Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Text-Visual Semantic Constrained AI-Generated Image Quality Assessment

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a two-module framework, combining MLLM-generated descriptions with frequency-weighted visual pooling, outperforms every prior method on three AI-generated image quality benchmarks.

desk verdict Solid AGIQA engineering with a plausible new combination, but the SOTA claim rests on a single-split evaluation and missing baselines. read the letter →

arxiv 2507.10432 v3 pith:ZZR374PL submitted 2025-07-14 cs.CV

classification cs.CV
keywords AI-generatedimagequalityassessmenttext-imagealignmentmultimodallargelanguagemodelfrequencydomainanalysiscontrastsensitivityfunctionmixtureofexpertsperceptualdistortion
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 AI-generated image quality can be scored well by separating two questions: whether the image fulfills its text prompt, and whether fine details are perceptually clean. For the first question, it replaces the direct image-prompt match used by CLIP and BLIP with a comparison between the original prompt and a description of the image written by a multimodal large language model, since AI images often depict subjects those cross-modal models never saw in training. For the second, it pools visual features with weights derived from the human contrast sensitivity function, so small artifacts in high-frequency detail change the score. If the claim is right, one trained model can rank AI images against human opinion at 0.9051-0.9113 SRCC and 0.9158-0.9558 PLCC on three public benchmarks, improving on every previous method it compares against.

What carries the argument

At the core is the two-branch decomposition plus a frequency-weighted pooling rule. TSAM uses the MLLM to produce a "descriptive prompt" that stands in for the image, then cross-attends its BLIP embeddings against the original prompt's embeddings to yield the Semantic Consistency Information. FFDPM derives a patch-level weight $1+W_h$ from the Contrast Sensitivity Function applied to the FFT of image patches, and multiplies these HVS weights with learned spatial and channel weights before pooling ViT features into the Visual Quality Information. The MoER head, four experts selected by a gating network, fuses the two information vectors into the final score, trained end-to-end with smooth L1 loss.

What would settle it

Evaluate the trained model on a separately held-out split chosen with several different random seeds, or on images from a generator that appears in none of AGIQA-1K, AGIQA-3K, or AIGCIQA2023; if the SRCC/PLCC margins over MoE-AGIQA shrink below the 0.67-6.36 percentage-point gains reported, the original advantage came mainly from validation-set selection rather than the new modules.

Watch

Extended reading notes

Core claim

SC-AGIQA's central claim is that a generated image's quality decomposes into semantic consistency and perceptual distortion, and that each needs its own machinery. The semantic branch translates the image into a textual description with an MLLM and measures agreement between that description and the original prompt in a BLIP embedding space, avoiding the direct image-text alignment that CLIP and BLIP get wrong on AI content. The perceptual branch weights Vision Transformer patch features by learned importance multiplied by a contrast-sensitivity response to each patch's frequency content, so the final visual-quality vector is sensitive to subtle artifacts. A Mixture-of-Experts regressor combines the two vectors, and the paper reports that the full pipeline sets new highs of 0.9051/0.9558 on AGIQA-1K, 0.9070/0.9361 on AGIQA-3K, and 0.9113/0.9158 on AIGCIQA2023.

Load-bearing premise

The load-bearing premise is that the fixed 20% validation split used for early stopping and model selection fairly represents the population of future AI-generated images, since the reported numbers are computed on that same split and the paper reports a single fixed random seed.

Editorial extensions

If this is right

  • Separating semantic and perceptual checks means improvements in one branch do not need to distort the other; the ablations show removing either branch reduces correlation with human scores.
  • MLLM-generated descriptions can act as a stable textual proxy for image content, so later quality models can compare two texts instead of aligning an image with a text, a much easier task for cross-modal encoders.
  • Frequency-weighted pooling gives an explicit place where HVS priors enter the score, so distortions that are invisible to semantic alignment can still change the outcome.
  • The MoER head allows different quality regimes to weight alignment and detail differently, and the paper's consistent gains on all three datasets suggest this regime-specialization helps beyond any single-regression head.

