Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Per-Query Visual Concept Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A single 4-second gradient update improves personalized text-to-image generation across six methods and three backbones, by aligning identity and prompt features.

desk verdict A useful, modest per-query add-on for personalization; the gains look real for DB and TI, but the DIFT-correspondence mechanism is under-tested and some baselines are re-implemented. read the letter →

arxiv 2508.09045 v1 pith:QKAFACV4 submitted 2025-08-12 cs.CV

Ori Malca , Dvir Samuel , Gal Chechik This is my paper
classification cs.CV
keywords text-to-imagepersonalizationper-queryadaptationdiffusionmodelsself-attentionfeaturescross-attentionmapsDIFTcorrespondencepromptadherenceidentitypreservation
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

This paper tries to establish that any existing text-to-image personalization checkpoint can be strengthened by a per-query fine-tuning step: for a fixed prompt and noise seed, one gradient update with self- and cross-attention losses raises both identity preservation and prompt adherence. The update costs about four seconds on an NVIDIA H100 and works on top of six personalization methods, including DreamBooth, LoRA, and Textual Inversion, as well as UNet- and transformer-based backbones. A sympathetic reader would care because the method turns personalization into a cheap, disposable correction at generation time rather than a permanent fine-tune, and it reports consistent gains on standard image- and text-alignment metrics.

What carries the argument

DIFT-matched attention features: self-attention features ('PDM features') extracted from a single clean denoising step localize identity, cross-attention maps localize the concept token [V], and DIFT supplies point correspondences between reference and generated latents. The losses pull identity and prompt features together while a standard noise-prediction loss anchors the update. The fixed prompt and seed make the update specific to a single query, and the use of t=1000 makes the feature extraction cheap.

What would settle it

Take a prompt with heavy occlusion or reflection, compute DIFT matches between reference and generated latents, and compare the matched pairs against human-annotated correspondences; if a large fraction are incorrect, or if applying the single update to those queries lowers CLIP-I/DINO-I relative to no update, then the correspondence premise fails.

Watch

Extended reading notes

Core claim

The central claim is that visual identity is carried by self-attention features (the PDM features originally designed for retrieval and segmentation) while prompt compliance is carried by cross-attention maps, so aligning these two feature types between a reference image and a single generated draft, after fixing the prompt and seed, is enough to improve generation. The authors compute one noised denoising step at t=1000 for both latents, match points with DIFT, and minimize L = λSA LSA + λCA LCA + λLDM LLDM. They report that a single gradient update of the already-personalized model improves CLIP-I, DINO-I, and CLIP-T across DreamBooth, LoRA, Textual Inversion, DBlend, AttnDB, and CLD, and

Load-bearing premise

The load-bearing premise is that the DIFT point correspondence computed once between the reference latent and the generated draft is semantically correct and stays valid after the single parameter update; if the matches are wrong, the attention losses push the model toward noise rather than identity.

Editorial extensions

If this is right

  • Adding the step to DreamBooth, LoRA, Textual Inversion, DBlend, AttnDB, or CLD improves subject alignment and prompt adherence on standard benchmarks.
  • The single-update regime allows the method to be used at inference time as a roughly 4-second per-query correction, rather than a separate training run.
  • LSA alone drives identity gains, LCA alone drives prompt adherence, and combining them gives both, while applying the losses during base training hurts performance.
  • Because it does not change the architecture or the base method, the same add-on applies to UNet-based models (SD, SDXL) and transformer-based models (FLUX).
  • The authors report that their per-query method outperforms previous per-query personalization methods AlignIT and PALP on complex prompts.

Reading between the lines

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

  • If the gains hold broadly, the same DIFT-anchored attention losses could serve as a general post-hoc regularizer for other diffusion-based generation tasks, such as subject-driven editing; the paper does not test this.
  • The per-query nature suggests a possible search or caching strategy: since the update is tied to a specific seed, users could jointly optimize prompts and seeds with the one-step correction.
  • A concrete test: perturb the reference pose or add occlusion to break DIFT correspondences; the identity gain should shrink or invert, which would confirm that the correspondence assumption is the active ingredient.
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. The paper proposes a per-query, seed-specific add-on for text-to-image personalization. Given an already-personalized checkpoint, a reference image, a target prompt, and a noise seed, the method generates a draft image, computes self- and cross-attention features from both the draft and reference latents, establishes point correspondences via DIFT, and performs a single gradient update with three losses: a self-attention identity loss (LSA), a cross-attention prompt-alignment loss (LCA), and the standard diffusion reconstruction loss (LLDM). The authors claim consistent improvements in subject alignment (CLIP-I, DINO-I) and prompt adherence (CLIP-T) on top of six personalization methods across UNet- and DiT-based backbones, including comparisons with two prior per-query methods, AlignIT and PALP, and report a runtime of about 4 seconds per query on an H100.

