pith. sign in

arxiv: 2607.00766 · v2 · pith:RFQIHUG3new · submitted 2026-07-01 · 💻 cs.CV

Decoupled Guidance: Disentangling Subject and Context Pathways in Text-to-Image Personalization

Seongmin Kim , Kyucheol Shin , Heesun Jung , Jinseo Kim , Sungyong Baik This is my paper

Pith reviewed 2026-07-03 21:29 UTC · model grok-4.3

classification 💻 cs.CV
keywords text-to-image personalizationdecoupled guidanceconditioning entanglementfidelity-editability trade-offdiffusion modelsattention mapsplug-and-playspatial mixing
0
0 comments X

The pith

Routing subject identity and scene context through separate guidance streams reduces their competition in text-to-image personalization.

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

The paper claims that the common trade-off between faithful subject rendering and flexible scene editing arises because both signals share one conditioning pathway and compete for attention resources. It supports this by showing that swapping the specific subject token for a generic one shifts attention maps and improves context adherence. The proposed solution, Decoupled Guidance, splits the signals into two independent streams that are later fused only inside semantically appropriate regions via spatial mixing. The framework requires no changes to existing model weights and works at inference time on multiple personalization methods and backbones, including flow-matching Diffusion Transformers.

Core claim

Conditioning entanglement occurs when subject identity and scene context are encoded through the same pathway, forcing them to compete for attention-map resources and producing a fidelity-editability trade-off. Replacing the target subject token with a generic token shifts attention allocation and alters context adherence, confirming the causal link. Decoupled Guidance counters this by maintaining two independent guidance streams and applying a spatial mixing mechanism that lets each stream operate only inside its relevant region.

What carries the argument

Decoupled Guidance (DeGu) with two independent guidance streams and a spatial mixing mechanism that fuses them without cross-region interference.

If this is right

  • Personalization performance improves consistently across existing methods without backbone modifications.
  • Users gain inference-time control to adjust the fidelity-editability balance for each generation.
  • The approach extends to flow-matching Diffusion Transformers as well as standard diffusion backbones.
  • Attention resources are allocated more efficiently because each stream stays confined to its semantic region.

Where Pith is reading between the lines

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

  • The same separation principle could be tested in other conditional generation settings where multiple signals compete for model capacity.
  • Automatic selection of the mixing strength based on prompt content might further reduce manual tuning.
  • The method suggests that attention-map diagnostics can serve as a general tool for diagnosing signal interference in generative models.

Load-bearing premise

The fidelity-editability trade-off is caused primarily by the shared conditioning pathway rather than by other model or training factors.

What would settle it

If applying the two-stream guidance produces no measurable gain on standard personalization metrics such as subject fidelity and text alignment, or if the token-replacement test shows no corresponding attention shifts, the entanglement explanation would be falsified.

Figures

Figures reproduced from arXiv: 2607.00766 by Heesun Jung, Jinseo Kim, Kyucheol Shin, Seongmin Kim, Sungyong Baik.