Reading between the lines

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

  • A natural next experiment the paper does not run is cross-dataset transfer: train on two benchmarks and test on the third, or test on a newer generator; the fixed-seed single split leaves this open.
  • The redundancy the authors flag in their limitation section suggests a fix: feed the original prompt to the MLLM when writing the description so aesthetic over-description can be suppressed, while guarding against hallucinated prompt details.
  • The HVS frequency weighting should transfer to ordinary no-reference image quality assessment, where the same contrast-sensitivity prior applies, so the method is a candidate drop-in for non-AI distortions too.
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 / 6 minor

Summary. The paper proposes SC-AGIQA, a framework for AI-generated image quality assessment (AGIQA) that combines two modules: a Text-assisted Semantic Alignment Module (TSAM), which uses an MLLM (Doubao v1.5) to generate a descriptive prompt from the image and compares it with the original prompt via a BLIP encoder, and a Frequency-domain Fine-Grained Degradation Perception Module (FFDPM), which pools ViT features using an adaptive, HVS-inspired frequency weighting scheme. The two representations are fused by a Mixture-of-Experts regressor. The method is evaluated on AGIQA-1K, AGIQA-3K, and AIGCIQA2023, where it is reported to outperform prior methods on SRCC, PLCC, and a proposed 'Main Score' metric, and ablation studies isolate the contribution of each module.

Significance. The central idea is well-motivated and the modular design is clear: using an MLLM to convert an image into a descriptive prompt for semantic comparison is a sensible way to bridge the distribution gap between natural-image pretraining and AI-generated content, and the HVS-inspired frequency weighting in AQAFP is a reasonable mechanism for improving detail perception. If the reported performance is confirmed under a proper evaluation protocol, SC-AGIQA would be a useful contribution to the AGIQA literature. The authors also release code, which aids reproducibility. The main weakness is the evaluation protocol: the headline numbers are computed on the validation split used for early stopping and model selection under a single random seed, which makes the claimed state-of-the-art results unverified as estimates of generalization. The missing baselines in some comparisons further weaken the cross-benchmark claim. These issues are load-bearing because the paper's central contribution is the reported superiority over existing methods, but they are fixable with additional experiments.

major comments (3)
  1. [§4.2–§4.3, Tables 1–3] The evaluation protocol uses the same 20% validation split for early stopping and model selection and then reports that split's SRCC and PLCC as the final results, under a single fixed random seed. Since the reported numbers are exactly the values used to select the checkpoint, they are optimistically biased estimates of generalization and do not constitute an independent test. The claimed improvements, e.g., +5.21 SRCC over MoE-AGIQA-v1 on AGIQA-1K where the validation set contains roughly 216 images, may fall within split-to-split variation. The paper should either use a held-out test split or perform repeated random splits (with different seeds) and report mean and standard deviation, so that the state-of-the-art claim is supported by unbiased estimates.
  2. [Tables 1 and 3, and §4.3] The comparison against state-of-the-art methods is incomplete: MA-AGIQA, a strong MLLM-based baseline, appears only in Table 2 (AGIQA-3K) and is omitted from Tables 1 and 3, and no MLLM-based method is compared on AIGCIQA2023. Since the paper explicitly claims superiority across all three benchmarks, the absence of these baselines (or an explicit justification for their omission) leaves the cross-dataset state-of-the-art claim under-supported.
  3. [Table 4, §4.4] The ablation study in Table 4 reports only point estimates with no error bars or significance tests. Several differences are small, for example on AIGCIQA2023 the full model's SRCC (0.9113) is only 0.54 points above the configuration without AQAFP (0.9016), and the configuration with AQAFP but without HP (0.9016) is only 0.24 points above the configuration with neither (0.8992). Without repeated runs or a significance test, the claimed contribution of each module is not established, and the same single-split issue that affects Tables 1–3 applies here.
minor comments (6)
  1. [§4.4] The ablation table uses 'SCI' to denote the semantic consistency information, while the module that produces it is called TSAM in Section 3.2; please align the notation for clarity.
  2. [Figure 4 caption] The caption refers to 'green text' to highlight matching concepts, but the figure is likely to be printed in grayscale; please use a marker that survives grayscale printing, such as underlining or bold.
  3. [Eq. (11)] The smooth L1 loss is written with the quadratic term 0.5(x)^2 without normalizing by beta; the standard formulation is 0.5*(x^2)/beta for |x|<beta. Since beta is fixed to 1.0 throughout the paper this is not a numerical error, but the formula should be corrected for generality.
  4. [§2.1 and references] Reference [14] is listed nine times in the related work section, and the arXiv identifier for Q-Instruct is mistyped as 'arXiv preprint arxiv:2010:11929' instead of 'arXiv:2310.11235'; please fix these presentation issues.
  5. [§4.1] The paper states that 'a fixed random seed' is used for reproducibility, but the seed value is not reported; please include the seed value or a link to the code that specifies it.
  6. [§4.2] The MLLM component uses the Doubao v1.5 commercial API, which may be unavailable or change over time; please provide the exact checkpoint name and access date to improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SC-AGIQA is a supervised regression on external MOS benchmarks; the headline SOTA claim rests on an evaluation-protocol generalization concern, not on a definitional or self-citational reduction.

