Pith. sign in

REVIEW 4 major objections 6 minor 50 references

Efficiency without Compromise: CLIP-aided Text-to-Image GANs with Increased Diversity

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A GAN family called SCAD removes conditional mode collapse in CLIP-aided text-to-image generators, reaching StyleGAN-T-level zero-shot FID at a fraction of the training cost while increasing per-prompt diversity.

desk verdict A genuinely useful efficiency result for text-to-image GANs, but the diversity metric is under-specified and some headline numbers don't line up. read the letter →

arxiv 2506.01493 v1 pith:5TYEHFC7 submitted 2025-06-02 cs.CV cs.LG

Yuya Kobayashi , Yuhta Takida , Takashi Shibuya , Yuki Mitsufuji This is my paper
classification cs.CVcs.LG
keywords text-to-imagegenerationgenerativeadversarialnetworksslicedWassersteindistanceconditionalmodecollapseper-promptdiversityCLIPmutualinformationtrainingefficiency
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 claims that the efficiency gains of CLIP-aided text-to-image GANs come at a hidden price: conditional mode collapse, where images for a single prompt are nearly identical. It proposes SCAD, a GAN family that adds a text-conditioned Slicing Adversarial Network loss, two specialized discriminators (semantic and fidelity), and a mutual-information regularizer to reverse that collapse while keeping training cheap. With these additions, it reports zero-shot FID scores on COCO comparable to StyleGAN-T at roughly one-tenth to one-hundredth of the training cost, and measurably higher per-prompt diversity than the GALIP baseline. It also introduces a metric, Per-Prompt Diversity (PPD), built on embedding distances, to quantify how varied a model's outputs are for a fixed prompt. A sympathetic reader would take these experiments as evidence that GANs remain a viable fast, low-cost alternative to diffusion models once diversity is explicitly addressed.

What carries the argument

The load-bearing object is a text-conditioned Slicing Adversarial Network discriminator. Instead of a single scalar direction, it uses a condition-dependent direction $\omega(c)$ and a spectrally normalized feature extractor $h(x,c)$, so the discriminator estimates sliced Wasserstein distance between real and fake distributions conditioned on the prompt; this metrizability is what the paper uses to avoid rough distance estimates that let modes collapse. Around this core sit two auxiliary mechanisms: a fidelity branch (PatchGAN) and a semantic branch (the SAN head) as expert discriminators, and a noise-predictor head that maximizes mutual information between input noise $z$ and generated images. The evaluation machinery is Per-Prompt Diversity, defined as the $\ell^p$ norm (with $p=10$) of DINOv2 image-embedding distances to the prompt-conditional mean, averaged over prompts to give mPPD.

What would settle it

Take the prompts from the paper's evaluation, ask human raters to rank which model's set of images is more varied, and check whether mPPD ranks the models the same way; additionally recompute mPPD with the same number of images per prompt (say N=40) for every model in Table III. If human rankings disagree with mPPD, or if equalizing N changes which model leads, the paper's central diversity claim is not settled.

Watch

Extended reading notes

Core claim

The central discovery is that conditional mode collapse in CLIP-aided generators has identifiable causes—over-reliance on the purified CLIP image feature and CLIP guidance that downweights the noise input—and that these causes can be attacked directly. The paper shows that replacing the standard discriminator with a text-conditioned SAN discriminator, splitting discrimination into a semantic branch and a fidelity branch, and adding a mutual-information regularizer that forces the generator's noise to be recoverable from its images jointly restores diversity without sacrificing fidelity. In the paper's own terms, this yields SCAD-DD, which reaches zero-shot FID 30k of 12.34 on COCO in about 190 A100 days, and SCAD-MI, which reaches 13.93 in about 10 A100 days at an early training stage—numbers comparable to StyleGAN-T's 13.90 at a fraction of its cost—while mPPD rises from 1.93 for GALIP to 2.26–2.32 for the proposed variants.