Figure 1
Figure 1. Figure 1: Personalization results from DeGu. Given a few reference images, our method faithfully preserves fine-grained subject identity while composing each subject into novel scenes. All images are generated with the SD v3.5 backbone. Abstract Text-to-image personalization aims to generate a user￾provided subject in novel scenes described by text. How￾ever, most existing methods encode subject identity (fi￾delity)… view at source ↗
Figure 3
Figure 3. Figure 3: Attention collapse is a systematic consequence of uni￾fied prompts. Per-token cross-attention magnitudes for a com￾positional prompt. Left: In the pre-trained model, attention dis￾tributes evenly across tokens. Right: After personalization via TI, the learned subject token S ∗ absorbs most of the cross-attention budget, suppressing context tokens such as “purple,” “rug,” and “forest” to near zero—an effect… view at source ↗
Figure 2
Figure 2. Figure 2: Conditioning entanglement in existing personaliza￾tion and the proposed DeGu framework. (a) Previous methods route subject identity and scene context through a single condi￾tioning pathway. The attention maps illustrate attention collapse in embedding optimization (detailed analysis in Section 3). (b) DeGu routes subject identity and scene context through two in￾dependent guidance streams, preventing the a… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of our proposed framework. (a) We isolate the subject token from the context tokens and introduce Context-Agnostic Embeddings (CAE), which trains the token embedding of our learnable token [S ∗ ] using the token sequence [SOT] [S ∗ ] [EOT] · · · [EOT] to obtain the fidelity condition cf via text encoder E to focus on capturing the subject identity. (b) At inference, our Decoupled Guidance Mixer (D… view at source ↗
Figure 5
Figure 5. Figure 5: Attention collapse as evidence of conditioning en￾tanglement. Both plots report ACS against fidelity (DINO) and editability (CLIP-T). (a) As TI training progresses, ACS rises— fidelity improves but editability degrades accordingly. (b) Increas￾ing learnable tokens (nT I = 1–5) also raises ACS with the same trade-off, confirming a structural consequence of unified prompts rather than a matter of capacity. c… view at source ↗
Figure 6
Figure 6. Figure 6: Independent guidance control through γe and γf . Left (γf = 3.0 fixed): Increasing γe strengthens text adherence while pre￾serving subject identity. Right (γe = 7.5 fixed): Increasing γf enhances subject-specific features while preserving text prompt adherence. This validates independent pathway operation. tation zt conditioned on text condition c from text encoder E. Classifier-Free Guidance (CFG) [17] st… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on SDXL. Each row shows one subject: reference (leftmost), standalone DeGu, SDXL baselines (DTI, MS-Diffusion), and plug-and-play variants (DTI+DeGu, MS+DeGu). DTI follows the prompt but loses subject-specific details; MS-Diffusion retains coarse cues yet misses fine identity details for unseen subjects and fails to compose novel scenes. DeGu decouples identity and context via dedica… view at source ↗
Figure 8
Figure 8. Figure 8: Mask quality across denoising timesteps. Each group shows the decoded intermediate latent and the corresponding mask. At early timesteps (high noise), Grounded SAM 2 produces unreliable segmentations because its input lies far outside its training distribution, while TCAM consistently localizes the subject by operating on latent-space attention maps. Reference Images a bowl on top of green grass with sunfl… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative effect of each component. w/o DGM, TCAM: CAE placed in unified guidance recreates attention col￾lapse. w/o TCAM: DGM decouples guidance but cf spatially dom￾inates, leaking reference backgrounds. DeGu (Ours): TCAM con￾fines each stream to its relevant region, achieving both high fidelity and compositional control [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
read the original abstract

Text-to-image personalization aims to generate a user-provided subject in novel scenes described by text. However, most existing methods encode subject identity (fidelity) and context (editability) through the same conditioning pathway, forcing the two to compete for attention-map resources. We refer to this phenomenon as conditioning entanglement and show that it induces a fidelity-editability trade-off. We further provide causal evidence by replacing the target subject token with a generic subject token, which produces shifts in attention allocation and corresponding changes in context adherence. To this end, we propose Decoupled Guidance (DeGu), a plug-and-play framework that routes subject identity and scene context through two independent guidance streams. We further introduce a spatial mixing mechanism that dynamically fuses these streams, ensuring each operates within its semantically relevant region without interference. Furthermore, DeGu can be readily applied to existing personalization methods without modifying the underlying backbone models, consistently improving the overall personalization performance while enabling inference-time control over the fidelity-editability balance, across diverse methods and backbones, including flow-matching Diffusion Transformers (DiTs).

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 0 minor

Summary. The manuscript identifies conditioning entanglement—where subject identity and scene context share a single conditioning pathway and compete for attention resources—as the root cause of the fidelity-editability trade-off in text-to-image personalization. It supplies causal evidence via a token-replacement experiment (substituting the target subject token with a generic subject token) that produces observable shifts in attention allocation and context adherence. It then introduces Decoupled Guidance (DeGu), a plug-and-play framework that routes subject and context through independent guidance streams, augmented by a spatial mixing mechanism to fuse them without interference. The approach is claimed to improve personalization performance across existing methods and backbones (including flow-matching DiTs) while permitting inference-time control of the fidelity-editability balance, all without modifying the underlying models.

