pith. sign in

arxiv: 2606.02129 · v1 · pith:RXFVVYBRnew · submitted 2026-06-01 · 💻 cs.CV

Equilibrated Diffusion: Frequency-aware Textual Embedding for Equilibrated Image Customization

Liyuan Ma , Xueji Fang , Guo-Jun Qi This is my paper

Pith reviewed 2026-06-28 15:09 UTC · model grok-4.3

classification 💻 cs.CV
keywords image customizationdiffusion modelsfrequency decompositiontextual embeddingsubject fidelitystyle disentanglementmask-guided generation
0
0 comments X

The pith

Equilibrated Diffusion separates low-frequency subject content from high-frequency styles in textual embeddings to reduce entanglement during image customization.

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

The paper establishes that conventional fine-tuning packs subject, style, and background into one embedding, which creates unwanted disturbances when generating new images from text. Equilibrated Diffusion instead decomposes the reference concept in frequency space, treating low frequencies as the subject identity and high frequencies as style, then optimizes separate embeddings for each. Independent tuning lets the model capture style without altering core subject features and improves generalization to unseen stylistic prompts. A mask-guided diffusion step further limits background drift while residual reference attention preserves structural consistency. If correct, this frequency separation yields higher subject fidelity and tighter text alignment than unified-embedding baselines.

Core claim

Decomposing concepts into frequency-specific embeddings, optimizing them independently, and merging the results lets the denoiser treat style as detachable from subject identity while mask guidance and residual attention maintain spatial fidelity and text adherence.

What carries the argument

Frequency-aware textual embedding that decomposes reference images into low-frequency subject and high-frequency style components for separate optimization before merging.

If this is right

  • Independent frequency embeddings allow the model to apply novel styles to the learned subject without retraining.
  • Merging the optimized embeddings keeps the original spatial customization capability of the base diffusion model intact.
  • Mask-guided diffusion restricts background alterations and improves prompt adherence during generation.
  • Residual reference attention inserted in spatial layers preserves subject structure across varied prompts.

Where Pith is reading between the lines

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

  • The same frequency split could be tested on video or 3D generation tasks where content-style separation is also desirable.
  • Different frequency cutoffs or wavelet bases might be compared to find the split that maximizes disentanglement for a given dataset.
  • The method might reduce the need for heavy fine-tuning data if the frequency prior already supplies useful separation.

Load-bearing premise

Low image frequencies inherently encode subject content while high frequencies encode style.

What would settle it

A controlled test in which low- and high-frequency embeddings are swapped at inference time yet produce no measurable drop in subject fidelity or rise in style leakage.

Figures

Figures reproduced from arXiv: 2606.02129 by Guo-jun Qi, Liyuan Ma, Xueji Fang.

Figure 1
Figure 1. Figure 1: Image customization results of our method, which enables the recreation of a specified conceptual subject under [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Method overview. In the training stage, the Frequency-aware Decoupled TextualEmbedding (FDTE) decomposes the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of Mask Guided Diffusion Process. To [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Framework of Residual Reference Attention (RRA). [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) Reference Attention is capable of attending to [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison with prevalent methods. Please zoom in for a better view. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization results of ablation study. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Numerical analysis of FDTE. 𝑝𝑙 , 𝑝ℎ, and 𝑝𝑜 represent the probability of choosing low-frequency, high-frequency and original images, respectively. image’s background in the result. Although MDL introduces con￾straints on background regions in loss objectives, it still struggles to mitigate the influence of background attributes. Conversely, MGDP exhibits superior capability in decoupling backgrounds from t… view at source ↗
Figure 7
Figure 7. Figure 7: The best and the second best results are bold [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