Load-bearing premise

The argument stands on the assumption that the Per-Prompt Diversity score, computed from embedding distances to the average image for a prompt, correctly reflects how much visual variety humans perceive, and that mPPD numbers stay comparable when different models are evaluated with different numbers of images per prompt.

Editorial extensions

If this is right

  • If the SCAD results hold, StyleGAN-T-level zero-shot FID is obtainable on 12M image-text pairs with roughly 190 A100 days (SCAD-DD), or 10 A100 days if a small FID gap is acceptable (SCAD-MI), making large-scale GAN training accessible to smaller budgets.
  • The per-prompt diversity of CLIP-aided generators is not fixed by architecture; the proposed losses and discriminators recover it, suggesting the same components can rescue other CLIP-aided GAN designs.
  • The PPD metric gives text-to-image researchers a cheap, linear-time way to measure conditional diversity without collecting multiple real images per prompt, and it can be applied to any generative model.
  • Since the paper's models are trained on a relatively small dataset (CC12M), its results imply the proposed training recipe scales; pushing to larger datasets may close the remaining FID gap to diffusion models.

Reading between the lines

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

  • Beyond the paper: the paper's mPPD comparisons may understate diffusion-model diversity because diffusion models were evaluated with N=20 samples per prompt while GANs used N=40, and an empirical-mean-based distance decreases with smaller N; re-evaluating all models at the same N could change rankings.
  • Beyond the paper: if PPD tracks human perception, the conditional-mode-collapse diagnosis likely extends beyond GALIP to any generator that routes randomness through a frozen CLIP encoder, and SCAD's recipe is a natural test for those models.
  • Beyond the paper: combining mutual-information regularization and dual discriminators in one model was tried and reportedly hard to stabilize, so a promising extension is to find a schedule or weighting that lets both techniques stack, since each independently improves diversity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes SCAD, a family of CLIP-aided text-to-image GANs that combine a pretrained CLIP image encoder in the generator with three modifications: a Slicing Adversarial Network (SAN) loss adapted to text conditioning, two expert discriminators (a semantic branch and a fidelity branch), and a mutual-information regularizer. The authors also introduce Per-Prompt Diversity (PPD), a DINOv2-embedding-based metric intended to measure the diversity of images produced from a single prompt. They report that SCAD-DD improves zero-shot FID and per-prompt diversity over GALIP on CC12M/COCO, reaches a zero-shot FID comparable to StyleGAN-T at roughly 10-190 A100 days of training, and claim this is achieved 'without compromise' in diversity.

Significance. If the results hold, the paper offers a practical step toward efficient text-to-image GANs: SCAD-DD reaches zFID 12.34 on COCO with 190 A100 days, and SCAD-MI (early) reaches 13.93 with 10 A100 days, compared with StyleGAN-T's 1,706 A100 days. The proposed PPD metric addresses a real gap in text-to-image evaluation, and the clean-fid based comparisons with GALIP are externally grounded. The main weakness is that the diversity metric as defined is not scale-invariant and is compared across different sample counts, so the central diversity claim is not yet established as written.

