Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Text-to-Image GAN with Pretrained Representations

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read TIGER, a text-to-image GAN, reports FID 5.48 on COCO and 9.38 on CUB by feeding frozen CLIP and DINO features into its discriminator while generating images in about 0.04 seconds.

desk verdict TIGER has a genuinely new architecture idea, but the headline FID is a test-set-selected best-of-many number, so treat the SOTA claim as unproven until code or a validation split appears. read the letter →

arxiv 2501.00116 v1 pith:FWHV4U7P submitted 2024-12-30 cs.CV cs.AIcs.LG

Xiaozhou You , Jian Zhang This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords text-to-imagesynthesisgenerativeadversarialnetworkspretrainedvisionrepresentationsCLIPDINOFIDzero-shotgenerationhigh-capacityfusionblock
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to show that a GAN can still be the fastest and most practical architecture for text-to-image synthesis if its discriminator is powered by frozen, pretrained vision representations instead of features learned from scratch. To that end it introduces TIGER, which stacks a CLIP sub-discriminator and a DINO sub-discriminator inside a vision-empowered discriminator, and pairs them with a high-capacity generator built from fusion blocks that mix text conditions into image features at multiple depths. On the standard benchmarks the paper reports a new best FID of 5.48 on COCO and 9.38 on CUB, with inference at about 0.04 seconds per image, and on a zero-shot COCO evaluation it reports FID 11.96 while using far less training data and fewer parameters than LDM or Parti-350M. A sympathetic reader would care because this points to a way of getting diffusion-like quality from a model that is roughly two orders of magnitude faster at generation time.

What carries the argument

The load-bearing mechanism is the vision-empowered discriminator: a bank of frozen pretrained vision models, here CLIP and DINO, each feeding its own sub-discriminator through an adapter, with the sub-discriminator losses summed as the total adversarial loss. The companion object is the high-capacity fusion block (HFBlock), a generator module that stacks two deep fusion modules, each an affine text-conditioning layer plus a 3x3 convolution, and one global fusion module made of a 3x3 depth-wise convolution, a 5x5 dilated depth-wise convolution with dilation 3, and a 1x1 convolution. A semantic contrastive loss, computed as the CLIP cosine similarity between generated image and text, is added to the generator objective, and each sub-discriminator uses hinge loss with MA-GP, the matching-aware gradient penalty. Together these pieces give the generator strong text-image alignment signals and give the discriminator a scene-understanding prior that does not require training from scratch.

What would settle it

Re-run the TIGER pipeline with all architecture choices, including CLIP layers 2, 5, 9, DINO layers 1, 5, 9, adapter pair A/B, and exactly two sub-discriminators, selected on a held-out validation split of COCO or CUB, then evaluate once on the test set; if FID rises materially above 5.48 on COCO or 9.38 on CUB, the published advantage is partly a product of test-set selection.

Watch

Extended reading notes

Core claim

TIGER's central claim is that text-to-image GANs close the quality gap with diffusion and autoregressive models when the discriminator is built from multiple frozen pretrained vision encoders rather than trained from scratch. The vision-empowered discriminator sums adversarial losses from two sub-discriminators, one processing CLIP features and one processing DINO features, with lightweight adapters that map multi-level features into a common space before concatenating text features and scoring image-text alignment. The high-capacity generator pairs these signals with fusion blocks whose deep fusion modules inject sentence embeddings through affine layers and whose global fusion module mixes spatial and channel information with depth-wise, dilated, and 1x1 convolutions. The paper reports that this combination reaches FID 5.48 on COCO and 9.38 on CUB, improving on the GAN, diffusion, and autoregressive baselines in its comparison tables, and that in the zero-shot setting it reaches FID 11.96 on COCO with 0.34 billion parameters, 12 million training pairs, and 0.04-second inference.

Load-bearing premise

The headline scores rest on the assumption that choosing the model components by repeatedly measuring FID on the COCO test set, as reported in Section 4.3 and Figure 6, did not inflate the final numbers, because the same test set is then used for the state-of-the-art comparison.

