions

If this is right

  • Training memory and communication drop to about half of a standard LoRA of the same rank, because only ΔR is updated while Q and R stay frozen.
  • Merging multiple customizations becomes a simple weighted sum of ΔR matrices, eliminating per-layer merging heuristics such as coefficient search.
  • The same recipe can be injected into any backbone that supports LoRA, so the authors' results on SDXL, SD3, and FLUX.1-dev are evidence, on their account, that disentanglement comes from the parameterization rather than from model-specific blocks.
  • Because different tasks' ΔR matrices are near-orthogonal, adding more attributes (object, scene, texture) should remain feasible without retraining earlier adapters, as the paper's multi-feature composition experiments suggest.
  • If the cosine-similarity measure is accepted, QR-LoRA provides a quantitative certificate of disentanglement rather than only a qualitative one.

Reading between the lines

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

  • A testable consequence not drawn in the paper is that the fixed-Q construction caps expressivity: any adaptation requiring weight movement outside the span of the top-r singular-vector directions of the pretrained matrix cannot be learned, and this ceiling should be measurable by projecting a converged full-rank update onto the frozen basis.
  • The near-zero cosine similarities were measured on single-image training pairs; on broader multi-image or multi-concept datasets, task-specific ΔR matrices may become more aligned, and the naive addition rule may need learned scaling or orthogonalization.
  • The same orthogonal-subspace trick could transfer to other parameter-efficient fine-tuning settings, such as multi-task language-model adapters or continual learning, where the bottleneck is cross-task interference; QR-LoRA's fixed basis is a natural prior for those problems.
  • The 0.2 cosine threshold is correlational evidence, not causal proof: an ablation that replaces the SVD-derived Q with a random orthogonal Q of the same shape would show whether the benefit comes from orthogonality itself or from the specific pretrained basis.
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

3 major / 5 minor

Summary. QR-LoRA proposes a parameter-efficient fine-tuning method for text-to-image diffusion models. Given a pretrained weight matrix W, it extracts a rank-r SVD core, defines a complement matrix Wcomp = W - Wcore, and constructs an orthogonal Q and upper-triangular R via reduced QR decomposition of V[:,:r]Σ[:r]. The method freezes Q and R and trains only a task-specific ΔR, so the effective weight update is (QΔR)^T; content and style adaptations are merged by element-wise addition of ΔR_c and ΔR_s. The paper evaluates QR-LoRA on SDXL, SD3, and FLUX.1-dev using qualitative comparisons, image-similarity metrics (DINO/CLIP), and a user study, and claims superior disentanglement with half the trainable parameters of conventional LoRA.

Significance. If the underlying subspace-capacity assumption holds, QR-LoRA is a simple and potentially attractive design: it is architecture-agnostic, reduces trainable parameters relative to LoRA for square weight matrices, and offers an extremely simple merging rule for content and style. The manuscript's strengths include a transparent algorithmic description, experiments across three diffusion backbones, and a user study. However, the central disentanglement claim currently rests on a parameter-space cosine-similarity analysis that is confounded by high dimensionality, and on quantitative differences that are often within one standard deviation. The frozen top-r singular subspace assumption is load-bearing and is not directly validated. These issues are addressable with additional analysis and experiments, so the work is potentially significant but needs revision.

