Pith. sign in

REVIEW 3 major objections 4 minor 64 references

Low-Biased General Annotated Dataset Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a diffusion model fine-tuned only on ImageNet class names can generate a synthetic pre-training dataset whose images carry less dataset bias than real ImageNet, improving downstream transfer accuracy for both ResNet…

desk verdict A serious empirical paper on synthetic low-bias dataset generation with a novel bi-level alignment, but the core mechanism is under-verified and the gains are modest without error bars. read the letter →

arxiv 2412.10831 v3 pith:H4LNVNRZ submitted 2024-12-14 cs.CV

Dengyang Jiang , Haoyu Wang , Lei Zhang , Wei Wei , Guang Dai , Mengmeng Wang , Jingdong Wang , Yanning Zhang This is my paper · ORCID
classification cs.CV
keywords low-biaseddatasetgenerationbiassyntheticImageNetdiffusionmodelfine-tuningCLIPsemanticalignmentbackbonepre-trainingtransferlearningqualityassuranceloss
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 a general pre-training dataset for vision backbones can be generated synthetically with less dataset bias than a manually collected one, and that this makes downstream models generalize better. The proposed framework, lbGen, fine-tunes a diffusion model using only the 1000 ImageNet class names as input, pulling generated images into a semantic space defined by CLIP text features and keeping them realistic with a quality-scoring loss. Pre-training ResNet50 and ViT-S on the resulting images improves average transfer accuracy across eight datasets by 1.7% and 2.1% over the same backbones pre-trained on real ImageNet, and the advantage grows when downstream training data is scarce. If the claim holds, it offers a cheap, scalable route to bias-controlled pre-training data without manual collection.

What carries the argument

The key machinery is a bi-level semantic alignment loss built on CLIP's shared image-text space. The first level is an adversarial loss: a small discriminator learns to tell apart CLIP embeddings of generated images from CLIP text embeddings of the 1000 class names, and the diffusion generator is trained to fool it, aligning the whole dataset to the class-name text distribution. The second level is a per-image cosine-similarity loss between each generated image and the phrase 'photo of <class>'. A quality assurance loss from the Q-ALIGN scorer keeps images faithful. These losses fine-tune Stable Diffusion (SD1.5) through reward-style gradients with LoRA, using only the class names as inputs.

What would settle it

One decisive check: train a linear probe to separate CLIP image features of the generated images from CLIP text features of the class names; if the probe succeeds at high accuracy, the adversarial loss has closed the modality gap rather than transferring semantic structure, and the low-bias explanation is unsupported. A second check: evaluate lbGen-pretrained backbones on a domain shift far outside the training categories (for example, medical or satellite imagery); if the transfer advantage over real-ImageNet pre-training disappears, the claimed debiasing is not general.

Watch

Extended reading notes

Core claim

The central discovery is that the bias in a manually collected dataset is not inseparable from its annotations: a diffusion model can be steered to generate a class-annotated dataset whose images are less tied to collection-specific backgrounds, styles, and object positions. The paper defines 'low-biased' through the CLIP text-feature distribution of the target dataset's class names, and uses a bi-level semantic alignment loss to enforce that the whole generated dataset matches that distribution adversarially while each image matches its own class description, with an additional quality assurance loss to prevent image degradation. The result is a low-biased ImageNet generated from scratch, and models pre-trained on it show stable generalization gains on transfer learning, detection, segmentation, and specific bias benchmarks, with the largest gains where manually labeled target data is scarce.

Load-bearing premise

The load-bearing premise is that CLIP's text-feature distribution of the 1000 ImageNet class names is a truly low-bias semantic space, so that adversarially aligning generated images to it removes dataset bias rather than merely closing the image-text modality gap.

Editorial extensions

If this is right

  • Pre-training on lbGen-generated data yields higher average transfer accuracy than real ImageNet: +1.7% for ResNet50 and +2.1% for ViT-S across eight transfer learning datasets.
  • The generalization advantage widens when downstream data is scarce, such as with only 20% of COCO training images, where lbGen pre-training gains 1.54% APbox over IN-Real, and similar gains appear on ADE20K.
  • Backbones pre-trained on lbGen show lower texture bias, context bias, and background bias on Cue Conflict, FOCUS, and Mixed-Rand/Mixed-Same benchmarks, indicating they rely less on shortcut features.
  • Because the generator is fine-tuned using only category names, the pipeline avoids manual image collection and image-text pairs, producing an ImageNet-scale dataset in about 52 hours on 8 A100 GPUs.