Significance. If the central claim holds, the method is a genuinely useful and cheap add-on: a single gradient update that improves both identity preservation and prompt alignment across a wide range of existing personalization pipelines and backbones. The paper's strengths are its broad empirical coverage (six personalization methods, SD/SDXL/FLUX backbones), the simple and fast procedure, and the explicit per-loss ablation in Table 2, which shows complementary roles for the self- and cross-attention terms. The main weaknesses are statistical: the headline 'significant' gains are not backed by paired significance tests or confidence intervals, the quantitative comparison against AlignIT/PALP appears only as a figure without numbers, and the load-bearing DIFT correspondence mechanism is never controlled. These issues are fixable and do not, at this stage, invalidate the practical claim, but they must be resolved before the mechanism and the SoTA claim can be accepted.

major comments (3)
  1. [§4.2.2, Tables 1 and 2] The abstract and §1 describe the gains as 'significant', but the only per-method numeric comparisons (Tables 1 and 2) report standard deviations that overlap with the reported deltas (e.g., DB DINO-I: 0.659±0.101 → 0.710±0.103; CLIP-I: 0.805±0.048 → 0.845±0.046). No paired tests or bootstrap confidence intervals are given. Since prompts and seeds are matched across w/ and w/o conditions, paired permutation tests or per-subject CIs should be reported. The comparison against AlignIT/PALP in Fig. 5 has no numeric table at all, so the 'significant improvements even over previous per-query methods' claim is currently unsupported.
  2. [§3.2, Eqs. (1)–(2); Table 2] The claimed novelty is using PDM/DIFT semantic correspondences to preserve identity, but no experiment varies the correspondence mechanism. LSA and LCA always use DIFT matches; Table 2 ablates the loss terms, not the matching. A random-match control (same losses with random point pairs) is needed to determine whether the gains depend on semantic matching or whether the losses act as generic regularizers. This is load-bearing because the paper's identity-preservation mechanism—and its generalization to complex prompts where DIFT may produce wrong matches—is otherwise unsupported.
  3. [§4.1.1, Eq. (4)] Reproducibility: Eq. (4) includes λLDM, but the implementation details only fix λSA=λCA=1; λLDM is never specified. In addition, Step A/B say a 'single denoising step' is performed with t=1000 and that it is 'solely for feature extraction' and 'does not change the latent'; this is underspecified and appears contradictory, since t=1000 corresponds to near-pure noise. Please specify the exact forward/backward calls, the reference prompt used for the cross-attention term, and the λLDM value or range. These details are necessary to replicate the single-update procedure.
minor comments (5)
  1. [§4.1.2] Typo: 'Textual Invresion' should be 'Textual Inversion'.
  2. [§4.2.2] Typo: 'align th' should be 'align the'.
  3. [§3.2, Step A] The phrase 'perform a single denoising step of z_ref^0, with a denoising parameter α determined with a parameter t=1000' is unclear. Does the step add noise or predict clean latents? How is α computed from t? Clarify the exact operation.
  4. [Tables 1 and 2] Naming inconsistency: the text uses LSA for the self-attention loss, while Table 2 uses λPDM and LPDM. Unify the notation.
  5. [§4.1.1] DBlend, AlignIT, and PALP are re-implemented without official code. Please release the re-implementation code or provide detailed configuration files so that the comparisons are verifiable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the add-on losses are external-supervision objectives and evaluation uses independent CLIP/DINO metrics.