major comments (3)
  1. [Section 3.2, Figures 2 and 3] The near-zero cosine similarity between ΔR matrices is not by itself evidence of feature disentanglement. For ΔR ∈ R^{64×m}, with m equal to an attention or feed-forward dimension, two independent random matrices would typically exhibit cosine similarities near zero without any orthogonalization; no random baseline is reported. Because all tasks share the same frozen Q and R, the reported similarity is computed in the internal parameter basis, not in the model's output or feature space. The paper should report an output-space interference measure, such as the effect of swapping ΔR_c and ΔR_s on generated images, or the cosine similarity of activation changes, and compare against a random-subspace control.
  2. [Algorithm 1, Equations (5) and (6)] The learned update is confined to the span of the top-r right singular vectors of W. Since Q is the orthogonal factor of the reduced QR decomposition of V[:,:r]Σ[:r], its columns span the top-r right singular subspace, and ΔW = (QΔR)^T has every row in that subspace. If a content or style adaptation requires components in the complement of this subspace, QR-LoRA cannot represent them. The convergence plot in Figure 8 and the qualitative results do not quantify this capacity limitation. A direct test is needed, such as projecting a fully trained LoRA update onto this subspace and reporting the residual norm, or comparing QR-LoRA with standard LoRA under identical rank, step count, and data across several tasks.
  3. [Table 1, Section 4.2] Most quantitative differences between QR-LoRA and the strongest baseline are smaller than one standard deviation. For example, on SDXL, DINO-S improves from 0.686±0.131 (ZipLoRA) or 0.689±0.075 (B-LoRA) to 0.694±0.084, and CLIP-S from 0.686±0.105 to 0.707±0.060; these differences are not distinguishable without repeated-seed results and significance tests. The user-study averages are reported without the number of participants, the number of comparisons, or variance. Reporting confidence intervals and paired significance tests would substantiate the claim of superior disentanglement.
minor comments (5)
  1. [References, Section 4.1 and Table 1] The paper cites B-LoRA as [28], but reference [28] is VB-LoRA by Li et al.; B-LoRA is reference [8] by Frenkel et al. This citation mismatch appears in the Related Work section, the experimental setup, and Table 1, and should be corrected throughout.
  2. [Section 3.1 and Section 4.1] The claim that QR-LoRA uses half of conventional LoRA's parameters assumes square weight matrices. For general m×n weights, the fraction is rm/[r(m+n)] = m/(m+n), which is only exactly 1/2 when m=n; this should be stated explicitly.
  3. [Figure 2 caption] The caption should specify how the maximum and minimum cosine similarities are computed, including which layers are included, whether matrices are flattened, and how the 'direct fine-tuning strategy' differs from the ΔR-only update strategy.
  4. [Section 4.2] The test set is described only as '64 randomly sampled generated images'; the number of content-style pairs and the sampling procedure should be clarified so that the metric standard deviations are interpretable.
  5. [Section 5] The initialization overhead is reported as '∼1, 2, and 5 minutes' for SDXL, SD3, and FLUX.1-dev; please specify the hardware, the number of layers, and the exact timing procedure.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential evidence for the ΔR disentanglement proxy; core empirical claims remain externally benchmarked.

  1. self definitional [Section 3.2 'Motivation and Analysis' and Section 3.4; Figures 2-3]
    "The effectiveness of this approach is validated by our empirical results in Figure 2 and Figure 3, which show that the maximum cosine similarity between ∆R matrices across all injection layers remains consistently below 0.2, with mean values stable around 0, indicating superior feature disentanglement."

    The paper uses near-zero cosine similarity between the trainable ΔR matrices both as the operational characterization of feature disentanglement ('i.e., ΔRc ⊥ ΔRs as shown in Figure 2') and as the evidence for that same property. Since only ΔR is trained and the orthogonal basis Q is fixed, the low pairwise similarity of ΔR matrices is a property of the chosen parameterization and of the very high-dimensional parameter space, not an independent consequence of the orthogonal Q matrix. This makes the internal disentanglement validation largely self-referential.