Reading between the lines

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

  • If CLIP's text space is genuinely low-bias, the same recipe could synthesize pre-training datasets for arbitrary category sets by supplying only their names, including categories absent from existing image collections.
  • The method's success is entangled with the quality assurance loss: dropping Lq degrades both image quality and transfer accuracy, suggesting that pure semantic alignment may push the generator toward degenerate images and that the debiasing effect partly depends on keeping images realistic.
  • The polysemy failure mode (e.g., 'crane' generating both birds and machines) shows that a single class name can define a multi-modal target in CLIP space; more specific descriptions would help but risk reintroducing human bias, so the core premise that text is low-bias needs validation per vocabulary.
  • A testable extension is to check whether the adversarial alignment matches image features to text features semantically or merely closes the image-text modality gap; if the latter, the gains might come from a style shift rather than from true debiasing.
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

3 major / 4 minor

Summary. The paper proposes lbGen, a framework for generating a synthetic ImageNet dataset that is claimed to have lower dataset bias than real ImageNet. The generator fine-tunes Stable Diffusion using only class names as input, with a bi-level semantic alignment loss based on CLIP (an adversarial dataset-level alignment to the class-name text distribution and a per-image cosine alignment to the class description) plus a Q-ALIGN quality assurance loss. The resulting dataset is used to pretrain ResNet50 and ViT-S, which are then evaluated on eight transfer datasets, COCO detection, ADE20K segmentation, and three bias benchmarks. The paper reports consistent improvements over real ImageNet and two synthetic baselines, particularly in low-data regimes.

Significance. If the results hold, lbGen is a simple and practical recipe for producing low-bias pretraining data from only category names, and the paper contributes a useful evaluation protocol for synthetic dataset bias. The paper ships code, provides detailed training hyperparameters, and honestly lists limitations (polysemy and limited categories) in Appendix G. The main experimental design is strong in breadth: two backbones, two perception tasks, eight transfer sets, and three bias benchmarks. However, the central mechanism—that CLIP text features define a low-bias semantic space and that adversarial alignment to them removes bias—is not independently validated, and the reported gains are all single-run numbers without variance.

major comments (3)
  1. [§3.2, Eq. (5)] The load-bearing assumption that the CLIP text-feature distribution of the 1000 ImageNet class names is a low-biased semantic space, and that optimizing Eq. (5) moves image features into that space, is not validated. CLIP image and text embeddings are known to occupy separate regions (the modality gap), so the discriminator in Eq. (5) may learn to separate modalities rather than semantic content, and the generator can reduce L_ent by shifting features along the image-to-text direction without removing class-background or style correlations. The large IN-val degradation in Table 1 (ResNet50: 46.1 vs 76.2 for IN-Real) is consistent with this failure mode. I request a diagnostic: report the discriminator accuracy on held-out real ImageNet images versus text features, and measure the CLIP image-text modality gap before and after fine-tuning. If the discriminator can separate modalities with high accuracy even after training, the argument should be reformulated.
  2. [Tables 1–3] All reported numbers are from a single run, with no standard deviations or significance tests. The headline gains are modest (1.4 to 2.1 percentage points average transfer accuracy) and some individual datasets show losses (e.g., Flowers for ViT-S in Table 1). To support the claim of stable improvement, the authors should provide at least three seeds per condition and a significance test (e.g., paired t-test or Wilcoxon signed-rank test) for the main comparisons in Tables 1 and 2.
  3. [§4.3, Table 3] The bias benchmarks show that lbGen-pretrained models are less texture-biased, context-biased, and background-biased. However, these outcomes do not isolate the proposed mechanism. The ablation in Table 4 is on ImageNet-100 and does not include a control that tests whether the CLIP text space is actually low-bias; for example, fine-tuning the generator with the same losses but with a deliberately biased text distribution (e.g., class names plus spurious background words) should yield worse transfer and higher bias if the mechanism is correct. Without such a control, the improvements could come from generic distributional changes of the diffusion output (e.g., reduced texture or clutter) rather than from the semantic alignment.
minor comments (4)
  1. [Throughout] Several typos should be corrected: 'non-travail bias' (Abstract and Section 1) and 'non-trial bias' (Section 2.1) should be 'non-trivial bias'; also fix 'sup-optimal' (Section 2.1), 'Expect for' (Section 3.2), 'training object' (Section 3.2), and 'showd' (Table 7).
  2. [References] References [42] and [43] are the same paper (Rombach et al., "High-Resolution Image Synthesis with Latent Diffusion Models"); please merge them or cite distinct versions.
  3. [Appendix B] In Eq. (11), the notation 'S⨿' appears to be a typo for 'S_q'; the quality score should be denoted consistently with the surrounding text.
  4. [Figure 1] Figure 1 is not explicitly referenced in the body text; please add a callout in Section 4.2 or integrate the figure more closely with the discussion of transfer learning results.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the 'low-biased' label is tied to CLIP text-space alignment, but the load-bearing generalization claims are validated on external, CLIP-free benchmarks and do not reduce to the training objective.