Editorial extensions

If this is right

  • If TIGER's reported numbers hold, GANs remain competitive with diffusion and autoregressive models on standard text-to-image benchmarks while generating images in about 0.04 seconds, roughly two orders of magnitude faster than the large models in the comparison.
  • Frozen pretrained vision representations, used in combination, can supply the scene-understanding and generalization signals that a from-scratch GAN discriminator lacks, which is the paper's main architectural claim.
  • The HFBlock design offers a concrete alternative to attention-based fusion: affine text injection per channel plus depth-wise and dilated convolutions for spatial and long-range mixing, without the quadratic cost of attention.
  • On zero-shot COCO evaluation, a 0.34-billion-parameter GAN trained on 12 million image-text pairs lands close to LDM and Parti-350M in FID while using a small fraction of their data, parameters, and inference time.
  • The semantic contrastive loss is credited with raising R-precision from 0.329 to 0.348 in the ablation, indicating that CLIP-based text-image alignment is a distinct contributor beyond the adversarial loss.

Reading between the lines

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

  • Editorial inference: because the paper selects components on the COCO test set, the cleanest validation of TIGER would be an independent replication that chooses layers and adapters on a validation split; until then, the 5.48 FID should be read as a reported number under a specific selection protocol.
  • Editorial inference: the large weight gap between sub-discriminators, with CLIP weighted 1 and DINO weighted 0.001, suggests CLIP carries nearly all the adversarial signal and DINO may act mainly as a regularizer; the paper does not analyze this imbalance.
  • Editorial inference: the same frozen-representation discriminator recipe could transfer to other conditional generation settings, such as text-to-video or layout-to-image, where scene understanding from a pretrained bank would be equally useful.
  • Editorial inference: the zero-shot experiment evaluates on COCO after training on CC12M, so its claim is about cross-dataset transfer within a standard benchmark protocol rather than fully open-ended zero-shot generalization to unseen concepts.
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 TIGER, a text-to-image GAN whose discriminator integrates multiple frozen pretrained vision representations (CLIP and DINO) through sub-discriminators and adapters, paired with a generator built from high-capacity fusion blocks (HFBlock) that contain deep fusion modules and a global fusion module. Training uses a CLIP-based semantic contrastive loss (Eq. 1) plus adversarial losses. The paper reports state-of-the-art FID scores of 5.48 on COCO and 9.38 on CUB in the standard setting, and competitive zero-shot results against larger diffusion/autoregressive models with much faster inference. The main evidence consists of Tables 1 and 2, supported by Table 3 and Figure 6 ablations.

Significance. If the reported FID numbers were neutral evaluations of a fixed architecture, the paper would make a meaningful contribution: it would show that a GAN with frozen CLIP and DINO features in the discriminator can outperform strong baselines at a fraction of the inference cost, and it provides a concrete recipe (layer selections, adapter pairing, and stopping at two sub-discriminators) that others could adopt. The component ablations also document a large search space. However, the central claim is currently not supported as stated because the architecture was selected on the COCO test set (§4.3), making the headline FID a fitted statistic rather than a prediction, and the R-precision metric is partially circular because it uses CLIP, the same encoder optimized by Eq. (1). The paper reports no error bars, no code, and no validation split, so the reproducibility and statistical grounding of the central claim are incomplete.