full rationale

The QR-LoRA reparameterization is algebraically exact: Eq. (5) defines Q and R from the top-r SVD factors of W, and Eq. (6) with ΔR=0 gives Wcomp + (QR)^T = W, so the initialization is an identity rather than a fitted prediction. The 'half the trainable parameters' claim is arithmetic (r·m vs. r(m+n) for LoRA) and follows directly from the design. Training ΔR with Q and R frozen is equivalent to LoRA with one factor fixed, so the parameter-efficiency and additive-merging rules are constructional. The only internal loop is the cosine-similarity evidence equating disentanglement with ΔRc ⊥ ΔRs and then measuring that same quantity; because the final DINO/CLIP and user-study benchmarks are external comparisons against prior methods, this loop is not load-bearing. Self-citations [34,35,66] are background or future-work references and do not support any load-bearing premise. No uniqueness theorem or ansatz is imported from the authors' prior work. Score 2 reflects the minor self-referential proxy, not substantive circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or formal entities. It relies on standard matrix decompositions and on two unproven domain assumptions about the sufficiency of the top-r singular subspace and about parameter-space orthogonality implying feature disentanglement.

free parameters (1)
  • rank r = 64
    Chosen by hand for all experiments; determines the dimension of the frozen subspace and hence the capacity of QR-LoRA updates. The parameter-reduction claim (half of LoRA) depends on this rank.
assumptions (4)
  • domain assumption The top-r right singular vectors of the pretrained weight matrix span the directions relevant for content and style adaptation
    Used in Section 3.3 to justify freezing Q and training only the residual matrix; implicitly assumes the desired task-specific update lies in this subspace.
  • domain assumption Low cosine similarity between residual matrices from different tasks indicates that the corresponding visual attributes do not interfere in generation
    Section 3.2 interprets near-zero parameter-space cosine similarity as evidence of feature disentanglement, but the link to generated-image disentanglement is not proven.
  • domain assumption The minimal Frobenius norm property of orthogonal parameterization supports the stability of Q across tasks
    Invoked in Section 1 with proof deferred to the Appendix; the main text does not state the property or its assumptions.
  • standard math SVD and QR decomposition exist and have the standard properties
    Assumed throughout Section 3.3 without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QR-LoRA: Efficient and Disentangled Fine-tuning via QR Decomposition for Customized Generation." pith.science (2026). https://pith.science/paper/2CFTOCFZ

@misc{pith2026250704599,
  author       = {Pith},
  title        = {Pith review of: QR-LoRA: Efficient and Disentangled Fine-tuning via QR Decomposition for Customized Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2CFTOCFZ}},
  note         = {Machine review of arXiv:2507.04599}
}
abstract

Existing text-to-image models often rely on parameter fine-tuning techniques such as Low-Rank Adaptation (LoRA) to customize visual attributes. However, when combining multiple LoRA models for content-style fusion tasks, unstructured modifications of weight matrices often lead to undesired feature entanglement between content and style attributes. We propose QR-LoRA, a novel fine-tuning framework leveraging QR decomposition for structured parameter updates that effectively separate visual attributes. Our key insight is that the orthogonal Q matrix naturally minimizes interference between different visual features, while the upper triangular R matrix efficiently encodes attribute-specific transformations. Our approach fixes both Q and R matrices while only training an additional task-specific $\Delta R$ matrix. This structured design reduces trainable parameters to half of conventional LoRA methods and supports effective merging of multiple adaptations without cross-contamination due to the strong disentanglement properties between $\Delta R$ matrices. Experiments demonstrate that QR-LoRA achieves superior disentanglement in content-style fusion tasks, establishing a new paradigm for parameter-efficient, disentangled fine-tuning in generative models. The project page is available at: https://luna-ai-lab.github.io/QR-LoRA/.

Figures

Figures reproduced from arXiv: 2507.04599 by the authors.