major comments (4)
  1. [§IV-B, Eq. (12); §V-C, Tab. III] PPD in Eq. (12) is defined as a sum over N, not an average: PPD(c) = sum_{i=1}^N ||s(G(z_i,c)) - \bar{s}_c||_p. The expectation of this quantity grows with N for any fixed per-image dispersion. The statement in §IV-B that 'using a larger number of evaluation samples does not lead to a better score; it only affects the accuracy of the estimation' is therefore incorrect under the printed definition. Tab. III compares mPPD for GANs evaluated with N=40 (and K=1000) against diffusion models evaluated with N=20 (and K=500), so the mPPD column is not comparable across those rows. Because mPPD is the only quantitative diversity evidence in the paper, the central claim of 'increased diversity' is not supported as reported. The authors should redefine PPD as a mean over N (or normalize by N) and re-run the diversity comparisons with equal N for all models.
  2. [Abstract; §I; §V-C2, Tab. III] The abstract states that 'SCAD-DD achieves a zero-shot FID 30k of 11.65 on COCO in about 190 A100 days,' while the Introduction states 12.34 and Tab. III reports SCAD-DD (ours) as zFID30k = 12.34. These numbers must be reconciled before publication.
  3. [§V-B2, Tab. II; §I Contributions] On the COCO ablation, SCAD-MI has mPPD 2.07, which is lower than the base SCAD ablation's 2.15. This contradicts the contribution statement that 'Our models (SCAD-MI and SCAD-DD) achieve richer diversity for a given prompt both qualitatively and quantitatively (e.g., in Fig. 1 and Tab. II)' and the claim in §V-B2 that the models achieved more than 10%-20% improvement on mPPD from GALIP. The paper should either explain why SCAD-MI has lower mPPD than its own base model despite being presented as a diversity-improving variant, or revise the claims to apply only to SCAD-DD.
  4. [§V-C2, Tab. III] mPPD rows for StyleGAN-T and GigaGAN are missing; the authors state that they could not evaluate these models. Consequently, the headline 'without compromise' claim relative to large-scale GANs is supported only by FID, not by the proposed diversity metric. The diversity comparison is effectively limited to GALIP and the authors' own ablations. Please either provide mPPD for these baselines with matched N and resolution, or soften the claim to state that diversity is evaluated only against GALIP.
minor comments (6)
  1. [§IV-B] Please clarify in the text whether PPD is intended as a sum or an average; the current notation and the surrounding discussion are inconsistent, and this affects the interpretation of all mPPD values.
  2. [§V-C2, Tab. III] The table footnote states that mPPD was evaluated at 224px native resolution while other metrics were evaluated at 256px; please explain why diversity measured at a different resolution is treated as comparable across models.
  3. [§II-B] The sentence 'we introduce 1) Slicing Adversarial Networks (SANs) [13] loss' is missing an article; it should read 'the SAN loss' or 'the Slicing Adversarial Network loss'.
  4. [§V-C2] The phrase 'almost same as StyleGAN-T in score-wise' is ungrammatical; consider rewording to 'nearly matches StyleGAN-T in FID'.
  5. [Fig. 4] The axis labels 'model-wise' and 'prompt-wise' are unclear; please describe both panels in the caption, including which quantity is on each axis.
  6. [§IV] The paper does not provide any human-subject validation of PPD as a perceptual diversity measure; acknowledging this limitation or adding a small user study would strengthen the metric's credibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the empirical SCAD results rest on external FID/CLIP benchmarks, and the self-cited SAN motivation is a non-load-bearing hypothesis; the PPD metric raises a separate comparability concern.

full rationale

The paper's central claim is empirical: SCAD-MI and SCAD-DD are compared with GALIP, StyleGAN-T, GigaGAN, and SDv1.4 using clean-fid zero-shot FID, FID-CLIP, CLIP score, and mPPD. These benchmarks are external to the training procedure; mPPD is computed with DINOv2 embeddings and is not a training objective, so the reported improvements over GALIP are measurements rather than values fitted into the model. The only notable self-citation is the SAN loss from Takida et al. [13], which shares co-authors with this paper; its metrizability theory motivates the discriminator design, but the paper treats the causal story as a hypothesis and the FID/CLIP gains do not depend on accepting that theory. No equation in the paper reduces a prediction to an input by construction, and no fitted parameter is renamed as a prediction. Two non-circularity caveats remain. First, Eq. (12) defines PPD as an unnormalized sum over N, and Tab. III compares N=40 GANs with N=20 diffusion models, so those mPPD comparisons are not quantitatively comparable as printed; this is a metric-validity concern, not a circular derivation. Second, the paper explicitly states it did not evaluate mPPD for StyleGAN-T and GigaGAN, so the diversity comparison against those large GANs is unsupported. These caveats affect the strength of the diversity claim but do not make the derivation circular.

