Pith. sign in

REVIEW 5 major objections 6 minor 35 references

AdaptaGen: Domain-Specific Image Generation through Hierarchical Semantic Optimization Framework

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read AdaptaGen claims that domain-specific image generation under extreme data scarcity improves when prompt engineering and model adaptation are optimized together, achieving better FID, IS, and CLIP scores than two recent baselines using…

desk verdict A useful integration of known components whose semantic claim is circular and whose tables have arithmetic errors; fixable, and a referee should see it. read the letter →

arxiv 2507.05621 v1 pith:7DDFNZ4O submitted 2025-07-08 cs.CV cs.MM

Suoxiang Zhang , Xiaxi Li , Hongrui Chang , Zhuoyan Hou , Guoxin Wu , Ronghua Ji This is my paper · ORCID
classification cs.CVcs.MM
keywords domain-specificimagegenerationfew-shotpromptoptimizationLow-RankAdaptation(LoRA)semanticconsistencyhallucinationmitigationT5captiontransformationCLIPsimilarity
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

AdaptaGen claims that domain-specific image generation under extreme data scarcity is best treated as one pipeline in which prompt engineering and model adaptation are optimized together, not as separate stages. The paper proposes a three-stage framework: BLIP-2 generates multiple candidate captions per image, a CLIP similarity matrix selects the best one, LoRA adapts a diffusion model with those captions, and a two-phase T5 transformation recombines captions during synthesis to preserve core meaning while increasing diversity. Using only 16 images per category across 40 categories from four datasets, the authors report an average FID-1k of 80.90 versus 156.80 for DoRA and 109.64 for DataDream, with higher IS and CLIP scores. A sympathetic reader would care because the claim, if true, would give resource-constrained specialized fields like agriculture and medicine a practical recipe for generating large sets of semantically consistent training images from a handful of examples.

What carries the argument

The load-bearing mechanism is a similarity matrix built from CLIP image and text embeddings, $S(i,j) = \frac{(E_I(x_i))^T E_T(c_{ij})}{\|E_I(x_i)\|\|E_T(c_{ij})\|}$, used to rank BLIP-2's multi-perspective candidate captions; this selected caption then conditions LoRA adaptation of a Stable Diffusion model, and at generation time a two-phase T5 transformation (controlled randomization with a temperature $\tau$, plus cross-image feature fusion when more outputs than captions are needed) rewrites captions before they condition the diffusion model. The matrix selection couples semantics to visual adaptation, and the T5 phase enforces domain constraints while adding diversity.

What would settle it

Re-evaluate the 40 categories with an independent semantic metric that was never used to select captions — human preference judgments or a second vision-language model different from CLIP — and check whether AdaptaGen's CLIP-score advantage (0.32 vs 0.29) survives; if the advantage shrinks or reverses, the claimed semantic-consistency gain is an artifact of selection-evaluation coupling.

Watch

Extended reading notes

Core claim

The central claim is that treating prompt selection and parameter adaptation as one coupled optimization, then constraining synthesis with a T5-based semantic transformation, reduces hallucinations and improves distribution alignment in few-shot domain-specific generation. Concretely, the paper asserts that on 40 categories from ImageNet100, Sun397, Food101, and PlantLeafDiseases, AdaptaGen outperforms DoRA and DataDream on FID (80.90 vs 156.80 and 109.64), IS (2.87 vs 2.28 and 2.14), and CLIP score (0.32 vs 0.29 and 0.29), with an ablation showing that removing the semantic transformation degrades FID to 126.64. The authors present these numbers as evidence that their hierarchical semantic optimization captures domain-specific terminology and visual relationships from very few examples.

Load-bearing premise

The argument rests on CLIP embedding similarity being a trustworthy measure of caption quality and of semantic consistency, even though the same CLIP model both picks the captions and scores the outputs; if that measure is biased, the reported semantic-consistency advantage over baselines may be circular.