Figure 1
Figure 1. Given a single content and a single style image, we present [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Comparison of cosine similarities between differ￾ent training strategies in QR-LoRA. Visualization of layer-wise cosine similarities between matrices obtained from training on a randomly selected image pair in SDXL model. The comparison demonstrates distinct characteristics between directly fine-tuning Q and R matrices (upper zoom-in) versus only tuning ∆R matri￾ces (lower zoom-in). See Appendix for comprehensive an… view at source ↗
Figure 4
Figure 4. Overview of QR-LoRA framework. Upper (Sec 3.3): Technical illustration of our method compared to traditional fine-tuning paradigms, highlighting our efficient parameter updates through orthogonal decomposition. Lower (Sec 3.4): Application pipeline of our framework for content-style disentanglement, consisting of a disentangled fine-tuning module and a streamlined feature merging module. fine-tuning framework that n… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison. Comparison of our QR-LoRA against state-of-the-art methods on SDXL and a naive baseline on SD3 and FLUX.1-dev models, demonstrating the model-agnostic nature and superior performance of our framework. Zoom in to view details [PITH_FULL_IMAGE:fi…
Figure 8
Figure 8. Figure 8: Convergence analysis. Comparison of training conver￾gence between QR-LoRA and LoRA. Scene Object Style Results Scene Object Results Scene Human Results Object1 Object2 Results Object Texture Results Object Texture Results Scene Object Results [PITH_FULL_IMAGE:figures/…
Figure 7
Figure 7. Figure 7: Ablation study on scaling coefficient λc and λs. Ablation. To further validate the robustness of our method, we experiment with different combinations of scal￾ing coefficients λc and λs during feature composition. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 46 canonical work pages

  1. [1]

    Cross-image attention for zero- shot appearance transfer

    Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch- Elor, and Daniel Cohen-Or. Cross-image attention for zero- shot appearance transfer. In ACM SIGGRAPH 2024 Confer- ence Papers, pages 1–12, 2024. 3

    2024

  2. [2]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 6, 7

    work page 2021

  3. [3]

    Training-free layout control with cross-attention guidance

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373, 2023. 3

    arXiv 2023

  4. [4]

    The approximation of one ma- trix by another of lower rank

    Carl Eckart and Gale Young. The approximation of one ma- trix by another of lower rank. Psychometrika, 1(3):211–218,

    work page

  5. [5]

    Diffusion self-guidance for control- lable image generation

    Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control- lable image generation. Advances in Neural Information Processing Systems, 36:16222–16239, 2023. 3

    work page 2023

  6. [6]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learn- ing, 2024. 2, 3, 6, 7

    work page 2024

  7. [7]

    Lora-x: Bridging foundation models with training-free cross-model adaptation

    Farzad Farhadzadeh, Debasmit Das, Shubhankar Borse, and Fatih Porikli. Lora-x: Bridging foundation models with training-free cross-model adaptation. In ICLR, 2025. 3

    work page 2025

  8. [8]

    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, 2025. 2, 3

    work page 2025

Show all 76 references
  1. [9]

    An image is worth one word: Personalizing text-to- image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patash- nik, Amit H Bermano, Gal Chechik, and Daniel Cohen- Or. An image is worth one word: Personalizing text-to- image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3

    2022 arXiv

  2. [10]

    Singular value de- composition and least squares solutions

    Gene H Golub and Christian Reinsch. Singular value de- composition and least squares solutions. In Handbook for Automatic Computation: Volume II: Linear Algebra , pages 134–151. Springer, 1971. 2

    1971

  3. [11]

    Matrix computations johns hopkins university press

    Gene H Golub and Charles F Van Loan. Matrix computations johns hopkins university press. Baltimore and London, 41: 62, 1996. 2

    1996

  4. [12]

    Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models

    Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Chen Yunpeng, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, Yixiao Ge, Shan Ying, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. NeurIPS,

  5. [13]

    Svdiff: Compact param- eter space for diffusion fine-tuning

    Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact param- eter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 7323–7334, 2023. 2

    2023

  6. [14]

    Prompt-to-prompt im- age editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im- age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3

    2022 arXiv

  7. [15]

    Style aligned image generation via shared atten- tion

    Amir Hertz, Andrey V oynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared atten- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 4775–4785,

  8. [16]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In NeurIPS, pages 6840–6851,