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

No new physical entities are introduced; the contributions are architectural and statistical.

free parameters (5)
  • MI loss weight lambda = 1.0
    Chosen ad hoc for SCAD-MI; the paper does not ablate this value and SCAD-DD uses 0.0.
  • CLIP guidance weight mu = 4.0
    Inherited from GALIP; not tuned here and not ablated.
  • Gradient penalty weights = k1=0.5, k2=0.1, l1=6, l2=1
    Inherited from GALIP defaults; the paper does not ablate them.
  • PPD norm order p = 10
    Chosen after checking several values; the authors report the ordering is stable across p, so this is a low-sensitivity choice.
  • PPD sample counts (N, K) = N=40/K=1000 for GANs; N=20/K=500 for diffusion
    Evaluation-protocol choices; the differing N biases cross-model mPPD comparisons.
assumptions (5)
  • domain assumption Metrizability conditions (injectivity, separability, direction optimality) from SAN hold for text-to-image discriminators
    Section III-A2 extends SAN to text-to-image based on this assumption, without verifying the conditions in the new setting.
  • domain assumption CLIP image embeddings are insensitive to pixel-level details, so the two-stage generator's feat2 acts as a purified image prompt
    Section II-A's interpretation motivates the architecture and the mode-collapse hypothesis, but is not tested independently.
  • domain assumption DINOv2 embedding dispersions track human-perceived per-prompt diversity
    PPD (Section IV-B) is defined on this premise; no human evaluation or comparison with LPIPS diversity is reported.
  • domain assumption The official GALIP checkpoint and clean-fid evaluation are faithful baselines
    All comparisons inherit this; the paper notes GALIP's evaluation script differs from clean-fid and then adopts clean-fid.
  • standard math The Barber-Agakov variational lower bound is a valid estimator for the MI regularizer
    Used in Eq. (7) to justify the noise predictor; standard in InfoGAN-style methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficiency without Compromise: CLIP-aided Text-to-Image GANs with Increased Diversity." pith.science (2026). https://pith.science/paper/5TYEHFC7

@misc{pith2026250601493,
  author       = {Pith},
  title        = {Pith review of: Efficiency without Compromise: CLIP-aided Text-to-Image GANs with Increased Diversity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TYEHFC7}},
  note         = {Machine review of arXiv:2506.01493}
}
read the original abstract

Recently, Generative Adversarial Networks (GANs) have been successfully scaled to billion-scale large text-to-image datasets. However, training such models entails a high training cost, limiting some applications and research usage. To reduce the cost, one promising direction is the incorporation of pre-trained models. The existing method of utilizing pre-trained models for a generator significantly reduced the training cost compared with the other large-scale GANs, but we found the model loses the diversity of generation for a given prompt by a large margin. To build an efficient and high-fidelity text-to-image GAN without compromise, we propose to use two specialized discriminators with Slicing Adversarial Networks (SANs) adapted for text-to-image tasks. Our proposed model, called SCAD, shows a notable enhancement in diversity for a given prompt with better sample fidelity. We also propose to use a metric called Per-Prompt Diversity (PPD) to evaluate the diversity of text-to-image models quantitatively. SCAD achieved a zero-shot FID competitive with the latest large-scale GANs at two orders of magnitude less training cost.

Figures

Figures reproduced from arXiv: 2506.01493 by the authors.