Editorial extensions

If this is right

  • If correct, practitioners can adapt text-to-image models for a new specialized category with 16 images and automated captions, removing manual prompt writing.
  • The framework claims to reduce semantic, attribute, spatial, and structural hallucinations, which would make synthetic data more usable for downstream tasks in domains like plant disease diagnosis.
  • The two-phase semantic transformation is the key component: removing it raises FID from 80.90 to 126.64, so caption recombination drives most of the reported benefit.
  • The integrated treatment of prompt optimization and LoRA adaptation outperforms weight-decomposition (DoRA) and template-based dataset generation (DataDream) under the same data budget on the tested metrics.

Reading between the lines

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

  • Because CLIP similarity selects captions and CLIP score is the semantic metric, part of the CLIP improvement could reflect selection-evaluation alignment; an independent semantic judge (human ratings or a different vision-language model) would be needed to confirm genuine semantic gains.
  • The method's success on 10 randomly chosen categories per dataset suggests but does not establish that it scales to full datasets or to truly specialized vocabularies; testing all classes of PlantLeafDiseases would be a natural next check.
  • The same two-phase caption recombination could transfer to other conditional generators or to data-augmentation loops for classification training, since it is model-agnostic beyond the diffusion backbone.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. This paper presents AdaptaGen, a three-stage pipeline for few-shot domain-specific image generation. Stage 1 (Sec. 3.1) uses BLIP-2 to produce multiple candidate captions per training image and CLIP image-text similarity (Eqs. 2-3) to select one caption per image; Stage 2 (Sec. 3.2) adapts Stable Diffusion with LoRA; Stage 3 (Sec. 3.3) applies a two-phase T5-based caption transformation with controlled randomization and cross-image fusion to generate diverse images. The experiments use 16 images per category across 40 categories from Food101, ImageNet100, Sun397, and PlantLeafDiseases, comparing FID-1k, Inception Score, and CLIP Score with DoRA and DataDream baselines. The authors report consistent improvements and claim reduced hallucinations and improved semantic consistency, with best overall FID-1k 80.90 vs 156.80/109.64, IS 2.87, and CLIP score 0.32.

Significance. The underlying idea—treating prompt optimization and LoRA-based adaptation as coupled and adding a caption-transformation stage for diversity—is relevant and potentially useful for data-scarce domains, and the reported FID/IS improvements, if reproducible, would be practically meaningful. The paper also addresses a real gap by considering the full generation pipeline rather than model adaptation alone. However, the current validation is not commensurate with the claims: the headline semantic-consistency metric is produced by the same CLIP model used for prompt selection, no uncertainty quantification or significance testing is reported, the hallucination-reduction claims rest on aggregate proxies rather than direct measurements, and the method description is too incomplete to reproduce. These issues are load-bearing for the central contributions, so the paper cannot be accepted on its current evidence.