  9. [17]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4

    2021 arXiv

  10. [18]

    Zero-shot text-guided object gen- eration with dream fields

    Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object gen- eration with dream fields. In CVPR, pages 867–876, 2022. 2

    2022

  11. [19]

    Decor:decomposition and projection of text embeddings for text-to-image cus- tomization, 2024

    Geonhui Jang, Jin-Hwa Kim, Yong-Hyun Park, Junho Kim, Gayoung Lee, and Yonghyun Jeong. Decor:decomposition and projection of text embeddings for text-to-image cus- tomization, 2024. 3

    2024

  12. [20]

    Visual style prompting with swapping self- attention

    Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self- attention. arXiv preprint arXiv:2402.12974, 2024. 2, 3

    2024 arXiv

  13. [21]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, pages 26565–26577, 2022. 2

    2022

  14. [22]

    Understanding diffusion objectives as the elbo with simple data augmentation

    Diederik Kingma and Ruiqi Gao. Understanding diffusion objectives as the elbo with simple data augmentation. Ad- vances in Neural Information Processing Systems, 36, 2024. 3

    2024

  15. [23]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 6

    2014 arXiv

  16. [24]

    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 (CVPR), 2023. 3

    2023

  17. [25]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2023. 2, 3, 6, 7

    2023

  18. [26]

    Stylestudio: Text-driven style transfer with selective control of style elements

    Mingkun Lei, Xue Song, Beier Zhu, Hao Wang, and Chi Zhang. Stylestudio: Text-driven style transfer with selective control of style elements. In CVPR, 2025. 3

    2025

  19. [27]

    The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021. 2

    2021 arXiv

  20. [28]

    Vb-lora: Extreme parameter efficient fine-tuning with vector banks

    Yang Li, Shaobo Han, and Shihao Ji. Vb-lora: Extreme parameter efficient fine-tuning with vector banks. In The 38th Conference on Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 6, 7

    2024

  21. [29]

    Svft: Parameter-efficient fine-tuning with singular vectors

    Vijay Lingam, Atula Tejaswi, Aditya Vavre, Aneesh Shetty, Gautham Krishna Gudur, Joydeep Ghosh, Alex Dimakis, Eunsol Choi, Aleksandar Bojchevski, and Sujay Sanghavi. Svft: Parameter-efficient fine-tuning with singular vectors. arXiv preprint arXiv:2405.19597, 2024. 2, 3

    2024 arXiv

  22. [30]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. arXiv preprint arXiv:2210.02747, 2022. 3

    2022 arXiv

  23. [31]

    Unziplora: Separating content and style from a single image

    Chang Liu, Viraj Shah, Aiyu Cui, and Svetlana Lazebnik. Unziplora: Separating content and style from a single image. arXiv preprint arXiv:2412.04465, 2024. 2, 3

    2024 arXiv

  24. [32]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3

    2022 arXiv

  25. [33]

    Adamole: Fine-tuning large lan- guage models with adaptive mixture of low-rank adaptation experts

    Zefang Liu and Jiahua Luo. Adamole: Fine-tuning large lan- guage models with adaptive mixture of low-rank adaptation experts. arXiv preprint arXiv:2405.00361, 2024. 8

    2024 arXiv

  26. [34]

    Tuning-free long video generation via global-local collaborative diffu- sion, 2025

    Yongjia Ma, Junlin Chen, Donglin Di, Qi Xie, Lei Fan, Wei Chen, Xiaofei Gou, Na Zhao, and Xun Yang. Tuning-free long video generation via global-local collaborative diffu- sion, 2025. 2

    2025

  27. [35]

    Adams bashforth moulton solver for inversion and editing in rectified flow, 2025

    Yongjia Ma, Donglin Di, Xuan Liu, Xiaokai Chen, Lei Fan, Wei Chen, and Tonghua Su. Adams bashforth moulton solver for inversion and editing in rectified flow, 2025. 2