full rationale

The paper's derivation chain is: take 1000 ImageNet class names, extract CLIP text features, treat that distribution as a low-biased semantic reference, fine-tune a diffusion model with a bi-level semantic alignment loss plus a quality-assurance loss, generate images, pre-train backbones on the generated dataset, and evaluate generalization on transfer, detection, segmentation, and specific-bias benchmarks. The central claim—that lbGen data improves downstream generalization and reduces measurable biases—is an empirical result established on external benchmarks (Cue Conflict, FOCUS, Mixed-Rand/Mixed-Same, COCO, ADE20K) that do not use CLIP or the training losses. The paper explicitly states the key premise as an assumption: 'we assume that the semantic space defined by language can be a low-biased representation.' Calling the CLIP text-feature distribution 'low-biased' and then aligning generated images to it is a definitional naming choice, not a circular derivation of the paper's actual contribution. The entire-alignment and individual-alignment losses (Eqs. 5 and 6) optimize closeness to CLIP text features, but the paper does not claim that this closeness itself is the evidence of low bias; rather, it uses external generalization and bias metrics as evidence. There are no fitted parameters renamed as predictions, no load-bearing self-citations, no uniqueness theorem imported from the authors' own prior work, and no known result merely renamed. The unvalidated assumptions about CLIP's text space being bias-free and about the image-text modality gap are legitimate correctness or robustness concerns, but they are not circularity under the stated criteria.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central 'low-biased' property rests on the assumption that CLIP's text space is bias-free and that adversarial alignment in that space transfers to real-world bias reduction. The paper provides external bias benchmarks, so the assumption is testable, but it is not independently established. No new physical entities are introduced; the losses and the discriminator are standard components.

free parameters (3)
  • lambda_1 (quality loss weight) = 0.1
    Scaling factor for the quality assurance loss in Eq. (8); chosen by hand, not fitted to the target result.
  • LoRA rank = 128
    Rank of low-rank adaptation for the UNet; a training hyperparameter that affects capacity but is not fitted to the claim.
  • gradient enable steps = 5
    Number of denoising steps with gradients enabled during fine-tuning, following Deep Reward and CoMat; an approximation that could affect the final generator.
assumptions (4)
  • domain assumption The CLIP text feature space of the ImageNet class names is a low-biased semantic representation.
    Used in Sec. 3.2 to define the target distribution for generated images; the paper assumes text is closer to ideal semantic information and free of the biases found in images.
  • domain assumption CLIP image and text features can be aligned across the modality gap by the adversarial discriminator.
    The entire-dataset alignment loss (Eq. 5) treats image features and text features as samples from the same distribution; the known CLIP modality gap is not addressed.
  • domain assumption Fine-tuning a pre-trained diffusion model with a REINFORCE-style gradient and partial gradient steps produces a valid low-bias generator.
    The gradient update in Eq. (4) follows prior work but is an approximation; the paper does not prove convergence or characterize the bias of the resulting generator.
  • domain assumption The standard pre-training and evaluation protocol is a faithful measure of generalization and bias.
    The paper uses linear probing, Mask R-CNN, UperNet, and bias benchmarks to measure the effect; this is standard practice but an assumption about how to quantify bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Low-Biased General Annotated Dataset Generation." pith.science (2026). https://pith.science/paper/H4LNVNRZ