Significance. If the empirical improvements and the causal isolation of entanglement hold under scrutiny, the work would offer a practical, backbone-agnostic advance for a widely studied problem. The plug-and-play design and demonstrated compatibility with DiTs are concrete strengths that could see rapid adoption.

major comments (1)
  1. [Abstract] Abstract (causal evidence paragraph): the token-replacement test correlates attention reallocation with changes in context adherence but does not establish that shared-pathway entanglement is the dominant driver of the fidelity-editability trade-off, as opposed to other factors such as training-data biases, loss weighting, or non-attention conditioning routes. Because this isolation is load-bearing for the motivation and claimed remedy of DeGu, additional experiments that rule out alternative explanations are required.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (causal evidence paragraph): the token-replacement test correlates attention reallocation with changes in context adherence but does not establish that shared-pathway entanglement is the dominant driver of the fidelity-editability trade-off, as opposed to other factors such as training-data biases, loss weighting, or non-attention conditioning routes. Because this isolation is load-bearing for the motivation and claimed remedy of DeGu, additional experiments that rule out alternative explanations are required.

    Authors: The token-replacement experiment performs a controlled intervention: the model, training procedure, loss function, and all tokens except the subject token remain fixed. The only change is replacement of the target subject token with a generic one, which produces measurable shifts in attention maps and context adherence. Because data biases, loss weighting, and non-attention routes are identical in both conditions, the observed differences are attributable to competition within the shared conditioning pathway. We therefore maintain that the experiment supplies causal evidence for the role of entanglement. That said, we agree the abstract language could be more precise. In the revision we will (i) rephrase the abstract to state that the experiment demonstrates the contribution of pathway entanglement rather than claiming it is the sole dominant driver, and (ii) add a short discussion paragraph in Section 3 explicitly noting the controlled nature of the intervention and why alternative factors are held constant. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper's central claim rests on an empirical causal test (token replacement shifting attention and context adherence) plus an architectural proposal for independent guidance streams and spatial mixing. Neither the entanglement diagnosis nor the performance gains reduce by construction to a fitted parameter, self-defined quantity, or self-citation chain; the method introduces new routing and fusion mechanisms whose outputs are not equivalent to the inputs. No load-bearing uniqueness theorem, ansatz smuggling, or renaming of known results appears. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; full text would be required to populate the ledger.

pith-pipeline@v0.9.1-grok · 5732 in / 978 out tokens · 23591 ms · 2026-07-03T21:29:19.761447+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

57 extracted references

  1. [1]

    A neural space-time representation for text-to- image personalization.ACM TOG, 42, 2023

    Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to- image personalization.ACM TOG, 42, 2023. 2, 3, 4, 6, 7, 8

    2023

  2. [2]

    Break-a-scene: Extracting multiple concepts from a single image

    Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen- Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. InACM SIGGRAPH Asia,

  3. [3]

    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 ICCV, 2021. 6, 4

    2021

  4. [4]

    Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models.ACM TOG, 42, 2023

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models.ACM TOG, 42, 2023. 3

    2023

  5. [5]

    Disenbooth: Identity- preserving disentangled tuning for subject-driven text-to- image generation

    Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Identity- preserving disentangled tuning for subject-driven text-to- image generation. InICLR, 2024. 2, 3

    2024

  6. [6]

    Scaling rec- tified 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, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rec- tified flow transformers for high-resolution image synthesis. InICML, 2024. 4

    2024

  7. [7]

    Catastrophic forgetting in connectionist networks.Trends in Cognitive Sciences, 3, 1999

    Robert M French. Catastrophic forgetting in connectionist networks.Trends in Cognitive Sciences, 3, 1999. 3

    1999

  8. [8]

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

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. InICLR, 2023. 2, 3, 4, 5, 6, 7, 8

    2023

  9. [9]

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

    Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yun- peng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low- rank adaptation for multi-concept customization of diffusion models. InNeurIPS, 2023. 3

    2023

  10. [10]

    Hybridbooth: Hybrid prompt inversion for efficient subject-driven generation

    Shuang Guan, Yichen Ge, Yu-Wing Tai, Jian Yang, Weijian Li, and Min You. Hybridbooth: Hybrid prompt inversion for efficient subject-driven generation. InECCV, 2025. 3