full rationale

SC-AGIQA is a supervised regression model trained end-to-end with Smooth L1 loss against human MOS. The derivation chain consists of three constructed modules: TSAM compares an MLLM-generated description P_d with the original prompt P_o via a BLIP/ImageReward encoder and cross-attention; FFDPM pools ViT features with data-driven and CSF-based HVS weights; MoER regresses the concatenated SCI and VQI vectors to a score. None of these components defines the predicted score in terms of the ground-truth MOS except through the standard training loss, and the headline SRCC/PLCC values are measured on the 20% validation partition against human labels, not manufactured from fitted parameters. The borrowings from MoE-AGIQA and ImageReward are explicitly attributed and externally trained; ImageReward is an independently published preference model, and MoE-AGIQA is an independent baseline, so citing them is not load-bearing self-citation. The paper's own limitation discussion about P_d redundancy is a design caveat, not a circular step. The main risk is methodological: the fixed 80/20 split with early stopping on the validation set means the reported numbers are the same numbers used for model selection, so they may be optimistically biased and should not be read as hold-out generalization evidence. That is a generalization/overfitting concern, not a circularity of the derivation; no equation or fitted parameter is equivalent by construction to the claimed result.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; the architecture components (TSAM, FFDPM, AQAFP, MoER) are methods, not entities. The free parameters are standard hyperparameters plus the unstated random seed, and the axioms concern the reliability of the pretrained components and the evaluation protocol. The most consequential assumption is that the validation split used for tuning is also a valid final test set.

free parameters (6)
  • Random data split seed = not reported
    The 80/20 partition is generated with a fixed random seed; the choice of seed determines the entire evaluation, and only one seed is used.
  • Smooth L1 loss beta = 1.0
    Set by hand; standard choice for smooth L1 and not ablated.
  • Number of MoER experts = 4
    Chosen without ablation; affects model capacity and gating behavior.
  • Top-K experts in MoER = 3
    Chosen by hand; dynamic expert selection depends on this value.
  • Number of inference crops = 15
    Predictions are averaged over 15 random crops at inference; this choice affects stability and final scores.
  • Learning rate schedule = 1e-5 base, warmup 3 epochs, decay 0.1 every 3 epochs
    Tuned on the same validation split whose scores are then reported as results.
assumptions (5)
  • domain assumption The MLLM (Doubao v1.5) generates P_d that accurately reflects the image's content and aesthetics without omitting key details or hallucinating.
    Eq. (2) in Section 3.2; the TSAM's semantic comparison is only as valid as this description.
  • domain assumption BLIP/ImageReward encoder embeddings, combined through cross-attention, produce a similarity measure between P_o and P_d that correlates with human text-image alignment judgments.
    Section 3.2, Eq. (3)-(4); the S_CI vector is the only semantic signal used.
  • domain assumption The Mannos-Sakrison contrast sensitivity function correctly models human sensitivity to the distortion types found in AI-generated images.
    Section 3.3, Eq. (8); CSF weights are applied to FFT patch magnitudes.
  • domain assumption The mean opinion scores in AGIQA-1K, AGIQA-3K, and AIGCIQA2023 are faithful and consistent ground truth for image quality and text-image alignment.
    Section 4.1; all training and evaluation use these labels without cross-dataset calibration.
  • ad hoc to paper The single 20% validation split used for early stopping and model selection is representative of future unseen AI-generated images.
    Sections 4.1-4.2; no separate test set is held out, so the general claim rests on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-Visual Semantic Constrained AI-Generated Image Quality Assessment." pith.science (2026). https://pith.science/paper/ZZR374PL

@misc{pith2026250710432,
  author       = {Pith},
  title        = {Pith review of: Text-Visual Semantic Constrained AI-Generated Image Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZR374PL}},
  note         = {Machine review of arXiv:2507.10432}
}
read the original abstract