major comments (3)
  1. [§4.3, Table 3, Figure 6] The final architecture is selected by lowest FID on the COCO test set, so the reported FID of 5.48 is a best-of-many statistic rather than a neutral evaluation of a fixed design. Section 4.3 states that experiments were "deployed on the COCO test set," and Figure 6 explicitly says "select the best performing model (lower FID on COCO test set) in each round." Table 3 shows that the choices are consequential: CLIP layer selections range from 5.48 to 10.42, DINO layer selections from 5.48 to 13.97, adapter order from 5.48 to 7.94, and pretrained model selections from 5.48 to 11.40. With no validation split, no seed/error bars, and no released code, the claimed state-of-the-art FID 5.48 may reflect selection on the evaluation set. The authors should either fix the architecture before touching the test set, or provide a validation split and report the performance of the chosen configuration with multiple seeds and error bars.
  2. [§3.4 Eq. (1) vs. §4 Evaluation Details] The R-precision metric is computed with CLIP cosine similarity between image and text, which is exactly what the semantic contrastive loss L_CLIP in Eq. (1) optimizes. Therefore the reported R-precision in Table 1 is a partially fitted quantity, not an independent measure of text-image alignment. This undermines the comparison of R-precision against baselines that do not use such a loss. To make the metric meaningful, the authors should either evaluate R-precision with a different encoder (e.g., a separately trained text-image retrieval model) or clearly disqualify R-precision as an independent metric in the central comparison.
  3. [§4.1 Table 1 and §4.3] The CUB result is less contaminated than COCO because the architecture was chosen on COCO, but it is still not independent evidence: the configuration is a test-selected one transferred to another dataset, and the paper reports no error bars or multiple seeds for CUB either. The claim "state-of-the-art performance on two challenging datasets" therefore rests on a single run of a test-selected model. The authors should report mean and standard deviation over at least three training runs for both datasets, and ideally show that the chosen configuration also performs well when selected on a held-out validation split.
minor comments (6)
  1. [§2 Related Works] There is a repeated sentence in the Introduction: "Among them, text-to-image synthesis is one of the most attractive areas" appears twice in the first paragraph.
  2. [§3.4 Eq. (3)] The text says "where λi is is the corresponding hyperparameter"—there is a duplicated "is".
  3. [Table 3] In the row "GFM w/o 3×3 DW-Conv, 1×1 Conv," the FID is written as "7,72" with a comma instead of a decimal point (should be 7.72).
  4. [§4.3] The sentence "we deploy our experiments on the COCO test set" uses "deploy" where "conduct" or "run" is intended; more importantly, this wording highlights the test-set selection issue raised in the major comments.
  5. [References] The reference to [Radford et al., 2015] in Eq. (1) seems mismatched: the CLIP model is introduced in [Radford et al., 2021], while [Radford et al., 2015] is the DCGAN paper. Please correct the citation in Section 3.4.
  6. [Availability] The paper states that "more experiments and analyses are conducted in the Supplementary Material," but no supplementary material or code link is provided in the arXiv submission. The authors should make the supplementary material available and, ideally, release code for reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

COCO FID 5.48 is a test-set-selected best-of-many result, and R-precision is scored with the same CLIP similarity that the model is trained to maximize.

  1. fitted input called prediction [Section 4.3 (Ablation Study), Figure 6 caption, and Table 3.]
    "to verify the superiority of each component in our proposed TIGER, we deploy our experiments on the COCO test set [Lin et al., 2014]. ... In the 1st round, we add the first pretrained representation ... The CLIP model performs best in the 1st round. In the 2nd round, we keep the selection from the previous round unchanged and add the second pretrained vision model ... Next, the DINO model is selected for our model. ... In the end, we chose the 2nd, 5th, 9th layer of CLIP and the 1st, 5th, 9th layer of DINO."

    The final architecture yielding the headline FID 5.48 (Table 1) is produced by selecting each component to minimize FID computed on the COCO test set, the same split used for the state-of-the-art comparison. Table 3 shows the selection is consequential: CLIP layer choices give FID 5.48 versus 6.85, 10.42, and 7.96; DINO choices give 5.48 versus 11.40, 9.47, and 13.97; and adapter order gives 5.48 versus 5.96, 6.84, and 7.94. With no validation split, no seed or error bars, and no released code, the reported 5.48 is the best of many test-set evaluations, not a neutral prediction of a fixed architecture. The state-of-the-art claim is therefore statistically forced by the selection procedure rather than independently verified.

  2. self definitional [Section 3.4, Equation (1); Section 4, Evaluation Details.]
    "we compute the cosine similarity between text features and image features extracted by the CLIP model ... LCLIP = Cos(I, T), (1) ... For R-precision, we use CLIP [Radford et al., 2021] to calculate the cosine similarity between original image and given description."