major comments (5)
  1. [Section 3.1, Eqs. (2)-(3), and Table 3] The prompt-selection stage chooses the caption maximizing CLIP image-text similarity, and Table 3's semantic-consistency metric is a CLIP score on generated images. Any pipeline that reranks its text inputs with CLIP will tend to score higher on CLIP than fixed-template baselines, so the reported 0.32 vs 0.29 advantage is expected by construction rather than evidence of more truthful or less hallucinated captions. To support the semantic-consistency and hallucination-reduction claims, please provide an independent evaluation (e.g., human ratings, a second VLM or embedding model not used for selection, or a caption-retrieval task) and show that the CLIP gain persists under that metric.
  2. [Section 4.2 and Tables 1-3] All quantitative comparisons are point estimates with no error bars, repeated seeds, or significance tests. Under the paper's own data-scarce setting (16 training images per category) the variance of FID/IS estimates is likely high, and the claim of 'superior performance across 40 categories' is not supported by dataset-level averages alone. The text repeatedly refers to per-category improvements (e.g., Sec. 4.2.1 and Sec. 4.3.3), but no per-category results are shown; please report the complete breakdown and assess statistical reliability.
  3. [Section 4.2.4 vs 4.3.2] The ablation variant Ours(w/o) is defined in Sec. 4.2.4 as lacking the two-phase caption semantic transformation and in Sec. 4.3.2 as lacking the feature fusion mechanism, yet Sec. 3.3 presents feature fusion as part of that same two-phase mechanism. As implemented, the ablation removes both the controlled-randomization phase and the cross-image fusion phase together, so it cannot isolate the contribution of feature fusion to the FID improvement from 126.64 to 80.90. Please ablate the two phases separately.
  4. [Section 4.3.3] The paper identifies four hallucination types (semantic, attribute, spatial, structural) but none is directly measured. The discussion infers hallucination reduction only from FID, IS, and CLIP scores, which are aggregate distributional and alignment proxies and cannot distinguish a genuinely correct domain-specific image from one that happens to be CLIP-aligned. Without human evaluation, domain-expert assessment, or a task-specific correctness metric, the conclusion that AdaptaGen 'effectively mitigates hallucinations' is overstated.
  5. [Section 3.3 and experimental setup] The method is underspecified for reproduction. Eq. (7) applies a temperature τ to T5 without explaining the sampling/decoding procedure; the 'intelligent content synthesis' fusion step has no algorithmic description; the number of candidate captions m and the BLIP-2 prompt templates are not given; the random selection of 10 categories per dataset is unseeded; and no code, splits, or category lists are provided. Eq. (8) uses a single optimized caption c* even though the second phase describes combining captions. Please provide pseudocode, hyperparameters, and a public implementation or detailed protocol.
minor comments (6)
  1. [Page 1 and Table 2] Page 1 contains formatting errors in the title and running header ('DOMAIN-SPECIFICIMAGEGENERATION', 'HIERARCHICALSEMANTICOPTIMIZATIONFRAMEWORK'), and Table 2's DoRA row concatenates numbers ('2.973.352.26'); please fix typography.
  2. [Tables 1-3] The column order of datasets is inconsistent: Table 1 uses Food101/ImageNet/Sun397/PLDiseases while Tables 2 and 3 use Sun397/PLDiseases/Food101/ImageNet; please harmonize.
  3. [References] References [3] and [29], and [4] and [30], are duplicates; reference [32] appears to misattribute the authors of 'Towards a Unified View of Parameter-Efficient Transfer Learning'. Please correct the bibliography.
  4. [Table 2] The 'Original dataset' row in Table 2 is confusing: Inception Score is not a standard metric for real images, and it is unclear how the row was computed and why it is a meaningful reference.
  5. [Section 3.2, Eq. (6)] Eq. (6) writes the LoRA scaling as α r/d; since r and d are not defined consistently, please state the standard scaling (usually α/r) and define all symbols.
  6. [Section 4.3.3] There is an internal inconsistency in Sec. 4.3.3: the Food101 bullet reports a CLIP score of 0.32, but Table 3 gives 0.31 for Food101 and 0.32 for ImageNet100; please correct.

Circularity Check

1 steps flagged · score 6.0 of 10

CLIP is both the prompt-selection objective and the semantic-consistency metric, so the reported consistency gain is partly built into the method.

  1. self definitional [Section 3.1, Eqs. (2)-(3); Section 4.1, Evaluation Metrics; Table 3]
    "where EI (xi) and ET (cij) represent image and text embeddings from CLIP’s encoders. Through this matrix optimization, we select the most semantically relevant description for each image: c∗ i = arg max j∈{1,...,m} S(i,j) (3) ... CLIP Score: We measure semantic alignment between generated images and their corresponding prompts."