With the rapid advancements in Artificial Intelligence Generated Image (AGI) technology, the accurate assessment of their quality has become an increasingly vital requirement. Prevailing methods typically rely on cross-modal models like CLIP or BLIP to evaluate text-image alignment and visual quality. However, when applied to AGIs, these methods encounter two primary challenges: semantic misalignment and details perception missing. To address these limitations, we propose Text-Visual Semantic Constrained AI-Generated Image Quality Assessment (SC-AGIQA), a unified framework that leverages text-visual semantic constraints to significantly enhance the comprehensive evaluation of both text-image consistency and perceptual distortion in AI-generated images. Our approach integrates key capabilities from multiple models and tackles the aforementioned challenges by introducing two core modules: the Text-assisted Semantic Alignment Module (TSAM), which leverages Multimodal Large Language Models (MLLMs) to bridge the semantic gap by generating an image description and comparing it against the original prompt for a refined consistency check, and the Frequency-domain Fine-Grained Degradation Perception Module (FFDPM), which draws inspiration from Human Visual System (HVS) properties by employing frequency domain analysis combined with perceptual sensitivity weighting to better quantify subtle visual distortions and enhance the capture of fine-grained visual quality details in images. Extensive experiments conducted on multiple benchmark datasets demonstrate that SC-AGIQA outperforms existing state-of-the-art methods. The code is publicly available at https://github.com/mozhu1/SC-AGIQA.

Figures

Figures reproduced from arXiv: 2507.10432 by the authors.

Figure 1
Figure 1. (a) Comparing natural images and AGIs; note AGIs can depict non-existent subjects (purple text examples). (b) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed SC-AGIQA framework. (A) The Text-assisted Semantic Alignment Module (TSAM) leverages [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The principle of the Adaptive Quality-Aware Fea [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The 𝑃𝑜 and 𝑃𝑑 of images corresponding to different text-image alignment scores. As the text - image alignment scores increase, the semantic consistency between 𝑃𝑜 and 𝑃𝑑 becomes more pronounced. (green) baselines. We obtain scores of 0.9070 (SRCC), 0.9361 (PLCC), and 0…
Figure 5
Figure 5. Figure 5: HVS-based pooling weights (𝑊𝑝 ) generated within the FFDPM’s AQAFP module. Calculated as 𝑊𝑝 = 1 + CSF(FFT(𝐼𝑐,patches)), these weights (visualized below each cropped image 𝐼𝑐 ) prioritize perceptually salient frequency in￾formation for fine-grained visual quality assess…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 37 canonical work pages

  1. [1]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer, ”IEEE Signal Processing Letters , vol. 20, no. 3, pp. 209–212, 2012

  2. [2]

    Blind image quality assessment: From natural scene statistics to perceptual quality,

    A. K. Moorthy and A. C. Bovik, “Blind image quality assessment: From natural scene statistics to perceptual quality, ” IEEE Transactions on Image Processing , vol. 20, no. 12, pp. 3350–3364, 2011

  3. [3]

    No-reference image quality assessment in the spatial domain,

    A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain, ”IEEE Transactions on Image Processing , vol. 21, no. 12, pp. 4695–4708, 2012

  4. [4]

    A feature-enriched completely blind image quality evaluator,

    L. Zhang, L. Zhang, and A. C. Bovik, “A feature-enriched completely blind image quality evaluator, ”IEEE Transactions on Image Processing , vol. 24, no. 12, pp. 5357–5370, 2015

  5. [5]

    Blind image quality assessment using a deep bilinear convolutional neural network,

    W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “Blind image quality assessment using a deep bilinear convolutional neural network, ”IEEE Transactions on Circuits and Systems for Video Technology , vol. 30, no. 1, pp. 36–47, 2020

  6. [6]

    Arniqa: Learning distortion manifold for image quality assessment,

    L. Agnolucci, L. Galteri, M. Bertini, and A. Del Bimbo, “Arniqa: Learning distortion manifold for image quality assessment, ” inProceedings of the Winter Conference on Applications of Computer Vision . IEEE, 2024, pp. 189–198

  7. [7]

    A multiscale approach to deep blind image quality assessment,

    M. Liu, J. Huang, D. Zeng, X. Ding, and J. Paisley, “A multiscale approach to deep blind image quality assessment, ”IEEE Transactions on Image Processing , vol. 32, pp. 1656–1669, 2023

  8. [8]

    Topiq: A top-down approach from semantics to distortions for image quality assessment,