    2025

  28. [36]

    Pissa: Prin- cipal singular values and singular vectors adaptation of large language models

    Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Prin- cipal singular values and singular vectors adaptation of large language models. In NeurIPS, 2024. 2

    2024

  29. [37]

    Null-text inversion for editing real im- ages using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real im- ages using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3

    2023

  30. [38]

    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, pages 4296–4304, 2024. 2, 3

    2024

  31. [39]

    PACE: marrying the generalization of PArameter-efficient fine-tuning with con- sistency regularization

    Yao Ni, Shan Zhang, and Piotr Koniusz. PACE: marrying the generalization of PArameter-efficient fine-tuning with con- sistency regularization. In The Thirty-eighth Annual Confer- ence on Neural Information Processing Systems, 2024. 2

    2024

  32. [40]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. pages 8162–8171. PMLR, 2021. 2

    2021

  33. [41]

    Finding and editing multi-modal neurons in pre-trained transformers

    Haowen Pan, Yixin Cao, Xiaozhi Wang, Xun Yang, and Meng Wang. Finding and editing multi-modal neurons in pre-trained transformers. In Findings of the Association for Computational Linguistics ACL 2024, pages 1012–1037, Bangkok, Thailand and virtual meeting, 2024. Association f...

    2024

  34. [42]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195–4205, 2023. 2

    2023

  35. [43]

    Orthogonal adaptation for modular customization of diffusion models

    Ryan Po, Guandao Yang, Kfir Aberman, and Gordon Wet- zstein. Orthogonal adaptation for modular customization of diffusion models. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 7964–7973, 2024. 3

    2024

  36. [44]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 6, 7

    2023 arXiv

  37. [45]

    Learning transferable visual models from natural language supervi- sion

    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 supervi- sion. In International conference on machine learning, ...

    2021

  38. [46]

    Dreambooth3d: Subject-driven text-to-3d generation

    Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In CVPR, pages 2349– 2359, 2023. 8

    2023

  39. [47]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, pages 10684– 10695, 2022. 2, 3

    2022

  40. [48]

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

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: 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 2250...

    2023

  41. [49]

    Ziplora: Any subject in any style by effectively merging loras

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svet- lana 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, 2025. 2, 3, 4, 6, 7

    2025

  42. [50]

    Unleashing the power of task-specific directions in parameter efficient fine-tuning

    Chongjie Si, Zhiyi Shi, Shifan Zhang, Xiaokang Yang, Hanspeter Pfister, and Wei Shen. Unleashing the power of task-specific directions in parameter efficient fine-tuning. In The Thirteenth International Conference on Learning Rep- resentations, 2025. 2

    2025

  43. [51]

    Loraclr: Contrastive adaptation for customization of diffusion models

    Enis Simsar, Thomas Hofmann, Federico Tombari, and Pinar Yanardag. Loraclr: Contrastive adaptation for customization of diffusion models. CVPR, 2025. 3

    2025

  44. [52]

    Styledrop: Text-to-image synthesis of any style

    Kihyuk Sohn, Lu Jiang, Jarred Barber, Kimin Lee, Nataniel Ruiz, Dilip Krishnan, Huiwen Chang, Yuanzhen Li, Irfan Essa, Michael Rubinstein, et al. Styledrop: Text-to-image synthesis of any style. Advances in Neural Information Pro- cessing Systems, 36, 2024. 2, 3, 6

    2024

  45. [53]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In ICLR, 2021. 2

    2021

  46. [54]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. In ICLR, 2021. 2, 3

    2021

  47. [55]

    Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks

    Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5227–5237,

  48. [56]

    Genmo Team. Mochi 1. https: // github. com/ genmoai/models, 2024. 8

    2024

  49. [57]

    Hydralora: An asymmetric lora architec- ture for efficient fine-tuning

    Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. Hydralora: An asymmetric lora architec- ture for efficient fine-tuning. In Advances in Neural Infor- mation Processing Systems (NeurIPS), 2024. 2

    2024

  50. [58]

    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. 3

    1921

  51. [59]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3

    2017

  52. [60]

    Diffusers: State-of-the-art diffu- sion models

    Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffu- sion models. https://github.com/huggingface/ diffusers...