    The training loss directly maximizes LCLIP = Cos(I,T), the CLIP cosine similarity between the generated image and the text. Evaluation R-precision is then reported using the same CLIP encoder and the same cosine-similarity operation between image and text. A model optimized for this objective is being scored on the exact quantity it was trained to maximize, so the reported R-precision (0.348 on COCO, 0.341 on CUB) is partly a measure of training-objective fit rather than an independent text-image alignment metric. A different, non-training scorer would be needed for an unbiased comparison.

full rationale

The central quantitative claim of the paper is the COCO FID 5.48, and this number is not a neutral evaluation of a fixed architecture. Section 4.3 explicitly deploys the ablations on the COCO test set, and Figure 6 describes progressive selection of pretrained models by lower FID on that same test set; layer and adapter choices are likewise selected after measuring FID on the test set (Table 3). Consequently the reported SOTA FID is a best-of-selection statistic, and the CUB and zero-shot results inherit a configuration selected on COCO rather than independently fixed. The R-precision metric is additionally aligned with the CLIP training objective in Equation (1), so that secondary metric is partly self-scored. There are no load-bearing self-citations; the circularity is in the evaluation protocol rather than in the derivation of the architecture. Because the model itself is a genuinely novel combination of components, the paper retains independent methodological content, but the headline empirical claims are partially forced by construction and selection.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

No new physical or formal entities are introduced. The key hidden costs are tuned hyperparameters (lambda_CLIP, lambda_1, lambda_2, n), test-set-selected layer and adapter choices, and the use of CLIP features both as a training signal and as the evaluator of text-image alignment.

free parameters (7)
  • lambda_CLIP = 4
    Weight of the CLIP semantic contrastive loss (Section 3.4, Training Details).
  • lambda_1, lambda_2 = 1 and 0.001
    Weights for the two sub-discriminator losses (Section 3.4, Training Details).
  • Number of deep fusion modules per HFBlock = 2
    Set as default in Section 3.2, not ablated.
  • Selected CLIP layers = 2, 5, 9
    Chosen as best on COCO test set in Section 4.3, Table 3.
  • Selected DINO layers = 1, 5, 9
    Chosen as best on COCO test set in Section 4.3, Table 3.
  • Adapter pair = A for CLIP, B for DINO
    Selected by FID on COCO test set in Section 4.3, Table 3.
  • Number of sub-discriminators = 2
    Stopped at two because the third round failed to converge (Section 4.3).
assumptions (3)
  • domain assumption Frozen CLIP and DINO features are informative for discriminating real vs generated images in a GAN.
    Introduced in Section 3.3; not formally justified beyond prior work.
  • domain assumption FID computed with Inception V3 is a valid measure of image generation quality.
    Standard evaluation choice in Section 4, Evaluation Details.
  • domain assumption CLIP cosine similarity between image and text is a valid measure of text-image alignment.
    Used in Eq (1) as a training loss and in Evaluation Details as R-precision.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-to-Image GAN with Pretrained Representations." pith.science (2026). https://pith.science/paper/FWHV4U7P

@misc{pith2026250100116,
  author       = {Pith},
  title        = {Pith review of: Text-to-Image GAN with Pretrained Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWHV4U7P}},
  note         = {Machine review of arXiv:2501.00116}
}
read the original abstract

Generating desired images conditioned on given text descriptions has received lots of attention. Recently, diffusion models and autoregressive models have demonstrated their outstanding expressivity and gradually replaced GAN as the favored architectures for text-to-image synthesis. However, they still face some obstacles: slow inference speed and expensive training costs. To achieve more powerful and faster text-to-image synthesis under complex scenes, we propose TIGER, a text-to-image GAN with pretrained representations. To be specific, we propose a vision-empowered discriminator and a high-capacity generator. (i) The vision-empowered discriminator absorbs the complex scene understanding ability and the domain generalization ability from pretrained vision models to enhance model performance. Unlike previous works, we explore stacking multiple pretrained models in our discriminator to collect multiple different representations. (ii) The high-capacity generator aims to achieve effective text-image fusion while increasing the model capacity. The high-capacity generator consists of multiple novel high-capacity fusion blocks (HFBlock). And the HFBlock contains several deep fusion modules and a global fusion module, which play different roles to benefit our model. Extensive experiments demonstrate the outstanding performance of our proposed TIGER both on standard and zero-shot text-to-image synthesis tasks. On the standard text-to-image synthesis task, TIGER achieves state-of-the-art performance on two challenging datasets, which obtain a new FID 5.48 (COCO) and 9.38 (CUB). On the zero-shot text-to-image synthesis task, we achieve comparable performance with fewer model parameters, smaller training data size and faster inference speed. Additionally, more experiments and analyses are conducted in the Supplementary Material.