full rationale

The paper proposes a per-query add-on that updates a personalized checkpoint with LSA, LCA, and LLDM, then evaluates with CLIP-I, DINO-I, and CLIP-T. These metrics are external to the proposed losses and are not the same feature spaces used in the update, so the improvement is not guaranteed by construction. LSA and LCA supervise the model using the reference image at inference time, but the evaluation is a comparison against baselines without the add-on, and the target metrics remain independent. The only same-group citation is PDM [28], used to extract identity-sensitive self-attention features. That citation is to a previously published, externally evaluated method for segmentation and retrieval; borrowing its features is legitimate and does not make the present result a restatement of the cited work. No equation in the paper reduces to another by construction, no fitted parameter is renamed as a prediction, and no claim is justified solely by a self-citation. The untested assumption that DIFT correspondences remain valid after one gradient step is a potential correctness risk, not a circularity. Therefore the derivation chain is self-contained.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new entities are introduced. The main free parameters are three scalar loss weights and the feature-extraction timestep t=1000. The method rests on domain assumptions about DIFT correspondence quality, PDM feature semantics, cross-attention map interpretability, and the safety of a single gradient update. These are reasonable for the diffusion personalization setting but are not derived from first principles.

free parameters (4)
  • lambda_SA = 1
    Tradeoff weight for the self-attention identity loss, set by hand in Section 4.1.1.
  • lambda_CA = 1
    Tradeoff weight for the cross-attention prompt loss, set by hand in Section 4.1.1.
  • lambda_LDM = 1 (implied)
    Weight for the reconstruction loss; not explicitly listed in implementation but treated as a tradeoff coefficient in Eq. 4.
  • feature_extraction_timestep_t = 1000
    Selected via ablation in Appendix B, Figure 8 on the DreamBooth benchmark and used for all reported main results.
assumptions (5)
  • domain assumption DIFT establishes correct point correspondences between zref0 and zgen0.
    Section 3.2 Step C; all identity losses are computed over DIFT matched pairs, so incorrect matches would corrupt the training signal.
  • domain assumption PDM self-attention features encode visual identity of the concept.
    Section 3.2 Step A; adopted from prior work [28] by the same group and not independently validated in this paper.
  • domain assumption Cross-attention maps for the [V] token encode prompt-sensitive alignment.
    Section 3.2 Steps A and B; LCA penalizes map differences at matched points, assuming these maps reflect concept placement and prompt control.
  • domain assumption A single gradient update with the base training learning rate improves generation without overfitting or corrupting the denoising process.
    Section 3.2 Step D and Section 4.1.1; stated empirically without a stability proof or analysis of failure cases.
  • domain assumption At inference the user has the reference image, the target prompt with [V], and the noise seed.
    Section 3.2 preamble; the per-query setting assumes these inputs are available for every generation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Per-Query Visual Concept Learning." pith.science (2026). https://pith.science/paper/QKAFACV4

@misc{pith2026250809045,
  author       = {Pith},
  title        = {Pith review of: Per-Query Visual Concept Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QKAFACV4}},
  note         = {Machine review of arXiv:2508.09045}
}
read the original abstract

Visual concept learning, also known as Text-to-image personalization, is the process of teaching new concepts to a pretrained model. This has numerous applications from product placement to entertainment and personalized design. Here we show that many existing methods can be substantially augmented by adding a personalization step that is (1) specific to the prompt and noise seed, and (2) using two loss terms based on the self- and cross- attention, capturing the identity of the personalized concept. Specifically, we leverage PDM features -- previously designed to capture identity -- and show how they can be used to improve personalized semantic similarity. We evaluate the benefit that our method gains on top of six different personalization methods, and several base text-to-image models (both UNet- and DiT-based). We find significant improvements even over previous per-query personalization methods.

Figures

Figures reproduced from arXiv: 2508.09045 by the authors.

