Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Translation of Text Embedding via Delta Vector to Suppress Strongly Entangled Content in Text-to-Image Diffusion Models

T0 review · 4 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 The paper claims that translating a target token's text embedding by the embedding of the unwanted word—added to keys, subtracted from values in cross-attention—suppresses strongly entangled content that other methods leave in generated ima

desk verdict A simple delta-vector mechanism that empirically beats existing suppression baselines, but the paper doesn't show the key alignment that makes it work — worth reviewing, not yet fully convincing. read the letter →

arxiv 2508.10407 v2 pith:7XGI3PFE submitted 2025-08-14 cs.CV

Eunseo Koh , Seunghoo Hong , Tae-Young Kim , Simon S. Woo , Jae-Pil Heo This is my paper
classification cs.CV
keywords text-to-imagediffusioncontentsuppressiondeltavectorcross-attentiontextembeddingpersonalizedmodelszero-shoteditingentangledattributes
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 addresses a specific failure of text-to-image models: when a prompt names something whose identity is entangled with an attribute, such as a famous person and a mustache, the attribute appears even if the user asks for it not to. The authors claim the entanglement can be cut directly in the text-embedding path by translating the target word's embedding with a delta vector equal to the embedding of the unwanted word, added to the key path and subtracted from the value path in cross-attention. They name this Selective Suppression with $\Delta$ Vector (SSDV) and report that, on their strongly-entangled-prompt benchmark, it suppresses content that negative prompting, prompt editing, semantic guidance, and inpainting-based methods leave in place. For personalized fine-tuned models, they optimize the delta using inversion and attention-mask losses to suppress content while preserving the subject's identity. The reason to care is a mechanism claim: a single zero-shot vector operation on the text embedding, routed through softmax competition, can localize and erase content that resists other ablation strategies.

What carries the argument

The load-bearing object is the delta vector $\delta$, defined as the bare embedding of the negative-content word, replicated to the target token's width, together with the signed translation of the target embedding in the key and value branches. Its power comes from two facts the paper uses: linearity of $f_k$ and $f_v$ (Eq. 4), which decomposes the modified projection into the original feature plus $\alpha f(\delta)$, and softmax competition between the translated key embedding and the original target token, which concentrates the suppression on the region where the unwanted content would be generated. The 'selective' part is the same softmax: the extra token draws attention to the content

What would settle it

Take any benchmark pair, generate images with and without the unwanted attribute, and compute the cosine similarity between $f_v(\delta)$, the projected embedding of the bare unwanted word, and the mean value-feature difference between the two image sets. If the similarity is close to zero, the value subtraction in Eq. 8 cannot be the mechanism that removes the attribute; likewise, re-running SSDV with $\alpha_k=0$ while keeping $\alpha_v<0$ would show whether the key-side localization is actually doing the work.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that strongly entangled unwanted content persists because the target token's value feature still carries the visual signature of the attribute, even when the prompt wording tells the model not to draw it. The fix is to translate the target embedding as $e_w^* = e_w + \alpha\delta$, with $\delta$ the token embedding of the negative content (mean-pooled over tokens if needed), and to use two versions of the translated embedding inside cross-attention: $e_k^* = e_w + \alpha_k\delta$ with $\alpha_k > 0$ for the key path, and $e_v^* = e_w + \alpha_v\delta$ with $\alpha_v < 0$ for the value path. Because the key/value projections are linear, $f_v(e_v^

Load-bearing premise

Everything rests on the assumption that the projected embedding of the bare unwanted word points in the same direction as the visual features that draw that content into the image; if it points elsewhere, subtracting it cannot remove the content.