    The prompt-selection stage ranks candidate captions by exactly the CLIP image-text cosine similarity that later defines the headline semantic metric. Eq. (2) defines S(i,j) using CLIP encoders, and Eq. (3) chooses the caption that maximizes that similarity; the selected caption then conditions LoRA training and generation. Section 4.1 then reports 'CLIP Score' as 'semantic alignment between generated images and their corresponding prompts', i.e. the same CLIP similarity applied to the same selected prompts. Because the evaluation metric and the optimization objective share the same quantity, part of AdaptaGen's higher CLIP score (0.32 vs 0.29) is expected by construction from the reranking step, so it does not independently establish better semantic consistency or reduced hallucination.

full rationale

The FID-1k and IS results are measured with Inception-based metrics external to the method and are not circular. The main circularity is confined to the semantic-consistency/hallucination claim: the same CLIP embedding similarity is used both to select optimal captions (Eqs. 2-3) and to score generated images (Table 3), so the reported CLIP improvement is partly an artifact of optimizing the evaluation metric. This does not make the entire framework definitionally circular, because distribution alignment and diversity remain externally testable, but it does mean the paper's central semantic-consistency and hallucination-reduction claim is not independently supported. There is no load-bearing self-citation chain or ansatz smuggled in via citation; the issue is the conflation of an optimization objective with an evaluation metric.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical or mathematical entities; the 'feature fusion mechanism' is an unspecified algorithmic component, not an invented entity. The central claim rests on domain assumptions about CLIP validity, BLIP-2 coverage, T5 preservation, LoRA sufficiency, and FID reliability, plus several unreported hyperparameters.

free parameters (7)
  • tau_base = 0.8
    Hand-chosen T5 temperature for caption transformation (Eq. 7); no sensitivity analysis or justification provided.
  • delta_tau = 0.2
    Hand-chosen temperature range for controlled randomization (Eq. 7); arbitrary value.
  • LoRA_scale_s = not reported
    Eq. 8 leaves the LoRA scaling factor unspecified; it controls adaptation strength and is critical for generation quality.
  • guidance_scale_omega = not reported
    Diffusion guidance scale in Eq. 8 is not reported, yet it strongly affects image-text alignment and diversity.
  • inference_steps_n = not reported
    Number of diffusion inference steps in Eq. 8 is not reported, affecting quality and runtime.
  • LoRA_rank_r = not reported
    Rank of the low-rank matrices in Eq. 5-6 is not reported; it controls the capacity of domain adaptation with only 16 images per category.
  • candidate_descriptions_m_per_image = not specified
    The number of BLIP-2 candidate captions generated per image (Eq. 1) is not given, though it defines the prompt optimization space.
assumptions (5)
  • domain assumption CLIP embeddings measure semantic similarity between images and text.
    Eq. 2 defines prompt selection by CLIP cosine similarity; if CLIP is not a faithful semantic measure, the selected captions are not necessarily optimal.
  • domain assumption BLIP-2 candidate captions cover the relevant semantic dimensions.
    Section 3.1 assumes multi-perspective templates capture object, scene, subject, and context, but the templates are not shown or validated.
  • ad hoc to paper T5 transformation preserves domain semantics under temperature perturbation.
    Section 3.3 assumes transformed captions retain core themes while adding variation; this is asserted with arbitrary temperature values and no independent validation.
  • domain assumption LoRA with 16 images per category is sufficient to adapt Stable Diffusion v2.1.
    The entire training design relies on this premise; no analysis of overfitting, rank sensitivity, or data scaling is provided.
  • domain assumption FID-1k computed with the stated protocol is a reliable distributional metric.
    The real reference set size is not stated, and FID is known to be unstable with small reference sets; the paper does not address this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AdaptaGen: Domain-Specific Image Generation through Hierarchical Semantic Optimization Framework." pith.science (2026). https://pith.science/paper/7DDFNZ4O

@misc{pith2026250705621,
  author       = {Pith},
  title        = {Pith review of: AdaptaGen: Domain-Specific Image Generation through Hierarchical Semantic Optimization Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7DDFNZ4O}},
  note         = {Machine review of arXiv:2507.05621}
}
read the original abstract