Image customization learns target subjects from reference concept images and generates conditioned images per text prompts, mainly modifying styles or backgrounds. Prevailing methods adopt fine-tuning to pack diverse concept attributes into a unified latent embedding, yet entangled attributes hinder elimination of irrelevant disturbances from style and background. To address this issue, we propose Equilibrated Diffusion, a frequency-driven approach that disentangles tangled concept features for balanced customization and consistent text-visual matching. Unlike conventional methods learning full concepts with shared embeddings and unified tuning, our work utilizes the inherent link between image frequency components and semantics: low frequencies represent subject content and high frequencies correspond to styles. We decompose concepts in frequency space and optimize each embedding independently. This separate optimization enables the denoiser to capture style detached from subject identity and generalize better to unseen stylistic prompts. Merging multi-frequency embeddings preserves the model's original spatial customization ability. We further deploy mask-guided diffusion to restrict irrelevant background changes and boost text alignment. Residual Reference Attention (RRA) is inserted into spatial attention to retain subject structure and identity consistency. Experiments prove Equilibrated Diffusion exceeds mainstream baselines on subject fidelity and text adherence, verifying our method's superiority.

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

2 major / 0 minor

Summary. The paper proposes Equilibrated Diffusion, a frequency-driven method for image customization that decomposes reference concepts into separate low-frequency (subject content) and high-frequency (style) textual embeddings, optimizes them independently to reduce entanglement, and augments this with mask-guided diffusion and Residual Reference Attention (RRA) to improve text alignment and identity preservation. It claims this yields superior subject fidelity and text adherence over mainstream baselines.

Significance. If the frequency-semantics correspondence holds and produces verifiable disentanglement, the approach could meaningfully advance subject-driven customization by mitigating the attribute entanglement common in unified fine-tuning methods, enabling better generalization to novel stylistic prompts while preserving spatial control.

major comments (2)
  1. [Abstract] Abstract: The method's core strategy of independent optimization rests on the unvalidated assertion of an 'inherent link' where low frequencies encode subject identity and high frequencies encode styles. No derivation from the diffusion forward process, no frequency-domain analysis of the UNet, and no ablation isolating the effect of this split on disentanglement are provided, making the architectural novelty and superiority claims dependent on an untested premise.
  2. [Abstract] Abstract: The claim that 'Experiments prove Equilibrated Diffusion exceeds mainstream baselines on subject fidelity and text adherence' is unsupported by any quantitative metrics, error bars, dataset specifications, baseline details, or ablation results in the provided text, preventing verification of the central empirical assertion.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and detailed comments. We address each major point below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The method's core strategy of independent optimization rests on the unvalidated assertion of an 'inherent link' where low frequencies encode subject identity and high frequencies encode styles. No derivation from the diffusion forward process, no frequency-domain analysis of the UNet, and no ablation isolating the effect of this split on disentanglement are provided, making the architectural novelty and superiority claims dependent on an untested premise.

    Authors: We acknowledge that the abstract presents the frequency-semantics correspondence as an inherent link without supporting derivation or analysis. While this correspondence draws from established image processing observations (low frequencies capture global structure and semantics, high frequencies capture textures and style), we agree that a direct connection to the diffusion forward process and UNet behavior requires explicit validation. In the revised manuscript we will add a concise frequency-domain analysis of the diffusion process together with an ablation that isolates the contribution of the low/high-frequency split to disentanglement. These additions will be placed in the method or experiments section and referenced from the abstract. revision: yes

  2. Referee: [Abstract] Abstract: The claim that 'Experiments prove Equilibrated Diffusion exceeds mainstream baselines on subject fidelity and text adherence' is unsupported by any quantitative metrics, error bars, dataset specifications, baseline details, or ablation results in the provided text, preventing verification of the central empirical assertion.