Editorial extensions

If this is right

  • On base text-to-image models, the delta is obtained zero-shot: no per-prompt training and no weight update, just one text-encoder forward pass to read off the unwanted word's embedding.
  • The paper's ablation results imply that the positive key translation and the negative value translation are jointly necessary: applying the delta to only one branch leaves the content in the image.
  • For personalized models, the optimization-based delta produces suppression that preserves the subject's identity better than the zero-shot delta, according to the paper's user study.
  • The mechanism is designed to be local: suppression acts in the region attended by the competing key token, so unrelated objects in the image can survive the edit.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to build $\delta$ from the difference of two embeddings—for example, 'with glasses' minus 'without glasses'—rather than the bare 'glasses' token; that difference may align with the visual feature direction more directly.
  • The inverse sign convention (subtract from keys, add to values) would plausibly insert an entangled attribute rather than suppress it, making the same machinery a zero-shot attribute-insertion tool.
  • The localization argument depends on softmax competition, so attributes that are spatially diffuse, such as lighting, texture, or style, may not be suppressible without global side effects; testing SSDV on such attributes would delimit the claim.
  • A quantitative check the paper does not report—cosine similarity between $f_v(\delta)$ and the value-space direction separating images with and without the content—would directly test the mechanism's load-bearing assumption.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper addresses the problem of suppressing strongly entangled negative content (e.g., a mustache when generating 'Charlie Chaplin') in text-to-image diffusion models. It proposes a 'delta vector' in text embedding space, obtained zero-shot as the CLIP embedding of the negative-content word, and modifies the text embedding of the target token by adding/subtracting this vector. A selective suppression method (SSDV) applies the delta to key and value embeddings in cross-attention with opposite signs, using a replicated token to localize suppression. For personalized models, the paper also proposes an optimization-based delta. Experiments on a self-constructed SEP-Benchmark and on DreamBooth models report improvements over several baselines in CLIP, IFID, DetScore, and user preference. The central mechanism is asserted in Sec. 4.3.1 and evaluated qualitatively; the paper does not provide a quantitative alignment check between the projected delta and the visual feature of the negative content.

Significance. If the mechanism is sound, the zero-shot delta-vector approach would be a simple and training-free method for a persistent failure mode of T2I models, and the cross-attention competition idea could be useful for localized content suppression. The paper also extends to personalized models, where prior suppression methods struggle. Code is promised, which is commendable. However, the empirical evidence is suggestive rather than conclusive: the core alignment assumption is untested, the main benchmark is small and self-constructed, and the evaluation metrics partly conflate suppression with image distortion. The contribution is potentially valuable but needs stronger validation before the claims can be accepted.

major comments (4)
  1. [Sec. 4.3.1, Eqs. (4)-(5)] The load-bearing premise is that f_v(delta), where delta is the bare CLIP embedding of the negative word, points along the visual feature of that content in the image. Eq. (4) only shows that the value projection is linear; it does not establish that f_v(delta) is the direction that removes the mustache/glasses/berry. The paper never measures the alignment between f_v(delta) and the actual visual feature, and the attention-map evidence (Fig. 5c) is admittedly unclear. Please add a quantitative check--e.g., cosine similarity between f_v(delta) and the value-feature change that best suppresses the content, or attention-map IoU with ground-truth negative-content masks--and a negative control using a random delta vector. Without this, the observed improvement could stem from a generic embedding perturbation rather than selective suppression.
  2. [Sec. 5.1, Table 1] The IFID metric is defined as the FID between original and suppressed image distributions, with higher values claimed to indicate stronger suppression. This conflates suppression with distortion: a method that removes the subject itself (as Inst-Inpaint does in the 'S* toy duck - yellow' row) achieves high IFID despite being undesirable. The paper needs a separate preservation metric (e.g., CLIP similarity to the original image or subject-identity score, or LPIPS) to show that the suppression improvement is not bought by global image corruption. The user study alone does not disentangle these factors.
  3. [Sec. 5.3, Table 3] The ablation study is performed on a single example, reports no error bars or statistical significance, and uses an 'OwlViT' metric that is never defined in the main text. Please run the ablation over the full SEP-Benchmark and report mean and standard deviation, and clarify what OwlViT measures. In addition, the attention map in Fig. 5(c) is described as 'unclear,' which weakens the localization claim; please provide a quantitative localization analysis with masks rather than a single qualitative map.
  4. [Sec. 4.4.1, Eqs. (10)-(14)] The optimization objective sets both alpha_k and alpha_v to positive values, while the deployed SSDV uses alpha_v < 0. This apparent inconsistency is not explained. Also, the optimized delta is evaluated only by a user preference study (Table 2); please report the same quantitative metrics used in Table 1 (CLIP, IFID, DetScore) for optimized versus zero-shot delta, so readers can assess the claimed improvement quantitatively.
minor comments (5)
  1. [Sec. 5.1] The definition of IFID should be reworded: higher FID does not inherently mean 'stronger suppression'; it means larger distribution shift. The authors should clarify the intended interpretation and connect it to prior work.
  2. [Sec. 4.2] The phrase 'the delta vector's influence on the word embedding increases when alpha > 0 and decreases when alpha < 0' is imprecise: adding alpha*delta with alpha<0 still changes the embedding, just in the opposite direction. Please rephrase.
  3. [Sec. 5.2.1] The claim that NP is 'second-best scores' should specify for which metric; in Table 1, NP is second-best in CLIP and DetScore but not in IFID or Prefer.
  4. [Eq. (5)] When the negative content consists of multiple tokens, mean-pooling is mentioned but no justification is given. Please add a brief validation or reference for this choice.
  5. [General] Some notation is inconsistent: the paper uses both 'e*_k' and 'e*_v' for modified embeddings but the superscripts are not defined clearly in Eqs. (7)-(8). Also, Figure 2 is dense and hard to read; consider enlarging or splitting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the zero-shot delta is a frozen CLIP embedding, Eq. 4-6 are linearity identities rather than manufactured suppression, and the main empirical claims are backed by external detection and user-study metrics.