Domain-specific image generation aims to produce high-quality visual content for specialized fields while ensuring semantic accuracy and detail fidelity. However, existing methods exhibit two critical limitations: First, current approaches address prompt engineering and model adaptation separately, overlooking the inherent dependence between semantic understanding and visual representation in specialized domains. Second, these techniques inadequately incorporate domain-specific semantic constraints during content synthesis, resulting in generation outcomes that exhibit hallucinations and semantic deviations. To tackle these issues, we propose AdaptaGen, a hierarchical semantic optimization framework that integrates matrix-based prompt optimization with multi-perspective understanding, capturing comprehensive semantic relationships from both global and local perspectives. To mitigate hallucinations in specialized domains, we design a cross-modal adaptation mechanism, which, when combined with intelligent content synthesis, enables preserving core thematic elements while incorporating diverse details across images. Additionally, we introduce a two-phase caption semantic transformation during the generation phase. This approach maintains semantic coherence while enhancing visual diversity, ensuring the generated images adhere to domain-specific constraints. Experimental results confirm our approach's effectiveness, with our framework achieving superior performance across 40 categories from diverse datasets using only 16 images per category, demonstrating significant improvements in image quality, diversity, and semantic consistency.

Figures

Figures reproduced from arXiv: 2507.05621 by the authors.

Figure 1
Figure 1. Comparison of prompt handling strategies: AdaptaGen’s semantic transformation approach versus [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The three-stage hierarchical semantic optimization framework of AdaptaGen, illustrating matrix-based prompt [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages

  1. [1]

    Multimodal datasets: misogyny, pornography, and malignant stereotypes

    Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. InProceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021

    work page 2021

  2. [2]

    Ap-adapter: Improving generalization of automatic prompts on unseen text-to-image diffusion models.Advances in Neural Information Processing Systems, 37:98320–98346, 2024

    Yuchen Fu, Zhiwei Jiang, Yuliang Liu, Cong Wang, Zexuan Deng, Zhaoling Chen, and Qing Gu. Ap-adapter: Improving generalization of automatic prompts on unseen text-to-image diffusion models.Advances in Neural Information Processing Systems, 37:98320–98346, 2024

    work page 2024

  3. [3]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn 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

    2022

  4. [4]

    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. InInternational conference on machine learning, pages 8821–8831. Pmlr, 2021

    2021

  5. [5]

    DoRA: Weight-decomposed low-rank adaptation

    Xiang Liu, Kevin Lin, Haotian Zhang, Xiangyu Xiao, Di Qi, and Kate Saenko. DoRA: Weight-decomposed low-rank adaptation. InInternational Conference on Learning Representations, 2024

    work page 2024

  6. [6]

    AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning

    Zhixiong Zhang, Jiaqi Sun, Xin Yang, Hao Wang, Ying Nian Wu, and Song-Chun Zhu. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning. InInternational Conference on Learning Representations, 2023

    work page 2023

  7. [7]

    QLoRA: Efficient finetuning of quantized LLMs

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient finetuning of quantized LLMs. InAdvances in Neural Information Processing Systems, 2023

    2023

  8. [8]

    Design guidelines for prompt engineering text-to-image generative models

    Yian Liu and Lydia Chilton. Design guidelines for prompt engineering text-to-image generative models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, 2022

    work page 2022

Show all 35 references
  1. [9]

    Prompt engineering for text-to-image generative models

    Jonas Oppenlaender. Prompt engineering for text-to-image generative models. InProceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, 2022. 11 AdaptaGen: Domain-Specific Image Generation

    2022

  2. [10]

    DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models

    Guangcong Wang, Weixi Luo, Yongle Yu, Yizhou Wu, Xiaoyun Zhang, Shunxin Wang, Guanqing Zeng, and Jing Dong. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. InNeurIPS, 2022

    2022

  3. [11]

    Easily accessible text-to-image generation amplifies demographic stereotypes at large scale

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. InProceedings of the 2023 A...