    Authors: The abstract is written as a concise summary; the full manuscript contains the quantitative evaluations, including subject fidelity and text-alignment metrics, baseline comparisons, dataset details, and ablations. Nevertheless, we agree that the abstract claim would be more verifiable if it referenced key results. We will revise the abstract to include brief quantitative highlights (e.g., relative improvements on standard metrics) while still keeping it within length limits, and we will ensure the abstract explicitly points to the experiments section for full details, error bars, and dataset specifications. revision: yes

Circularity Check

0 steps flagged

No significant circularity: method applies stated frequency-semantics assumption without self-referential reduction or fitted predictions

full rationale

The abstract presents the core premise as an 'inherent link' between frequency components and semantics (low frequencies = subject content, high frequencies = styles) and describes independent optimization of embeddings as a direct consequence. No equations, derivations, or self-citations are shown that would make any claimed prediction equivalent to its inputs by construction. The approach is self-contained once the frequency-semantics correspondence is granted as an external modeling choice; experiments compare against baselines rather than relying on internal consistency alone. This is the normal case of an assumption-driven method without circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review is abstract-only; the central claim rests on the unverified frequency-to-semantics mapping and the assumption that separate optimization plus mask and attention additions will generalize without introducing new artifacts.

axioms (1)
  • domain assumption Low frequencies represent subject content and high frequencies correspond to styles
    Invoked in the abstract to justify frequency decomposition and independent embedding optimization.

pith-pipeline@v0.9.1-grok · 5737 in / 1229 out tokens · 19516 ms · 2026-06-28T15:09:31.018487+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

51 extracted references · 21 canonical work pages · 9 internal anchors

  1. [1]

    Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. 2023. A Neural Space-Time Representation for Text-to-Image Personalization.ACM Transactions on Graphics (TOG)42, 6 (2023), 1–10

    2023

  2. [2]

    Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischin- ski. 2023. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia 2023 Conference Papers. 1–12

    2023

  3. [3]

    Yufei Cai, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Hu Han, and Wangmeng Zuo. 2024. Decoupled textual embeddings for customized image generation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 909–917

    2024

  4. [4]

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. 2023. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision. 22560–22570

    2023

  5. [5]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision. 9650–9660

    2021

  6. [6]

    Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. 2023. Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. InThe Twelfth International Conference on Learning Representations

    2023

  7. [7]

    Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. 2024. Subject-driven text-to-image generation via apprenticeship learning.Advances in Neural Information Processing Systems 36 (2024)

    2024

  8. [8]

    Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. 2023. Custom-edit: Text-guided image editing with customized diffusion models.arXiv preprint arXiv:2305.15779(2023)

    work page arXiv 2023

  9. [9]

    Siying Cui, Jiankang Deng, Jia Guo, Xiang An, Yongle Zhao, Xinyu Wei, and Ziyong Feng. 2024. IDAdapter: Learning Mixed Features for Tuning-Free Person- alization of Text-to-Image Models.arXiv preprint arXiv:2403.13535(2024)

    work page arXiv 2024

  10. [10]

    Ziyi Dong, Pengxu Wei, and Liang Lin. 2022. DreamArtist: Towards Controllable One-Shot Text-to-Image Generation via Positive-Negative Prompt-Tuning.arXiv preprint arXiv:2211.11337(2022)

    work page arXiv 2022

  11. [11]

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. 2023. An Image is Worth One Word: Per- sonalizing Text-to-Image Generation using Textual Inversion. InThe Eleventh International Conference on Learning Representations. https://openreview.net/ forum?id=NAQvF08TcyG

    2023

  12. [12]

    Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. 2023. Encoder-based domain tuning for fast personalization of text- to-image models.ACM Transactions on Graphics (TOG)42, 4 (2023), 1–13

    2023

  13. [13]

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

    2023

  14. [14]

    Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. 2023. ViCo: Plug- and-play Visual Condition for Personalized Text-to-image Generation. (2023)

    2023

  15. [15]

    Feihong He, Gang Li, Mengyuan Zhang, Leilei Yan, Lingyu Si, and Fanzhang Li

  16. [16]

    arXiv preprint arXiv:2401.15636(2024)

    Freestyle: Free lunch for text-guided style transfer using diffusion models. arXiv preprint arXiv:2401.15636(2024)

    work page arXiv 2024

  17. [17]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  18. [18]