    2022

  53. [61]

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

    Haofan Wang, 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. 2, 3

    2024 arXiv

  54. [62]

    Uncovering the disentanglement capability in text- to-image diffusion models

    Qiucheng Wu, Yujian Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text- to-image diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , p...

    1900

  55. [63]

    Freeman, Fr ´edo Durand, and Song Han

    Guangxuan Xiao, Tianwei Yin, William T. Freeman, Fr ´edo Durand, and Song Han. Fastcomposer: Tuning-free multi- subject image generation with localized attention. Interna- tional Journal of Computer Vision, 2024. 3

    2024

  56. [64]

    Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assess- ment

    Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assess- ment. arXiv preprint arXiv:2312.12148, 2023. 2

    2023 arXiv

  57. [65]

    End-to-end chinese landscape painting creation using generative adversarial networks

    Alice Xue. End-to-end chinese landscape painting creation using generative adversarial networks. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 3863–3871, 2021. 6

    2021

  58. [66]

    Tv-3dg: Mastering text-to-3d customized gener- ation with visual prompt

    Jiahui Yang, Donglin Di, Baorui Ma, Xun Yang, Yongjia Ma, Wenzhang Sun, Wei Chen, Jianxun Cui, Zhou Xue, Meng Wang, et al. Tv-3dg: Mastering text-to-3d customized gener- ation with visual prompt. arXiv preprint arXiv:2410.21299,

    arXiv

  59. [67]

    Zero-shot contrastive loss for text-guided diffusion image style transfer

    Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873–22882, 2023. 3

    2023

  60. [68]

    Deconfounded video moment retrieval with causal intervention

    Xun Yang, Fuli Feng, Wei Ji, Meng Wang, and Tat-Seng Chua. Deconfounded video moment retrieval with causal intervention. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, page 1–10, New York, NY , USA, 2021. A...

    2021

  61. [69]

    Video moment retrieval with cross-modal neural architecture search

    Xun Yang, Shanshan Wang, Jian Dong, Jianfeng Dong, Meng Wang, and Tat-Seng Chua. Video moment retrieval with cross-modal neural architecture search. IEEE Transac- tions on Image Processing, 31:1204–1216, 2022. 8

    2022

  62. [70]

    Robust video question answer- ing via contrastive cross-modality representation learning

    Xun Yang, Jianming Zeng, Dan Guo, Shanshan Wang, Jian- feng Dong, and Meng Wang. Robust video question answer- ing via contrastive cross-modality representation learning. Science China Information Sciences, 67(10):202104, 2024

    2024

  63. [71]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 8

    2024 arXiv

  64. [72]

    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,

    arXiv

  65. [73]

    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 ICCV, pages 3836–3847, 2023. 2, 3

    2023

  66. [74]

    Multi-lora composition for image generation

    Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. arXiv preprint arXiv:2402.16843, 2024. 2, 4

    2024 arXiv

  67. [75]

    Egotextvqa: Towards egocentric scene-text aware video question answering

    Sheng Zhou, Junbin Xiao, Qingyun Li, Yicong Li, Xun Yang, Dan Guo, Meng Wang, Tat-Seng Chua, and Angela Yao. Egotextvqa: Towards egocentric scene-text aware video question answering. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) , pages 3363– ...

    2025

  68. [76]

    Cached multi-lora composition for multi- concept image generation

    Xiandong Zou, Mingzhu Shen, Christos-Savvas Bouganis, and Yiren Zhao. Cached multi-lora composition for multi- concept image generation. In ICLR, 2025. 2, 4

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.