Figure 1
Figure 1. Given a pretrained text-to-image personalization checkpoint, our method enhances per-query generation in terms of image alignment and text alignment, with just a single gradient update (∼4 seconds on an NVIDIA H100 GPU). It is compatible with a wide range of personalization techniques (e.g., DreamBooth [26], LoRA [16], Textual Inversion [8], DBlend [22]) and supports various diffusion backbones, including UNet-based… view at source ↗
Figure 2
Figure 2. Method Overview [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison. We present images generated by various personalization methods without our method and with it, including DB [26], TI [8], DBlend [22], CLD [10], AttnDB [19], and LoRA [16], across 4 different backbones (FLUX [13], SDXL [20], SD2.1, SD1.5 [25]). The left column shows one example of the target concept, and then every pair of columns shows the generated images with and without our method. Each p… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Effect of adding our method to baseline personalization models. shown are image alignment (top) and text alignment (bot￾tom) of several personalization approaches with (w/) and without (w/o) the integration of our method. Tuning-time: We describe two sets of experiment…
Figure 5
Figure 5. Figure 5: Quantitative Comparison with previous per-query meth￾ods. Image alignment (top) and text alignment (bottom) of vari￾ous personalization approaches (including, DB ▲, LoRA ■, and TI ), with (w/) and without (w/o) the integration of different per￾query methods (including,…
Figure 6
Figure 6. Figure 6: Quality-vs-time tradeoff [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison. We present images generated by 3 personalization methods (including DB [26], TI [8], LoRA [16]) using 3 per-query methods (including AlignIT [1], PALP [3], Ours), across 3 different backbones (FLUX [13], SDXL [20], SD2.1 [25]). Our method demons…
Figure 8
Figure 8. Figure 8: Noise weight Ablation Study. Sensitivity of CLIP-I and CLIP-T metrics to the magnitude of the noise, in terms of the t parameter used when calculating features. Section 3 describes the feature extraction process. In the main paper, we use a fixed value of t = 1000. Her…
Figure 10
Figure 10. Figure 10: Additional qualitative comparison. We present images generated by various personalization methods without our method and with it, including DB [26], TI [8], DBlend [22], CLD [10], AttnDB [19], and LoRA [16], across 4 different backbones (FLUX [13], SDXL [20], SD2.1, S…
Figure 11
Figure 11. Figure 11: Additional qualitative comparison. We present images generated by 3 personalization methods (including DB [26], TI [8], LoRA [16]) using 3 per-query methods (including AlignIT [1], PALP [3], Ours), across 3 different backbones (FLUX [13], SDXL [20], SD2.1 [25]). Our m…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 17 canonical work pages

  1. [1]

    Alignit: Enhancing prompt alignment in cus- tomization of text-to-image models

    Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Alignit: Enhancing prompt alignment in cus- tomization of text-to-image models. In2025 IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 4882–4890. IEEE, 2025. 2, 4, 6, 8, 12

    work page 2025

  2. [2]

    A neural space-time representation for text-to- image personalization

    Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to- image personalization. arXiv preprint arXiv:2305.15391 ,

    arXiv

  3. [3]

    Palp: Prompt aligned personalization of text-to- image models

    Moab Arar, Andrey V oynov, Amir Hertz, Omri Avrahami, Shlomi Fruchter, Yael Pritch, Daniel Cohen-Or, and Ariel Shamir. Palp: Prompt aligned personalization of text-to- image models. arXiv preprint arXiv:2401.06105, 2024. 2, 3, 4, 6, 8, 12

    arXiv 2024

  4. [4]

    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

    2021

  5. [5]

    Subject-driven text-to-image generation via apprenticeship learning

    Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. Advances in Neural Information Processing Sys- tems, 36:30286–30305, 2023. 3

    work page 2023

  6. [6]

    Custom-edit: Text-guided image edit- ing with customized diffusion models

    Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image edit- ing with customized diffusion models. arXiv preprint arXiv:2305.15779, 2023. 4

    arXiv 2023

  7. [7]

    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 learning,

  8. [8]

    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. 1, 2, 3, 4, 5, 6, 8, 12

    arXiv 2022

Show all 39 references
  1. [9]

    Encoder-based domain tuning for fast personalization of text-to-image models.ACM Transactions on Graphics (TOG), 42(4):1–13, 2023

    Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models.ACM Transactions on Graphics (TOG), 42(4):1–13, 2023. 3

    2023

  2. [10]

    Classdiffu- sion: More aligned personalization tuning with explicit class guidance

    Jiannan Huang, Jun Hao Liew, Hanshu Yan, Yuyang Yin, Yao Zhao, Humphrey Shi, and Yunchao Wei. Classdiffu- sion: More aligned personalization tuning with explicit class guidance. arXiv preprint arXiv:2405.17532, 2024. 4, 5, 12

    2024 arXiv

  3. [11]

    Taming encoder for zero fine-tuning image customization with text-to-image diffusion models

    Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 3

    2023 arXiv

  4. [12]

    Multi-concept customization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1931–1941, 2023. 2, 3, 4

    1931

  5. [13]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Ku- lal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M¨uller, Dustin Podell, Robin Rombach, Harry S...