@misc{pith2026241210831,
  author       = {Pith},
  title        = {Pith review of: Low-Biased General Annotated Dataset Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4LNVNRZ}},
  note         = {Machine review of arXiv:2412.10831}
}
read the original abstract

Pre-training backbone networks on a general annotated dataset (e.g., ImageNet) that comprises numerous manually collected images with category annotations has proven to be indispensable for enhancing the generalization capacity of downstream visual tasks. However, those manually collected images often exhibit bias, which is non-transferable across either categories or domains, thus causing the model's generalization capacity degeneration. To mitigate this problem, we present a low-biased general annotated dataset generation framework (lbGen). Instead of expensive manual collection, we aim at directly generating low-biased images with category annotations. To achieve this goal, we propose to leverage the advantage of a multimodal foundation model (e.g., CLIP), in terms of aligning images in a low-biased semantic space defined by language. Specifically, we develop a bi-level semantic alignment loss, which not only forces all generated images to be consistent with the semantic distribution of all categories belonging to the target dataset in an adversarial learning manner, but also requires each generated image to match the semantic description of its category name. In addition, we further cast an existing image quality scoring model into a quality assurance loss to preserve the quality of the generated image. By leveraging these two loss functions, we can obtain a low-biased image generation model by simply fine-tuning a pre-trained diffusion model using only all category names in the target dataset as input. Experimental results confirm that, compared with the manually labeled dataset or other synthetic datasets, the utilization of our generated low-biased dataset leads to stable generalization capacity enhancement of different backbone networks across various tasks, especially in tasks where the manually labeled samples are scarce.

Figures

Figures reproduced from arXiv: 2412.10831 by the authors.

Figure 1
Figure 1. Performance on eight transfer learning datasets. Pre-training on our generated low-biased general annotated dataset can bring stable generalization capacity enhancement of different backbones. Abstract Pre-training backbone networks on a general annotated dataset (e.g., ImageNet) that comprises numerous manually collected images with category annotations has proven to be indispensable for enhancing the generalizatio… view at source ↗
Figure 2
Figure 2. Visualization of some randomly sampled images from 4 datasets. It is hard to tell from which dataset exhibits low bias through these images. However, models trained on these four datasets demonstrate a significant disparity in their generalization capabilities. 1. Introduction Deep neural networks have achieved great success in vari￾ous computer vision tasks [34, 37, 62]. One indispensable premise of such success li… view at source ↗
Figure 3
Figure 3. Overview of our training method. The generator first generates an image according to the class name. Then the image is sent to bi-level semantic guidance module and quality assurance module respectively for loss calculation. 3. Approach The overall training framework of lbGen is shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Scaling down the number of training images of eight transfer learning datasets. The benefits of using pre-trained models on our lbGen images are even more pronounced when there is less data for training. 4.2. Generalization across Downstream Tasks Transfer Learning. Tr…
Figure 5
Figure 5. Figure 5: Impact of individual image alignmentloss. We observe that ambiguity problem between classes when discarding Lin. Effect of Bi-Level Semantic Alignment Loss. According to [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Visualization of generated images prompted by polysemy [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 38 canonical work pages

  1. [1]

    Leaving reality to imagi- nation: Robust classification via generated datasets

    Hritik Bansal and Aditya Grover. Leaving reality to imagi- nation: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 2, 3, 5, 6, 7, 12

    arXiv 2023

  2. [2]

    Food-101–mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, pro- ceedings, part VI 13, pages 446–461. Springer, 2014. 5, 13

    work page 2014

  3. [3]

    Enhancing Diffusion Models with Text-Encoder Reinforcement Learning

    Chaofeng Chen, Annan Wang, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Enhancing diffu- sion models with text-encoder reinforcement learning. arXiv preprint arXiv:2311.15657, 2023. 4

    work page Pith review arXiv 2023

  4. [4]

    Un- derstanding and mitigating the label noise in pre-training on downstream tasks

    Hao Chen, Jindong Wang, Ankit Shah, Ran Tao, Hongxin Wei, Xing Xie, Masashi Sugiyama, and Bhiksha Raj. Un- derstanding and mitigating the label noise in pre-training on downstream tasks. arXiv preprint arXiv:2309.17002, 2023. 2

    arXiv 2023

  5. [5]

    Mmdetection: Open mmlab detection toolbox and benchmark

    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. 12

    arXiv 1906

  6. [6]