full rationale

The paper's zero-shot delta is defined as the frozen CLIP text embedding of the negative content (Eq. 5); it is not fitted to the evaluation metrics, and no parameter is optimized on the SEP-Benchmark or on the CLIP/DetScore test quantities. Equations (4) and (6) are algebraic consequences of the linearity of the value and key projection layers: they decompose f_v(e_w + alpha_v delta) and the attention logit into a baseline term plus a delta term. They do not by themselves guarantee that f_v(delta) is the visual feature of the negative content, nor that subtracting it removes that content; that is an empirical premise that the paper tests qualitatively and quantitatively. The lack of a direct alignment measurement between f_v(delta) and the entagled visual feature is a correctness/valuation limitation -- and the paper itself admits the attention map in Fig. 5(c) 'may appear unclear' -- but an unverified premise is not circularity, because the outcome is not contained in the input by construction. For the personalized-model variant, the optimized delta is trained with mask-based losses (Eqs. 11, 13), so the method is explicitly fitting a subject-specific delta rather than making a zero-shot prediction; the user study compares identity preservation, which is not directly optimized, so the evaluation is not statistically forced. The only self-citation is reference [3] in the related-work enumeration of personalized methods; it is cited alongside many external works and is not load-bearing. External evidence such as DetScore (MMDetection) and the human preference study provides support that is independent of the text-embedding manipulation itself. Overall, the derivation chain does not reduce to its inputs, and no circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The paper's central claim rests on two constructs (delta vector, competition token), two hand-tuned coefficients with unreported magnitudes (alpha_k, alpha_v), an unspecified weighting coefficient (lambda_mask), and an assumed alignment between text-space direction and image-space visual features (Eq 4). These are modest costs for an applied method paper, but the reader pays for the tuned coefficients, the mask-fidelity assumption, and the CLIP-space evaluation, not for a derivation. No new physical entities are postulated.