Figure 1
Figure 1. Images generated from GALIP and the proposed model SCAD-DD trained with CC12M dataset (in 224 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of CLIP-aided generators. Two decoders (Dec1, Dec2) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of our discriminator. It incorporates an objective function originating from SAN. This illustration is based on the semantic branch mainly. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Validation of PPD with various p (the order of a norm). The left figure shows the averaged PPD of 10 prompts in SCAD-MI (blue line) and GALIP (gray line). The right figure shows PPDs with different 10 prompts in SCAD-MI. where s¯c is an average of N embeddings. We test…
Figure 5
Figure 5. Figure 5: Samples from the models trained with the COCO dataset (83k images) in 224 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: 256×256 images generated by our proposed models. An suffix ”, movie light, high-resolution, best quality” is added to the left two prompts, and ”, high-resolution, masterpiece, artstation” is added to the right prompt. implementation5 is reported to be not reproducible…
Figure 7
Figure 7. Figure 7: Interpolation of the latent space with SCAD-DD in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 37 canonical work pages

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems (H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, eds.), vol. 33, pp. 6840– 6851, Curran Associates, Inc., 2020

    2020

  2. [2]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pp. 10684–10695, June 2022

    work page 2022

  3. [3]

    Hierarchical text-conditional image generation with clip latents,

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with clip latents,” 2022

    2022

  4. [4]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing Systems (Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, eds.), vol. 27, Curran Associates, Inc., 2014

    work page 2014

  5. [5]

    Tackling the generative learning trilemma with denoising diffusion GANs,

    Z. Xiao, K. Kreis, and A. Vahdat, “Tackling the generative learning trilemma with denoising diffusion GANs,” in International Conference on Learning Representations (ICLR) , 2022

    work page 2022

  6. [6]

    StyleGAN- t: Unlocking the power of GANs for fast large-scale text-to-image synthesis,

    A. Sauer, T. Karras, S. Laine, A. Geiger, and T. Aila, “StyleGAN- t: Unlocking the power of GANs for fast large-scale text-to-image synthesis,” in Proceedings of the 40th International Conference on Machine Learning (A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, eds.), vol. 202 of Proceedings of Machine Learning Research, pp....

    work page 2023

  7. [7]

    Scaling up gans for text-to-image synthesis,

    M. Kang, J.-Y . Zhu, R. Zhang, J. Park, E. Shechtman, S. Paris, and T. Park, “Scaling up gans for text-to-image synthesis,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    work page 2023

  8. [8]

    Bk-sdm: A lightweight, fast, and cheap version of stable diffusion,

    B.-K. Kim, H.-K. Song, T. Castells, and S. Choi, “Bk-sdm: A lightweight, fast, and cheap version of stable diffusion,” arXiv preprint arXiv:2305.15798, 2023

    arXiv 2023

Show all 50 references
  1. [9]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9650–9660, October 2021

    2021

  2. [10]

    Learning transferable visual models from natural language supervi- sion,

    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 transferable visual models from natural language supervi- sion,” in Proceedings of the 38th International Conference on Machine...

    2021

  3. [11]

    Projected gans con- verge faster,

    A. Sauer, K. Chitta, J. M ¨uller, and A. Geiger, “Projected gans con- verge faster,” in Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  4. [12]

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

    M. Tao, B.-K. Bao, H. Tang, and C. Xu, “Galip: Generative adversarial clips for text-to-image synthesis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 14214– 14223, 2023

    2023

  5. [13]

    SAN: Inducing metrizability of GAN with discrimi- native normalized linear layer,

    Y . Takida, M. Imaizumi, T. Shibuya, C.-H. Lai, T. Uesaka, N. Murata, and Y . Mitsufuji, “SAN: Inducing metrizability of GAN with discrimi- native normalized linear layer,” in The Twelfth International Conference on Learning Representations , 2024

    2024

  6. [14]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Journal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  7. [15]

    A question of terminology,

    J. Kreer, “A question of terminology,” IRE Transactions on Information Theory, vol. 3, no. 3, pp. 208–208, 1957

    1957

  8. [16]

    Mutual information neural estimation,

    M. I. Belghazi, A. Baratin, S. Rajeshwar, S. Ozair, Y . Bengio, A. Courville, and D. Hjelm, “Mutual information neural estimation,” in Proceedings of the 35th International Conference on Machine Learning (J. Dy and A. Krause, eds.), vol. 80 of Proceedings of Machine Learning R...