    Describing textures in the wild

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606–3613, 2014. 5, 13

    work page 2014

  7. [7]

    Mmsegmentation: Open- mmlab semantic segmentation toolbox and benchmark, 2020

    MMSegmentation Contributors. Mmsegmentation: Open- mmlab semantic segmentation toolbox and benchmark, 2020. 12

    work page 2020

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 2, 3, 5, 6, 7, 13

    work page 2009

Show all 64 references
  1. [9]

    How specific is the shape bias? Child development, 74(1):168–178, 2003

    Gil Diesendruck and Paul Bloom. How specific is the shape bias? Child development, 74(1):168–178, 2003. 7

    2003

  2. [10]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5

    2010 arXiv

  3. [11]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim En- tezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. In Forty-first International Conference on Mach...

    2024

  4. [12]

    Can biases in imagenet mod- els explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 22184–22194, 2024

    Paul Gavrikov and Janis Keuper. Can biases in imagenet mod- els explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 22184–22194, 2024. 2, 7

    2024

  5. [13]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 5, 7, 13

    2018 arXiv

  6. [14]

    Shortcut learning in deep neural networks

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Fe- lix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665–673, 2020. 2

    2020

  7. [15]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 5

    2014

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

    2016

  9. [17]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Gir- shick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 2, 7

    2017

  10. [18]

    Introducing eurosat: A novel dataset and deep learning benchmark for land use and land cover classifi- cation

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing eurosat: A novel dataset and deep learning benchmark for land use and land cover classifi- cation. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 204–207. IEE...

    2018

  11. [19]

    The many faces of robust- ness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kada- vath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust- ness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF internation...

    2021

  12. [20]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Edward J 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.arXiv preprint arXiv:2106.09685, 2021. 5

    2021 arXiv

  13. [21]

    Comat: Aligning text-to-image diffusion model with image- to-text concept matching

    Dongzhi Jiang, Guanglu Song, Xiaoshi Wu, Renrui Zhang, Dazhong Shen, Zhuofan Zong, Yu Liu, and Hongsheng Li. Comat: Aligning text-to-image diffusion model with image- to-text concept matching. arXiv preprint arXiv:2404.03653,

    arXiv

  14. [22]

    Focus: Familiar objects in common and uncommon settings

    Priyatham Kattakinda and Soheil Feizi. Focus: Familiar objects in common and uncommon settings. In Interna- tional Conference on Machine Learning, pages 10825–10847. PMLR, 2022. 5, 13

    2022

  15. [23]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Pro- ceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2013. 5, 13

    2013

  16. [24]

    Im- agenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Im- agenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3

    2012

  17. [25]

    The illustrated image captioning using transformers

    A Kumar. The illustrated image captioning using transformers. Ankur-NLP Enthusiast, 2022. 3

    2022

  18. [26]

    Image captions are natural prompts for text-to-image models

    Shiye Lei, Hao Chen, Sen Zhang, Bo Zhao, and Dacheng Tao. Image captions are natural prompts for text-to-image models. arXiv preprint arXiv:2307.08526, 2023. 3

    2023 arXiv

  19. [27]

    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. In International conference on machine learning, pages 19730–19742. PMLR,

  20. [28]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings...

    2014

  21. [29]

    Pseudo numerical methods for diffusion models on manifolds

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 12

    2022

  22. [30]

    A decade’s battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632,

    Zhuang Liu and Kaiming He. A decade’s battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632,

    arXiv

  23. [31]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 7

    2021

  24. [32]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 11976–11986,

  25. [33]

    Fine-grained visual clas- sification of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual clas- sification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5, 13

    2013 arXiv

  26. [34]

    Review the state-of-the-art technologies of semantic segmentation based on deep learning

    Yujian Mo, Yan Wu, Xinneng Yang, Feilin Liu, and Yujun Liao. Review the state-of-the-art technologies of semantic segmentation based on deep learning. Neurocomputing, 493: 626–646, 2022. 2

    2022

  27. [35]

    Automated flower classification over a large number of classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722–729. IEEE, 2008. 5, 13

    2008

  28. [36]

    Cats and dogs

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498–3505. IEEE, 2012. 5, 13

    2012

  29. [37]

    Learning to learn, from transfer learning to domain adaptation: A unifying perspec- tive