Figures

Figures reproduced from arXiv: 2501.00116 by the authors.

Figure 1
Figure 1. (a): Existing text-to-image GANs were trained from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The outstanding performance of our proposed TIGER. (a): On the standard text-to-image synthesis task, our TIGER achieves a new [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The architectures we investigate. (a): The high-capacity generator consists of several high-capacity fusion blocks to generate [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a): The architecture of sub discriminator in our vision-empowered discriminator. In our vision-empowered discriminator, each sub [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison between AttnGAN, DF-GAN, and our proposed TIGER conditioned on text descriptions from the test set [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Ablation experiments of different pretrained vision mod [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 29 canonical work pages

  1. [1]

    ediffi: Text-to-image diffusion models with an ensemble of expert denoisers

    [Balaji et al., 2022] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 ,

    arXiv 2022

  2. [5]

    Cogview: Mastering text-to-image generation via transformers

    [Ding et al., 2021] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), volume 34,

    work page 2021

  3. [6]

    Cogview2: Faster and better text-to-image generation via hierarchical transformers

    [Ding et al., 2022] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35:16890–16902,

    work page 2022

  4. [7]

    Vector quantized diffusion model for text- to-image synthesis

    [Gu et al., 2022] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text- to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696–10706,

    work page 2022

  5. [8]

    Visual at- tention network

    [Guo et al., 2022] Meng-Hao Guo, Cheng-Ze Lu, Zheng- Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual at- tention network. arXiv preprint arXiv:2202.09741,

    arXiv 2022

  6. [9]

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

    [Heusel et al., 2017] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochre- iter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), volume 30,

    work page 2017

  7. [11]

    Adam: A method for stochastic optimization

    [Kingma and Ba, 2015] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceed- ings of the International Conference on Learning Repre- sentations (ICLR),

    work page 2015

  8. [14]

    Text to image generation with semantic-spatial aware gan

    [Liao et al., 2022] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187–18196,

    work page 2022

Show all 40 references
  1. [15]

    Microsoft coco: Com- mon objects in context

    [Lin et al., 2014] Tsung-Yi Lin, Michael Maire, Serge Be- longie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Com- mon objects in context. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 740–755,

    2014

  2. [16]

    Swin transformer: Hierarchical vision transformer using shifted windows

    [Liu et al., 2021] 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 Inter- national Conference on Computer Vision (ICCV) , pages 1...

    2021

  3. [17]

    Glide: Towards photore- alistic image generation and editing with text-guided dif- fusion models

    [Nichol et al., 2021] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photore- alistic image generation and editing with text-guided dif- fusion models. arXiv preprint arXiv:2112.10741,

    2021 arXiv

  4. [18]

    Unsupervised representation learning with deep convolutional generative adversarial networks

    [Radford et al., 2015] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434,

    2015 arXiv

  5. [19]

    Learning transferable visual models from nat- ural language supervision

    [Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. In Proceedings of the In...

    2021

  6. [20]

    Zero-shot text-to-image generation

    [Ramesh et al., 2021] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the International Confer- ence on Machine Learning (ICML) , pages 8821–8831. PMLR,

    2021

  7. [21]

    Hierarchical text-conditional image generation with clip latents

    [Ramesh et al., 2022] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125,

    2022 arXiv

  8. [23]

    Dae- gan: Dynamic aspect-aware gan for text-to-image synthe- sis

    [Ruan et al., 2021] Shulan Ruan, Yong Zhang, Kun Zhang, Yanbo Fan, Fan Tang, Qi Liu, and Enhong Chen. Dae- gan: Dynamic aspect-aware gan for text-to-image synthe- sis. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV) , pages 13960–13969,

    2021

  9. [24]

    Improved techniques for training gans

    [Salimans et al., 2016] Tim Salimans, Ian Goodfellow, Woj- ciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), volume 29,