    2018

  9. [17]

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

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in Proceedings of the 31st International Conference on Neural Information Processing Systems , NIPS’17, (Red Hook, NY , U...

    2017

  10. [18]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, X. Chen, and X. Chen, “Improved techniques for training gans,” in Ad- vances in Neural Information Processing Systems (D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, eds.), vol. 29, Curran Associates, ...

    2016

  11. [19]

    Df- gan: A simple and effective baseline for text-to-image synthesis,

    M. Tao, H. Tang, F. Wu, X.-Y . Jing, B.-K. Bao, and C. Xu, “Df- gan: A simple and effective baseline for text-to-image synthesis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16515–16525, 2022

    2022

  12. [20]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 43, pp. 4217–4228, dec 2021

    2021

  13. [21]

    Understanding noise injection in gans,

    R. Feng, D. Zhao, and Z.-J. Zha, “Understanding noise injection in gans,” in Proceedings of the 38th International Conference on Machine Learning (M. Meila and T. Zhang, eds.), vol. 139 of Proceedings of Machine Learning Research , pp. 3284–3293, PMLR, 18–24 Jul 2021

    2021

  14. [22]

    The caltech-ucsd birds-200-2011 dataset,

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The caltech-ucsd birds-200-2011 dataset,” Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011

    2011

  15. [23]

    Improving the improved training of wasserstein GANs,

    X. Wei, Z. Liu, L. Wang, and B. Gong, “Improving the improved training of wasserstein GANs,” in International Conference on Learning Representations, 2018

    2018

  16. [24]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Computer Vision (ICCV), 2017 IEEE International Conference on , 2017

    2017

  17. [25]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pp. 12873– 12883, June 2021

    2021

  18. [26]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    2014

  19. [27]

    Generating sentences from a continuous space,

    S. Bowman, L. Vilnis, O. Vinyals, A. Dai, R. Jozefowicz, and S. Bengio, “Generating sentences from a continuous space,” pp. 10–21, 01 2016

    2016

  20. [28]

    Infogan: interpretable representation learning by information maximizing generative adversarial nets,

    X. Chen, Y . Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel, “Infogan: interpretable representation learning by information maximizing generative adversarial nets,” in Proceedings of the 30th International Conference on Neural Information Processing Systems , NIP...

    2016

  21. [29]

    The im algorithm: a variational approach to information maximization,

    D. Barber and F. Agakov, “The im algorithm: a variational approach to information maximization,” in Proceedings of the 16th International Conference on Neural Information Processing Systems, NIPS’03, (Cam- bridge, MA, USA), p. 201–208, MIT Press, 2003

    2003

  22. [30]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  23. [31]

    Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation,

    N. Ruiz, Y . Li, V . Jampani, Y . Pritch, M. Rubinstein, and K. Aberman, “Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023

    2023

  24. [32]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatu...

    2024

  25. [33]

    Vision transformers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski, “Vision transformers need registers,” in The Twelfth International Conference on Learning Representations, 2024

    2024

  26. [34]

    The role of imagenet classes in fr ´echet inception distance,

    T. Kynk ¨a¨anniemi, T. Karras, M. Aittala, T. Aila, and J. Lehtinen, “The role of imagenet classes in fr ´echet inception distance,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  27. [35]

    Openclip,

    G. Ilharco, M. Wortsman, R. Wightman, C. Gordon, N. Carlini, R. Taori, A. Dave, V . Shankar, H. Namkoong, J. Miller, H. Hajishirzi, A. Farhadi, and L. Schmidt, “Openclip,” July 2021. If you use this software, please cite it as below

    2021

  28. [36]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision – ECCV 2014 (D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, eds.), (Cham), pp. 740–755, Springer Internationa...