    Novi Patricia and Barbara Caputo. Learning to learn, from transfer learning to domain adaptation: A unifying perspec- tive. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1442–1449, 2014. 2, 3, 6

    2014

  30. [38]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3

    2023 arXiv

  31. [39]

    Video diffusion align- ment via reward gradients

    Mihir Prabhudesai, Russell Mendonca, Zheyang Qin, Kate- rina Fragkiadaki, and Deepak Pathak. Video diffusion align- ment via reward gradients. arXiv preprint arXiv:2407.08737,

    arXiv

  32. [40]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  33. [41]

    Hierarchical text-conditional image genera- tion with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image genera- tion with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2

    2022 arXiv

  34. [42]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 4

    2022

  35. [43]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 4, 5, 6, 7, 8, 12

    2022

  36. [44]

    Pho- torealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Pho- torealistic text-to-image diffusion models with deep language understanding. Advances in neural informati...

    2022

  37. [45]

    No reason for no supervision: Im- proved generalization in supervised models

    Mert Bulent Sariyildiz, Yannis Kalantidis, Karteek Ala- hari, and Diane Larlus. No reason for no supervision: Im- proved generalization in supervised models. arXiv preprint arXiv:2206.15369, 2022. 12

    2022 arXiv

  38. [46]

    Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones

    Mert Bülent Sarıyıldız, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition , pages 8011–8...

    2023

  39. [47]

    Transnext: Robust foveal visual perception for vision transformers

    Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17773– 17783, 2024. 7

    2024

  40. [48]

    Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images

    Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Simone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

    2024

  41. [49]

    Object name learn- ing provides on-the-job training for attention

    Linda B Smith, Susan S Jones, Barbara Landau, Lisa Gershkoff-Stowe, and Larissa Samuelson. Object name learn- ing provides on-the-job training for attention. Psychological science, 13(1):13–19, 2002. 7

    2002

  42. [50]

    Contrastive multiview coding

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16, pages 776–794. Springer, 2020. 7, 8, 13

    2020

  43. [51]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521–1528. IEEE, 2011. 2, 3

    2011

  44. [52]

    Generalizing from a few examples: A survey on few-shot learning

    Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1–34, 2020. 6, 12

    2020

  45. [53]

    Q-align: Teaching lmms for vi- sual scoring via discrete text-defined levels

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for vi- sual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 5, 12

    2023 arXiv

  46. [54]

    Deep reward supervisions for tuning text-to-image diffusion models

    Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. arXiv preprint arXiv:2405.00760, 2024. 5

    2024 arXiv

  47. [55]

    Sun database: Large-scale scene recog- nition from abbey to zoo

    Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recog- nition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485–3492. IEEE, 2010. 5, 13

    2010

  48. [56]

    Noise or signal: The role of image backgrounds in object recognition

    Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. ArXiv preprint arXiv:2006.09994, 2020. 5, 13

    2006 arXiv

  49. [57]

    Unified perceptual parsing for scene understanding

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418–434, 2018. 2, 7

    2018

  50. [58]

    Attngan: Fine- grained text to image generation with attentional generative adversarial networks

    Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine- grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition, pag...

    2018

  51. [59]

    mplug- owl2: Revolutionizing multi-modal large language model with modality collaboration

    Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug- owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ,...

    2024

  52. [60]

    Real-fake: Effective training data synthesis through dis- tribution matching

    Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through dis- tribution matching. arXiv preprint arXiv:2310.10402, 2023. 2, 3, 5, 6, 7, 12

    2023 arXiv

  53. [61]

    Understanding bias in large-scale visual datasets

    Boya Zeng, Yida Yin, and Zhuang Liu. Understanding bias in large-scale visual datasets. Advances in Neural Information Processing Systems, 37:61839–61871, 2025. 2, 3

    2025

  54. [62]

    Object detection with deep learning: A review

    Zhong-Qiu Zhao, Peng Zheng, Shou-tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE transactions on neural networks and learning systems , 30 (11):3212–3232, 2019. 2

    2019

  55. [63]

    Semantic under- standing of scenes through the ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fi- dler, Adela Barriuso, and Antonio Torralba. Semantic under- standing of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302–321, 2019. 5, 13

    2019

  56. [64]

    classnames

    Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm- gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802–5810, 2019. 3 Appendix of lbGen A. Loss Computa...

    2019

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.