    C. Chen, J. Mo, J. Hou, H. Wu, L. Liao, W. Sun, Q. Yan, and W. Lin, “Topiq: A top-down approach from semantics to distortions for image quality assessment, ” IEEE Transactions on Image Processing , vol. 33, pp. 2404–2418, 2024

Show all 38 references
  1. [9]

    Blindly assess image quality in the wild guided by a self-adaptive hyper network,

    S. Su, Q. Yan, Y. Zhu, C. Zhang, X. Ge, J. Sun, and Y. Zhang, “Blindly assess image quality in the wild guided by a self-adaptive hyper network, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 3667–3676

  2. [10]

    Data-efficient image quality assessment with attention-panel decoder,

    G. Qin, R. Hu, Y. Liu, X. Zheng, H. Liu, X. Li, and Y. Zhang, “Data-efficient image quality assessment with attention-panel decoder, ” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 2, 2023, pp. 2091–2100

  3. [11]

    Exploring clip for assessing the look and feel of images,

    J. Wang, K. C. Chan, and C. C. Loy, “Exploring clip for assessing the look and feel of images, ” inProceedings of the AAAI Conference on Artificial Intelligence , vol. 37, no. 2, 2023, pp. 2555–2563

  4. [12]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

    W. Zhang, G. Zhai, Y. Wei, X. Yang, and K. Ma, “Blind image quality assessment via vision-language correspondence: A multitask learning perspective, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 14 071–14 081

  5. [13]

    Q-bench: A benchmark for multi-modal foundation models on low-level vision from single images to pairs,

    Z. Zhang, H. Wu, E. Zhang, G. Zhai, and W. Lin, “Q-bench: A benchmark for multi-modal foundation models on low-level vision from single images to pairs, ” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, pp. 10 404– 10 418, 2024

  6. [14]

    Q-instruct: Improving low-level visual abilities for multi-modality foundation models,

    H. Wu, Z. Zhang, E. Zhang, C. Chen, L. Liao, A. Wang, K. Xu, C. Li, J. Hou, G. Zhai, G. Xue, W. Sun, Q. Yan, and W. Lin, “Q-instruct: Improving low-level visual abilities for multi-modality foundation models, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pa...

  7. [15]

    Clip-agiqa: Boosting the performance of ai-generated image quality assessment with clip,

    Z. Tang, Z. Wang, B. Peng, and J. Dong, “Clip-agiqa: Boosting the performance of ai-generated image quality assessment with clip, ” inProceedings of the Interna- tional Conference on Pattern Recognition . Springer, 2025, pp. 48–61

  8. [16]

    Moe-agiqa: Mixture-of- experts boosted visual perception-driven and semantic-aware quality assessment for ai-generated images,

    Y. Yang, J. Fu, W. Zhang, W. Cao, L. Liu, and H. Peng, “Moe-agiqa: Mixture-of- experts boosted visual perception-driven and semantic-aware quality assessment for ai-generated images, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop...

  9. [17]

    Sf-iqa: Quality and similarity integra- tion for ai generated image quality assessment,

    Z. Yu, F. Guan, Y. Lu, X. Li, and Z. Chen, “Sf-iqa: Quality and similarity integra- tion for ai generated image quality assessment, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 6692–6701

  10. [18]

    Bringing textual prompt to ai-generated image quality assessment,

    Z. Tang, Z. Wang, B. Peng, and J. Dong, “Bringing textual prompt to ai-generated image quality assessment, ” inProceedings of the IEEE International Conference on Multimedia and Expo. IEEE, 2024, pp. 1–6

  11. [19]

    Ai-generated image quality assessment based on task-specific prompt and multi-granularity similarity,

    J. Xia, L. He, F. Gao, K. Zhang, L. Li, and X. Gao, “Ai-generated image quality assessment based on task-specific prompt and multi-granularity similarity, ”arXiv preprint arXiv:2411.16087, 2024

  12. [20]

    Large multi-modality model assisted ai-generated image quality assessment,

    P. Wang, W. Sun, Z. Zhang, J. Jia, Y. Jiang, Z. Zhang, X. Min, and G. Zhai, “Large multi-modality model assisted ai-generated image quality assessment, ” inPro- ceedings of the ACM International Conference on Multimedia , 2024, pp. 7803–7812

  13. [21]