    2025

  11. [11]

    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. InICCV, 2023. 2, 3

    2023

  12. [12]

    Improved noise schedule for diffusion training.ICCV, 2025

    Tiankai Hang, Shuyang Gu, Xin Geng, and Baining Guo. Improved noise schedule for diffusion training.ICCV, 2025. 2

    2025

  13. [13]

    Vico: Plug-and-play visual condition for personal- ized text-to-image generation.arXiv preprint, 2023

    Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personal- ized text-to-image generation.arXiv preprint, 2023. 3

    2023

  14. [14]

    A data perspec- tive on enhanced identity preservation for diffusion person- alization

    Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Kun Wan, Helge Rhodin, and Ratheesh Kalarot. A data perspec- tive on enhanced identity preservation for diffusion person- alization. InWACV, 2025. 3

    2025

  15. [15]

    Hertz, R

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y . Pritch, and D. Cohen-Or. Prompt-to-prompt image editing with cross attention control.ICLR, 2023. 3

    2023

  16. [16]

    Training products of experts by minimiz- ing contrastive divergence.Neural Computation, 14, 2002

    Geoffrey E Hinton. Training products of experts by minimiz- ing contrastive divergence.Neural Computation, 14, 2002. 5, 1

    2002

  17. [17]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 5, 1

    2021

  18. [18]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. InNeurIPS, 2020. 2, 1

    2020

  19. [19]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InICLR,

  20. [20]

    Cus- tomized generation reimagined: Fidelity and editability har- monized

    Jianyi Jin, Yuxuan Shen, Zhaofeng Fu, and Jian Yang. Cus- tomized generation reimagined: Fidelity and editability har- monized. InECCV, 2025. 2, 3, 4

    2025

  21. [21]

    Directional textual inversion for personalized text-to- image generation.ICLR, 2026

    Kunhee Kim, NaHyeon Park, Kibeom Hong, and Hyunjung Shim. Directional textual inversion for personalized text-to- image generation.ICLR, 2026. 6, 7, 4, 9

    2026

  22. [22]

    Overcoming catastrophic forgetting in neu- ral networks.Proceedings of the National Academy of Sci- ences, 114, 2017

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska- Barwinska, et al. Overcoming catastrophic forgetting in neu- ral networks.Proceedings of the National Academy of Sci- ences, 114, 2017. 3

    2017

  23. [23]

    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. InCVPR, 2023. 2, 3

    2023

  24. [24]

    Blip-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing

    Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing. InNeurIPS, 2023. 2, 3

    2023

  25. [25]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InICML,

  26. [26]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2017. 3

    2017

  27. [27]

    Common diffusion noise schedules and sample steps are flawed.WACV, 2024

    Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed.WACV, 2024. 2

    2024

  28. [28]

    Flow matching for generative mod- eling.ICLR, 2023

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.ICLR, 2023. 2

    2023

  29. [29]

    Compositional visual generation with composable diffusion models

    Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. InECCV, 2022. 3, 5, 1, 2

    2022

  30. [30]

    Grounding dino: Marrying dino with grounded pre- training for open-set object detection.arXiv preprint, 2023

    Shilong Liu, Zewen Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre- training for open-set object detection.arXiv preprint, 2023. 6, 8, 5

    2023

  31. [31]

    Cones: Concept neurons in diffusion models for customized generation.ICML, 2023

    Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation.ICML, 2023. 3 14

    2023

  32. [32]

    Customizable image synthesis with multiple subjects

    Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Customizable image synthesis with multiple subjects. InNeurIPS, 2023. 3

    2023

  33. [33]

    Nichol, P

    A. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. McGrew, I. Sutskever, and M. Chen. Glide: Towards pho- torealistic image generation and editing with text-guided dif- fusion models.ICML, 2022. 2, 3

    2022

  34. [34]

    Cross initialization for face personaliza- tion of text-to-image models

    Lianyu Pang, Jian Yin, Haoran Xie, Qiping Wang, Qing Li, and Xudong Mao. Cross initialization for face personaliza- tion of text-to-image models. InCVPR, 2024. 2, 3