    Miao Hua, Jiawei Liu, Fei Ding, Wei Liu, Jie Wu, and Qian He. 2023. Dream- Tuner: Single Image is Enough for Subject-Driven Generation.arXiv preprint arXiv:2312.13691(2023)

    work page arXiv 2023

  19. [19]

    Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. 2023. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models.arXiv preprint arXiv:2304.02642(2023)

    work page arXiv 2023

  20. [20]

    Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. 2021. Focal frequency loss for image reconstruction and synthesis. InProceedings of the IEEE/CVF international conference on computer vision. 13919–13929

    2021

  21. [21]

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu

  22. [22]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Multi-concept customization of text-to-image diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1931–1941

    1931

  23. [23]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  24. [24]

    Liyuan Ma, Tingwei Gao, Haibin Shen, and Kejie Huang. 2023. Freqhpt: Frequency-aware attention and flow fusion for human pose transfer. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 3490–3495

    2023

  25. [25]

    Yiyang Ma, Huan Yang, Wenjing Wang, Jianlong Fu, and Jiaying Liu. 2023. Uni- fied multi-modal latent diffusion for joint subject and text conditional image generation.arXiv preprint arXiv:2303.09319(2023)

    work page arXiv 2023

  26. [26]

    Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Se- unggyu Chang. 2024. DreamMatcher: Appearance Matching Self-Attention for Semantically-Consistent Text-to-Image Personalization.arXiv preprint arXiv:2402.09812(2024)

    work page arXiv 2024

  27. [27]

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2021. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  28. [28]

    R OpenAI. 2023. Gpt-4 technical report. arxiv 2303.08774.View in Article2, 5 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. 2022. Learning spa- tiotemporal frequency-transformer for compressed video super-resolution. In European Conference on Computer Vision. Springer, 257–273

    2022

  30. [30]

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

    2021

  31. [31]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen

  32. [32]

    Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.061251, 2 (2022), 3

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  33. [33]

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. 2024. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10684–10695

    2022

  35. [35]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22500–22510

    2023

  36. [36]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems35 (2022), 36479–36494

    2022

  37. [37]

    Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. 2023. Instantbooth: Per- sonalized text-to-image generation without test-time finetuning.arXiv preprint arXiv:2304.03411(2023)

    work page arXiv 2023

  38. [38]

    Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. 2022. Inception transformer.Advances in Neural Information Processing Systems35 (2022), 23495–23509

    2022

  39. [39]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  40. [40]

    Yeji Song, Jimyeong Kim, Wonhark Park, Wonsik Shin, Wonjong Rhee, and Nojun Kwak. 2024. Harmonizing Visual and Textual Embeddings for Zero-Shot Text-to-Image Customization.arXiv preprint arXiv:2403.14155(2024)

    work page arXiv 2024

  41. [41]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  42. [42]

    Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. 2023. Key-locked rank one editing for text-to-image personalization. InACM SIGGRAPH 2023 Conference Proceedings. 1–11

    2023

  43. [43]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  44. [44]

    Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. 2023. Sketch-guided text- to-image diffusion models. InACM SIGGRAPH 2023 Conference Proceedings. 1–11

    2023

  45. [45]

    arXiv preprint arXiv:2303.09522 , year=

    Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. 2023.𝑃+: Extended Textual Conditioning in Text-to-Image Generation.arXiv preprint arXiv:2303.09522(2023)

    work page arXiv 2023

  46. [46]

    Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. 2024. In- stantid: Zero-shot identity-preserving generation in seconds.arXiv preprint arXiv:2401.07519(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  47. [47]

    Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. 2022. Semantic image synthesis via diffusion models. arXiv preprint arXiv:2207.00050(2022)

    work page arXiv 2022

  48. [48]

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo

  49. [49]

    InProceedings of the IEEE/CVF International Conference on Computer Vision

    Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 15943–15953

  50. [50]

    Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. 2023. Smart- brush: Text and shape guided object inpainting with diffusion model. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22428–22437

    2023

  51. [51]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023