    2016

  10. [25]

    Projected gans converge faster

    [Sauer et al., 2021] Axel Sauer, Kashyap Chitta, Jens M¨uller, and Andreas Geiger. Projected gans converge faster. Advances in Neural Information Processing Sys- tems, 34:17480–17492,

    2021

  11. [26]

    Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis

    [Sauer et al., 2023] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. arXiv preprint arXiv:2301.09515,

    2023 arXiv

  12. [27]

    A u-net based discriminator for generative adversarial networks

    [Schonfeld et al., 2020] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8207–8216,

    2020

  13. [29]

    Rethinking the inception architecture for computer vision

    [Szegedy et al., 2016] Christian Szegedy, Vincent Van- houcke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 2818–2826,

    2016

  14. [30]

    Df-gan: Deep fusion generative adversarial networks for text-to- image synthesis

    [Tao et al., 2020] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to- image synthesis. arXiv preprint arXiv:2008.05865,

    2020 arXiv

  15. [31]

    Df-gan: A sim- ple and effective baseline for text-to-image synthesis

    [Tao et al., 2022] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A sim- ple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 16515– 16525,

    2022

  16. [32]

    Galip: Generative adversarial clips for text-to-image synthesis

    [Tao et al., 2023] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14214–14223,

    2023

  17. [33]

    The caltech- ucsd birds-200-2011 dataset

    [Wah et al., 2011] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech- ucsd birds-200-2011 dataset

    2011

  18. [36]

    Scaling autoregressive models for content-rich text-to-image generation

    [Yu et al., 2022] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasude- van, Alexander Ku, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5,

    2022 arXiv

  19. [37]

    Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial net- works

    [Zhang et al., 2017] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial net- works. In Proceedings of the IEEE International Confer- ence o...

    2017

  20. [38]

    Cross-modal contrastive learning for text-to-image generation

    [Zhang et al., 2021] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 833–842,

    2021

  21. [39]

    Towards language-free training for text-to-image generation

    [Zhou et al., 2022] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxi- ang Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Re...

    2022

  22. [40]

    Dm-gan: Dynamic memory generative adversar- ial networks for text-to-image synthesis

    [Zhu et al., 2019] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversar- ial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 5802–5810, 2019

    2019

  23. [2011]

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

    [Xu et al., 2018] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xi- aodong He. Attngan: Fine-grained text to image gener- ation with attentional generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Patte...

    2018

  24. [2014]

    Ensembling off-the-shelf models for gan training

    [Kumari et al., 2022] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10651–10662,

    2022

  25. [2015]

    Kingma and Max Welling

    [Kingma and Welling, 2014] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In Proceedings of the International Conference on Learning Representa- tions (ICLR),

    2014

  26. [2016]

    High-resolution image synthesis with latent diffusion models

    [Rombach et al., 2022] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684–10695,

    2022

  27. [2017]

    Scaling up gans for text-to-image syn- thesis

    [Kang et al., 2023] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image syn- thesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10124– 10134,

    2023

  28. [2018]

    Re- current affine transformation for text-to-image synthesis

    [Ye et al., 2022] Senmao Ye, Fei Liu, and Minkui Tan. Re- current affine transformation for text-to-image synthesis. arXiv preprint arXiv:2204.10482,

    2022 arXiv

  29. [2020]

    Very deep convolutional networks for large-scale image recognition

    [Simonyan and Zisserman, 2015] Karen Simonyan and An- drew Zisserman. Very deep convolutional networks for large-scale image recognition. In Proceedings of the Inter- national Conference on Learning Representations (ICLR),

    2015

  30. [2021]

    Muse: Text-to-image generation via masked generative transformers

    [Chang et al., 2023] Huiwen Chang, Han Zhang, Jarred Bar- ber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704,

    2023 arXiv

  31. [2022]

    Emerging properties in self-supervised vi- sion transformers

    [Caron et al., 2021] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vi- sion transformers. In Proceedings of the IEEE/CVF in- ternational conference on computer vision , pages...

    2021

  32. [2023]

    Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts

    [Changpinyo et al., 2021] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pag...

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.