    2023

  4. [12]

    Prompt-free diffusion: Taking" text" out of text-to-image diffusion models

    Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking" text" out of text-to-image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8682–8692, 2024

    2024

  5. [13]

    Best prompts for text-to-image models and how to find them

    Nikita Pavlichenko and Dmitry Ustalov. Best prompts for text-to-image models and how to find them. In Proceedings of the SIGIR ’23 Conference, 2023

    2023

  6. [14]

    LoRA: Low-rank adaptation of large language models

    Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, 2022

    2022

  7. [15]

    VeRA: Vector-based random matrix adaptation

    Piotr Kopiczko, Enzo Tartaglione, Marco Grangetto, and Damian Borth. VeRA: Vector-based random matrix adaptation. InInternational Conference on Learning Representations, 2024

    2024

  8. [16]

    Optimizing prompts for text-to-image generation.Advances in Neural Information Processing Systems, 36:66923–66939, 2023

    Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation.Advances in Neural Information Processing Systems, 36:66923–66939, 2023

    2023

  9. [17]

    Datadream: Few-shot guided dataset generation

    Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. Datadream: Few-shot guided dataset generation. InEuropean Conference on Computer Vision, pages 252–268, 2024

    2024

  10. [18]

    BeautifulPrompt: Towards automatic prompt engineering for text-to-image synthesis

    Tianyi Cao, Chunhui Wang, Bang Liu, Zefan Wu, Jianqing Zhu, and Jie Huang. BeautifulPrompt: Towards automatic prompt engineering for text-to-image synthesis. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2023

    2023

  11. [19]

    Neuroprompts: An adaptive framework to optimize prompts for text-to-image generation

    Shachar Rosenman, Vasudev Lal, and Phillip Howard. Neuroprompts: An adaptive framework to optimize prompts for text-to-image generation. InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages ...

    2024

  12. [20]

    PromptCharm: Text-to-image generation through multi-modal prompting and refinement

    Zihao Wang, Yunhao Huang, Diyi Song, Lucy Ma, and Tongshuang Zhang. PromptCharm: Text-to-image generation through multi-modal prompting and refinement. InProceedings of the CHI Conference on Human Factors in Computing Systems, 2024

    2024

  13. [21]

    Automated black-box prompt engineering for personalized text-to-image generation.arXiv preprint arXiv:2403.19103, 2(5), 2024

    Yutong He, Alexander Robey, Naoki Murata, Yiding Jiang, Joshua Williams, George J Pappas, Hamed Has- sani, Yuki Mitsufuji, Ruslan Salakhutdinov, and J Zico Kolter. Automated black-box prompt engineering for personalized text-to-image generation.arXiv preprint arXiv:2403.19103,...

    2024 arXiv

  14. [22]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  15. [23]

    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, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  16. [24]

    Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

    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.Journal of machine learning research, 21(140):1–67, 2020

    2020

  17. [25]

    Imagenet: A large-scale hierarchical image database.2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database.2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009

    2009

  18. [26]

    Sun database: Large-scale scene recognition from abbey to zoo

    Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485–3492. IEEE, 2010

    2010

  19. [27]

    Food-101–mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. InEuropean Conference on Computer Vision, pages 446–461. Springer, 2014

    2014

  20. [28]

    Geetharamani and J

    G. Geetharamani and J. Arun Pandian. Identification of plant leaf diseases using a nine-layer deep convolutional neural network.Computers & Electrical Engineering, 76:323–338, 2019

    2019

  21. [29]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 12 AdaptaGen: Domain-Specific Image Generation

    2022

  22. [30]

    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. InInternational Conference on Machine Learning, 2021

    2021

  23. [31]

    Parameter-efficient transfer learning for NLP

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. InInternational Conference on Machine Learning, 2019

    2019

  24. [32]

    Towards a unified view of parameter- efficient transfer learning

    Junxian He, Chunting Zhou, Xuezhe Ma, Taylor L Berg, and Peilin Liang. Towards a unified view of parameter- efficient transfer learning. InInternational Conference on Learning Representations, 2021

    2021

  25. [33]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. InAnnual Meeting of the Association for Computational Linguistics, 2021

    2021

  26. [34]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing, 2021

    2021

  27. [35]

    BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models

    Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. InAnnual Meeting of the Association for Computational Linguistics, 2021. 13

    2021

Pith tools

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