  6. [14]

    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. Advances in Neural Information Pro- cessing Systems, 36:30146–30166, 2023. 3

    2023

  7. [15]

    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

  8. [16]

    Low-rank adaptation for fast text-to-image diffusion fine-tuning

    LoRA. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022. 1, 2, 3, 4, 5, 6, 8, 12

    2022

  9. [17]

    Unified multi-modal latent diffusion for joint subject and text conditional image generation.arXiv preprint arXiv:2303.09319, 2023

    Yiyang Ma, Huan Yang, Wenjing Wang, Jianlong Fu, and Jiaying Liu. Unified multi-modal latent diffusion for joint subject and text conditional image generation.arXiv preprint arXiv:2303.09319, 2023. 3

    2023 arXiv

  10. [18]

    Locating and editing factual associations in gpt

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Be- linkov. Locating and editing factual associations in gpt. Ad- vances in neural information processing systems, 35:17359– 17372, 2022. 2

    2022

  11. [19]

    Attndreambooth: To- wards text-aligned personalized text-to-image generation

    Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: To- wards text-aligned personalized text-to-image generation. Advances in Neural Information Processing Systems , 37: 39869–39900, 2024. 2, 4, 5, 12

    2024

  12. [20]

    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. 1, 2, 3, 4, 5, 8, 12

    2023 arXiv

  13. [21]

    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

  14. [22]

    Dreamblend: Advancing person- alized fine-tuning of text-to-image diffusion models

    Shwetha Ram, Tal Neiman, Qianli Feng, Andrew Stuart, Son Tran, and Trishul Chilimbi. Dreamblend: Advancing person- alized fine-tuning of text-to-image diffusion models. 2025. 1, 2, 3, 4, 5, 12

    2025

  15. [23]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International confer- ence on machine learning, pages 8821–8831. Pmlr, 2021. 2 9

    2021

  16. [24]

    Hierarchical text-conditional image gener- ation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    2022 arXiv

  17. [25]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 4, 5, 7, 8, 12

    2022

  18. [26]

    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

  19. [27]

    Photorealistic text-to-image diffusion models with deep language understanding

    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. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  20. [28]

    Where’s waldo: Diffusion features for person- alized segmentation and retrieval

    Dvir Samuel, Rami Ben-Ari, Matan Levy, Nir Darshan, and Gal Chechik. Where’s waldo: Diffusion features for person- alized segmentation and retrieval. Advances in Neural Infor- mation Processing Systems, 37:128160–128181, 2024. 2, 3, 6

    2024

  21. [29]

    In- stantbooth: Personalized text-to-image generation without test-time finetuning

    Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. In- stantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 8543–8552, 2024. 3

    2024

  22. [30]

    Emu: Generative pretraining in multimodality

    Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 2

    2023 arXiv

  23. [31]

    Emergent correspondence from image diffusion

    Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In NeurIPS, 2023. 3

    2023

  24. [32]

    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. In ACM SIGGRAPH 2023 conference proceedings , pages 1–11, 2023. 2

    2023

  25. [33]

    p+: Extended textual conditioning in text-to- image generation

    Andrey V oynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to- image generation. arXiv preprint arXiv:2303.09522, 2023. 2

    2023 arXiv

  26. [34]

    Instantid: Zero-shot identity-preserving generation in seconds

    Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3

    2024 arXiv

  27. [35]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943–15953, 2023. 2

    2023

  28. [36]

    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

  29. [37]

    Scaling autoregres- sive models for content-rich text-to-image generation

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gun- jan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin- fei Yang, Burcu Karagol Ayan, et al. Scaling autoregres- sive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2

    2022 arXiv

  30. [38]

    A survey on personalized content synthesis with diffusion models

    Xulu Zhang, Xiaoyong Wei, Wentao Hu, Jinlin Wu, Jiaxin Wu, Wengyu Zhang, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models. arXiv preprint arXiv:2405.05538, 2024. 1 10 A. Additional qualitative results We present additiona...

    2024

  31. [2025]

    1, 2, 3, 4, 5, 8, 12

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.