    Tier: Text-image encoder-based regression for aigc image quality assessment,

    J. Yuan, X. Cao, J. Che, Q. Wang, S. Liang, W. Ren, J. Lin, and X. Cao, “Tier: Text-image encoder-based regression for aigc image quality assessment, ”arXiv preprint arXiv:2401.03854, 2024

  14. [22]

    Aigc image quality assessment via image-prompt correspondence,

    F. Peng, H. Fu, A. Ming, C. Wang, H. Ma, S. He, Z. Dou, and S. Chen, “Aigc image quality assessment via image-prompt correspondence, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , 2024, pp. 6432–6441

  15. [23]

    Quality assessment for text-to-image generation: A survey,

    Y. Tian, Y. Liu, S. Wang, and S. Kwong, “Quality assessment for text-to-image generation: A survey, ”IEEE MultiMedia, 2025

  16. [24]

    Imagereward: Learning and evaluating human preferences for text-to-image generation,

    J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong, “Imagereward: Learning and evaluating human preferences for text-to-image generation, ” in Proceedings of the Advances in Neural Information Processing Systems , vol. 36, 2023, pp. 15 903–15 935

  17. [25]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation,

    Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy, “Pick-a-pic: An open dataset of user preferences for text-to-image generation, ” inProceedings of the Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 36 652– 36 663

  18. [26]

    Learning trans- ferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning trans- ferable visual models from natural language supervision, ” inProceedings of the International Conference on Machine Lear...

  19. [27]

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

    J. Li, D. Li, C. Xiong, and S. C. H. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, ” inPro- ceedings of the International Conference on Machine Learning . PMLR, 2022, pp. 12 888–12 900

  20. [28]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale, ”arXiv preprint arxiv:2010:11929, 2020

  21. [29]

    Imagenet: A large- scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large- scale hierarchical image database, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . IEEE, 2009, pp. 248–255

  22. [30]

    Aigciqa2023: A large-scale image quality assessment database for ai generated images: from the perspec- tives of quality, authenticity and correspondence,

    J. Wang, H. Duan, J. Liu, S. Chen, X. Min, and G. Zhai, “Aigciqa2023: A large-scale image quality assessment database for ai generated images: from the perspec- tives of quality, authenticity and correspondence, ” inProceedings of the CAAI International Conference on Artificia...

  23. [31]

    A perceptual quality assessment exploration for aigc images,

    Z. Zhang, C. Li, W. Sun, X. Liu, X. Min, and G. Zhai, “A perceptual quality assessment exploration for aigc images, ” inProceedings of the IEEE International Conference on Multimedia and Expo Workshops . IEEE, 2023, pp. 440–445

  24. [32]

    Agiqa-3k: An open database for ai-generated image quality assessment,

    C. Li, Z. Zhang, H. Wu, W. Sun, X. Min, X. Liu, G. Zhai, and W. Lin, “Agiqa-3k: An open database for ai-generated image quality assessment, ”IEEE Transactions on Circuits and Systems for Video Technology , vol. 34, no. 8, pp. 6833–6846, 2023

  25. [33]

    Aigiqa-20k: A large database for ai-generated image quality assessment,

    C. Li, T. Kou, Y. Gao, Y. Cao, W. Sun, Z. Zhang, Y. Zhou, Z. Zhang, W. Zhang, H. Wu, X. Liu, X. Min, and G. Zhai, “Aigiqa-20k: A large database for ai-generated image quality assessment, ” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 20...

  26. [34]

    Laion-5b: An open large-scale dataset for training next generation image-text models,

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, A. Kolesnikov, B. Schiele, S. Gelly, and J. Jitsev, “Laion-5b: An open large-scale dataset for training next generation image-text models, ” in Proceedings of the Advances in Neural Information Processing ...

  27. [35]

    Microsoft coco: Common objects in context,

    T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Dollar, “Microsoft coco: Common objects in context, ” inProceedings of the European Conference on Computer Vision . Springer International Publishing, 2014, pp. 740–755

  28. [36]

    The effects of a visual fidelity criterion of the encoding of images,

    J. Mannos and D. J. Sakrison, “The effects of a visual fidelity criterion of the encoding of images, ”IEEE Transactions on Information Theory , vol. 20, no. 4, pp. 525–536, 1974

  29. [37]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models, ” inPro- ceedings of the Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 6840–6851

  30. [38]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets, ” inProceedings of the Advances in Neural Information Processing Systems , vol. 27, 2014

Pith tools

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