free parameters (4)
  • alpha_k (key modification coefficient) = sign fixed positive (Eq 7); magnitude unspecified in main text
    Controls how strongly the competition token marks negative-content regions; magnitude is tuned per benchmark, only signs are ablated in Table 3.
  • alpha_v (value modification coefficient) = sign fixed negative (Eq 8); magnitude unspecified in main text
    Controls suppression strength vs identity preservation; magnitude not reported.
  • lambda_mask (weighting coefficient in Eq 14) = unspecified
    Balances the reconstruction loss and attention-mask loss in the optimized delta; its value is never stated.
  • local blending binary-mask threshold = unspecified
    Separates significant from insignificant attention differences in Sec 4.5; details deferred to Suppl.3.
assumptions (5)
  • domain assumption Cross-attention projections f_k and f_v are linear, so f_v(e_w + alpha*delta) = f_v(e_w) + alpha*f_v(delta) (Eq 4, 6)
    Stated in Sec 4.3.1; true for the linear projection layers of SD, but the paper relies on the algebraic decomposition to justify the whole mechanism.
  • ad hoc to paper The projected embedding of the bare negative word f_v(delta) is the visual feature of that content, so subtracting it suppresses the content
    Sec 4.3.1, 'By defining delta as the embedding of negative content C, it can effectively suppress the features responsible for generating the negative content.' This is the load-bearing empirical premise, never verified beyond qualitative attention maps.
  • domain assumption First upsampling-block attention maps of e*_k approximate semantic masks of the negative content (L_mask in Eq 13)
    Sec 4.4.1, leaning on cited refs [17,18,20]; the paper does not validate the mask fidelity for its own subjects.
  • standard math DDIM inversion (Eq 12) gives a faithful reconstruction path for the optimization loss (Eq 11)
    Sec 4.4.1; standard assumption in the editing literature (Null-text inversion [19]), invoked without re-derivation.
  • domain assumption CLIP cosine similarity between generated image and negative text is a valid measure of suppression
    Sec 5.1; used as the primary quantitative metric, but it lives in the same CLIP representation family the method manipulates.
invented entities (2)
  • delta vector (embedding displacement delta)
    purpose: Translate the target word's text embedding toward or away from the negative-content direction to suppress that content in the image
    Methodological construct, not an observed entity; it is defined as the frozen embedding of the negative word (Eq 5), so it has no falsifiable handle independent of the suppression behavior it is introduced to produce.
  • competition token e*_k (replicated delta-modified token appended to K)
    purpose: In the softmax over keys, absorb attention in regions where negative content would appear, preventing the entangled target token from attending there
    Introduced in Eq 7 to make suppression selective; its attention-locating property is asserted and shown only qualitatively (Fig 5, one map of which the authors call unclear).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Translation of Text Embedding via Delta Vector to Suppress Strongly Entangled Content in Text-to-Image Diffusion Models." pith.science (2026). https://pith.science/paper/7XGI3PFE

@misc{pith2026250810407,
  author       = {Pith},
  title        = {Pith review of: Translation of Text Embedding via Delta Vector to Suppress Strongly Entangled Content in Text-to-Image Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7XGI3PFE}},
  note         = {Machine review of arXiv:2508.10407}
}
read the original abstract

Text-to-Image (T2I) diffusion models have made significant progress in generating diverse high-quality images from textual prompts. However, these models still face challenges in suppressing content that is strongly entangled with specific words. For example, when generating an image of "Charlie Chaplin", a "mustache" consistently appears even if explicitly instructed not to include it, as the concept of "mustache" is strongly entangled with "Charlie Chaplin". To address this issue, we propose a novel approach to directly suppress such entangled content within the text embedding space of diffusion models. Our method introduces a delta vector that modifies the text embedding to weaken the influence of undesired content in the generated image, and we further demonstrate that this delta vector can be easily obtained through a zero-shot approach. Furthermore, we propose a Selective Suppression with Delta Vector (SSDV) method to adapt delta vector into the cross-attention mechanism, enabling more effective suppression of unwanted content in regions where it would otherwise be generated. Additionally, we enabled more precise suppression in personalized T2I models by optimizing delta vector, which previous baselines were unable to achieve. Extensive experimental results demonstrate that our approach significantly outperforms existing methods, both in terms of quantitative and qualitative metrics.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 21 canonical work pages

  1. [1]

    Sega: Instructing text-to-image models using semantic guidance

    Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36,

    work page

  2. [2]

    Mmdetection: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019. 6

    arXiv 1906

  3. [3]

    Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation

    Jiwoo Chung, Sangeek Hyun, Hyunjun Kim, Eunseo Koh, MinKyu Lee, and Jae-Pil Heo. Fine-tuning visual autore- gressive models for subject-driven generation.arXiv preprint arXiv:2504.02612, 2025. 3

    work page Pith review arXiv 2025

  4. [4]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto- Kaufman, and David Bau. Erasing concepts from diffusion models. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 2426–2436, 2023. 2, 6

    work page 2023

  5. [5]

    Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K. Wong. Vico: Detail-preserving visual condition for person- alized text-to-image generation. 2023. 6

    work page 2023

  6. [6]

    Prompt-to-prompt im- age editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    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. 1, 2, 3, 6, 7

    arXiv 2022

  7. [7]

    Clipscore: A reference-free evaluation met- ric for image captioning.arXiv preprint arXiv:2104.08718,

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

    arXiv

  8. [8]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bern- hard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6

    2017

Show all 36 references
  1. [9]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1, 2

    2020

  2. [10]

    Count- ing guidance for high fidelity text-to-image synthesis.arXiv preprint arXiv:2306.17567, 2023