    2024

  35. [35]

    Seediff: Off-the-shelf seeded mask generation from diffusion models

    Joon Hyun Park, Kumju Jo, and Sungyong Baik. Seediff: Off-the-shelf seeded mask generation from diffusion models. InAAAI, 2025. 3

    2025

  36. [36]

    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. InICLR, 2024. 2, 3, 4

    2024

  37. [37]

    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, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021. 3, 6, 4

    2021

  38. [38]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.JMLR, 2020. 3

    2020

  39. [39]

    Hierarchical text-conditional image gener- ation with CLIP latents.arXiv preprint, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with CLIP latents.arXiv preprint, 2022. 2, 3

    2022

  40. [40]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt- ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao- Yuan Wu, Ross Girshick, Piotr Doll´ar, and Christoph Feicht- enhofer. Sam 2: Segment anything in images and videos. arXiv preprint, 2024. 6, 8, 5

    2024

  41. [41]

    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. InCVPR, 2022. 2, 4

    2022

  42. [42]

    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. InCVPR, 2023. 2, 3, 6, 7, 4, 8

    2023

  43. [43]

    Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. InCVPR, 2024. 3

    2024

  44. [44]

    Photorealistic text-to-image diffusion models with deep lan- guage understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Sali- mans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep lan- guage understanding. InNeurIPS, 2022. 2, 3

    2022

  45. [45]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InICML, 2015. 2

    2015

  46. [46]

    Denois- ing diffusion implicit models.ICLR, 2021

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

    2021

  47. [47]

    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. InICLR, 2021. 2

    2021

  48. [48]

    Key-locked rank one editing for text-to-image personaliza- tion

    Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personaliza- tion. InACM SIGGRAPH, 2023. 3

    2023

  49. [49]

    Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion

    Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. InCVPR, 2024. 3

    2024

  50. [50]

    Concept decomposition for visual exploration and inspiration.ACM Transactions on Graphics, 42, 2023

    Yael Vinker, Andrey V oynov, Daniel Cohen-Or, and Ariel Shamir. Concept decomposition for visual exploration and inspiration.ACM Transactions on Graphics, 42, 2023. 3

    2023

  51. [51]

    P+: Extended textual conditioning in text-to- image generation.arXiv preprint, 2023

    Andrey V oynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to- image generation.arXiv preprint, 2023. 2, 3, 4, 6, 7, 8

    2023

  52. [52]

    Ms-diffusion: Multi-subject zero-shot image per- sonalization with layout guidance

    Xierui Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image per- sonalization with layout guidance. InICLR, 2025. 2, 3, 6, 7, 4, 9, 13

    2025

  53. [53]

    Elite: Encoding visual con- cepts into textual embeddings for customized text-to-image generation

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual con- cepts into textual embeddings for customized text-to-image generation. InICCV, 2023. 2, 3

    2023

  54. [54]

    Core: Context- regularized text embedding learning for text-to-image per- sonalization

    Feize Wu, Yun Pang, Junyi Zhang, Lianyu Pang, Jian Yin, Baoquan Zhao, Qing Li, and Xudong Mao. Core: Context- regularized text embedding learning for text-to-image per- sonalization. InAAAI, 2025. 6, 7, 4, 8

    2025

  55. [55]

    Cusconcept: Cus- tomized visual concept decomposition with diffusion mod- els

    Zhi Xu, Shaozhe Hao, and Kai Han. Cusconcept: Cus- tomized visual concept decomposition with diffusion mod- els. InWACV, 2025. 3

    2025

  56. [56]

    Prospect: Prompt spectrum for attribute- aware personalization of diffusion models.ACM Transac- tions on Graphics, 42, 2023

    Yuxing Zhang, Wenbo Dong, Fanyi Tang, Nian Huang, Haisu Huang, Changjie Ma, Tong-Yee Lee, Oliver Deussen, and Changhe Xu. Prospect: Prompt spectrum for attribute- aware personalization of diffusion models.ACM Transac- tions on Graphics, 42, 2023. 3

    2023

  57. [57]

    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. InCVPR, 2024. 2, 3 15

    2024