    2014

  29. [37]

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

    S. Changpinyo, P. Sharma, N. Ding, and R. 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 (CVPR) , pp. 3558–3568, June 2021

    2021

  30. [38]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” in NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applica- tions, 2021

    2021

  31. [39]

    Generative adversarial text to image synthesis,

    S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee, “Generative adversarial text to image synthesis,” in Proceedings of The 33rd International Conference on Machine Learning (M. F. Balcan and K. Q. Weinberger, eds.), vol. 48 of Proceedings of Machine Learning Res...

    2016

  32. [40]

    Stackgan: Text to photo-realistic image synthesis with stacked gener- ative adversarial networks,

    H. Zhang, T. Xu, H. Li, S. Zhang, X. Wang, X. Huang, and D. Metaxas, “Stackgan: Text to photo-realistic image synthesis with stacked gener- ative adversarial networks,” in 2017 IEEE International Conference on Computer Vision (ICCV) , pp. 5908–5916, 2017

    2017

  33. [41]

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

    T. Xu, P. Zhang, Q. Huang, H. Zhang, Z. Gan, X. Huang, and X. He, “Attngan: Fine-grained text to image generation with attentional generative adversarial networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    2018

  34. [42]

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

    H. Zhang, J. Y . Koh, J. Baldridge, H. Lee, and Y . Yang, “Cross-modal contrastive learning for text-to-image generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 833–842, June 2021

    2021

  35. [43]

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

    Y . Zhou, R. Zhang, C. Chen, C. Li, C. Tensmeyer, T. Yu, J. Gu, J. Xu, and T. Sun, “Towards language-free training for text-to-image generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 17907–17917, June 2022

    2022

  36. [44]

    Spectral normal- ization for generative adversarial networks,

    T. Miyato, T. Kataoka, M. Koyama, and Y . Yoshida, “Spectral normal- ization for generative adversarial networks,” in International Conference on Learning Representations , 2018

    2018

  37. [45]

    Progressive growing of GANs for improved quality, stability, and variation,

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of GANs for improved quality, stability, and variation,” in International Conference on Learning Representations , 2018

    2018

  38. [46]

    De- mystifying MMD GANs,

    M. Bi ´nkowski, D. J. Sutherland, M. Arbel, and A. Gretton, “De- mystifying MMD GANs,” in International Conference on Learning Representations, 2018

    2018

  39. [47]

    Assessing Generative Models via Precision and Recall,

    M. S. M. Sajjadi, O. Bachem, M. Lu ˇci´c, O. Bousquet, and S. Gelly, “Assessing Generative Models via Precision and Recall,” in Advances in Neural Information Processing Systems (NeurIPS) , 2018

    2018

  40. [48]

    Kynk ¨a¨anniemi, T

    T. Kynk ¨a¨anniemi, T. Karras, S. Laine, J. Lehtinen, and T. Aila, Improved precision and recall metric for assessing generative models . Red Hook, NY , USA: Curran Associates Inc., 2019

    2019

  41. [49]

    Feature likelihood score: Evaluating the generalization of generative models using samples,

    M. Jiralerspong, J. Bose, I. Gemp, C. Qin, Y . Bachrach, and G. Gidel, “Feature likelihood score: Evaluating the generalization of generative models using samples,” in Thirty-seventh Conference on Neural Infor- mation Processing Systems , 2023

    2023

  42. [50]

    Holistic evaluation of text-to-image models,

    T. Lee, M. Yasunaga, C. Meng, Y . Mai, J. S. Park, A. Gupta, Y . Zhang, D. Narayanan, H. B. Teufel, M. Bellagente, M. Kang, T. Park, J. Leskovec, J.-Y . Zhu, L. Fei-Fei, J. Wu, S. Ermon, and P. Liang, “Holistic evaluation of text-to-image models,” inThirty-seventh Conference o...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.