    Wonjun Kang, Kevin Galim, and Hyung Il Koo. Count- ing guidance for high fidelity text-to-image synthesis.arXiv preprint arXiv:2306.17567, 2023. 1

    2023 arXiv

  3. [11]

    Ablating con- cepts in text-to-image diffusion models

    Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating con- cepts in text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691–22702, 2023. 2

    2023

  4. [12]

    Multi-concept customization of text- to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shecht- man, and Jun-Yan Zhu. Multi-concept customization of text- to-image diffusion. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1931–1941, 2023. 2, 3, 5, 6

    1931

  5. [13]

    Multi-concept customization of text- to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shecht- man, and Jun-Yan Zhu. Multi-concept customization of text- to-image diffusion. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 1931–1941, 2023. 3

    1931

  6. [14]

    Styledif- fusion: Prompt-embedding inversion for text-based editing

    Senmao Li, Joost van de Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Styledif- fusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 3

    2023 arXiv

  7. [15]

    Get what you want, not what you don’t: Image content suppression for text- to-image diffusion models.arXiv preprint arXiv:2402.05375,

    Senmao Li, Joost van de Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Get what you want, not what you don’t: Image content suppression for text- to-image diffusion models.arXiv preprint arXiv:2402.05375,

    arXiv

  8. [16]

    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. InEuropean Conference on Computer Vision, pages 423–439, 2022. 1

    2022

  9. [17]

    Diffusion hyperfeatures: Searching through time and space for semantic correspondence.Ad- vances in Neural Information Processing Systems, 36:47500– 47510, 2023

    Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holyn- ski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence.Ad- vances in Neural Information Processing Systems, 36:47500– 47510, 2023. 6

    2023

  10. [18]

    Open-vocabulary atten- tion maps with token optimization for semantic segmentation in diffusion models

    Pablo Marcos-Manch ´on, Roberto Alcover-Couso, Juan C SanMiguel, and Jose M Mart ´ınez. Open-vocabulary atten- tion maps with token optimization for semantic segmentation in diffusion models. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition...

    2024

  11. [19]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 6

    2023

  12. [20]

    Emerdiff: Emerging pixel-level se- mantic knowledge in diffusion models.arXiv preprint arXiv:2401.11739, 2024

    Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, and Seung Wook Kim. Emerdiff: Emerging pixel-level se- mantic knowledge in diffusion models.arXiv preprint arXiv:2401.11739, 2024. 6

    2024 arXiv

  13. [21]

    Teaching clip to count to ten

    Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 3170–3180, 2023. 1

    2023

  14. [22]

    Localizing object-level shape variations with text-to-image diffusion models

    Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch- Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. InProceed- ings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 6

    2023

  15. [23]

    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. InInternational conference on machine learning, p...

    2021

  16. [24]

    Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    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. 1, 2

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3

    2022

  18. [26]

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

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 22...

  19. [27]

    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. InProceedings of the IEEE/CVF conference on computer vision and patt...

    2024

  20. [28]

    Photorealistic text-to-image diffusion models with deep lan- guage understanding.Advances in neural information pro- cessing systems, 35:36479–36494, 2022

    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 lan- guage understanding.Advances in neural informatio...

    2022

  21. [29]

    Dragdiffusion: Harnessing diffusion models for interactive point-based image editing

    Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Han- shu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2024

  22. [30]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

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

    2010 arXiv

  23. [31]

    Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019. 2

    2019

  24. [32]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 3

    2017

  25. [33]

    Imagereward: Learning and evaluating human preferences for text-to-image generation.Advances in Neural Information Processing Sys- tems, 36, 2024

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation.Advances in Neural Information Processing Sys- tems, 36, 2024. 2

    2024

  26. [34]

    Inst-inpaint: Instructing to remove objects with diffusion models.arXiv preprint arXiv:2304.03246, 2023

    Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models.arXiv preprint arXiv:2304.03246, 2023. 2, 6, 7

    2023 arXiv

  27. [35]

    Jedi: Joint- image diffusion models for finetuning-free personalized text- to-image generation

    Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting- Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint- image diffusion models for finetuning-free personalized text- to-image generation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognit...

    2024

  28. [36]

    Forget-me-not: Learning to forget in text-to- image diffusion models

    Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to- image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755–1764, 2024. 2

    2024

Pith tools

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