Pith. sign in

REVIEW 3 major objections 5 minor 49 references

A Diffusion Model Translator for Efficient Image-to-Image Translation

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that translating between two image domains with a pretrained diffusion model needs only a single intermediate noise level, where a lightweight translator maps the noisy source to the noisy target and a frozen DDPM…

desk verdict The recipe works and the speedups are real, but the paper's central proof that one intermediate step is sufficient doesn't hold; treat it as a strong empirical heuristic. read the letter →

arxiv 2502.00307 v1 pith:22HFX6IG submitted 2025-02-01 cs.CV

Mengfei Xia , Yu Zhou , Ran Yi , Yong-Jin Liu , Wenping Wang This is my paper · ORCID
classification cs.CV
keywords diffusionmodelsimage-to-imagetranslationdenoisingprobabilisticmodelintermediatetimestepdomaintransferefficientinferenceselectionpairedtraining
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 claims that image-to-image translation can be split into one lightweight translation step at a single intermediate diffusion timestep plus a frozen denoising diffusion model, instead of injecting the source image into every denoising step. The authors prove that at that timestep the optimal translator is a Gaussian whose mean is the noisy clean target, which turns training into a simple regression from noisy source to noisy target. They also give a practical rule for picking the timestep from the intersection of two SSIM-distance curves. In experiments on stylization, colorization, segmentation-to-image, and sketch-to-image, the resulting system reports better FID and LPIPS than Palette, Pix2Pix, and TSIT while cutting inference time from tens of seconds to under half a second. If the claim holds, any pretrained DDPM can be turned into a fast conditional generator by plugging in a small translator at the right noise level.

What carries the argument

The load-bearing object is the intermediate translation module $p_\theta(y_t|x_t)$ inserted at one timestep $t$ of two coupled forward diffusion chains. The coupling uses one shared Gaussian noise $z$: $x_t = \sqrt{\bar{\alpha}_t}x_0 + \sqrt{1-\bar{\alpha}_t}z$ and $y_t = \sqrt{\bar{\alpha}_t}y_0 + \sqrt{1-\bar{\alpha}_t}z$. Theorems 1 and 2 convert the conditional generation objective into the closed-form regression $\|f_\theta(x_t) - y_t\|^2$, which lets the translator be trained like an ordinary image-to-image network while the DDPM stays frozen. The timestep $t$ is selected before training by locating the intersection of the SSIM-distance curves $d(x_t, y_t)$ and $d(x_0, x_t)$, which the complexity argument identifies as the minimum of the system's bottleneck.

What would settle it

Train the DMT objective on a task with genuine one-to-many ambiguity, such as colorization from grayscale, and compare the trained $f_\theta(x_t)$ against $y_t$ for matched noise $z$; if the mean prediction differs systematically on inputs that admit multiple valid colors, the single-Gaussian optimum assumed in Theorem 2 is not the true minimizer of the KL objective.

Watch

Extended reading notes

Core claim

The central claim is that the conditional distribution $q(y_0|x_0)$ of a paired image translation task can be approximated by a Markov process that couples the forward diffusions of $x_0$ and $y_0$ with the same noise $z$ until a chosen timestep $t$, then applies a learned translator $p_\theta(y_t|x_t)$, then runs the pre-trained DDPM backward from $y_t$. Theorem 1 and Theorem 2 state that the variational lower bound for $\log p_\theta(y_0|x_0)$ reduces to a KL divergence between $q(y_t|y_0)$ and $p_\theta(y_t|x_t)$, and that the optimal $p_\theta$ is Gaussian with mean $\sqrt{\bar{\alpha}_t} y_0$. The authors reparameterize this mean as $f_\theta(x_t) - \sqrt{1-\bar{\alpha}_t}z$, so the training loss is simply $\|f_\theta(x_t) - y_t\|^2$ with $x_t, y_t$ computed from the same noise. A timestep-selection heuristic chooses $t$ where the SSIM distance between $(x_t, y_t)$ and between $(x_0, x_t)$ curves intersect, balancing domain gap against content preservation. The empirical section reports that this single-step translator, built on Pix2Pix or TSIT backbones, outperforms the strong baselines in FID and LPIPS on all four tested tasks and is 40--80x faster than Palette.

Load-bearing premise

The derivation assumes the noisy source image determines the clean target image exactly, so that the optimal translator is the single Gaussian with mean equal to the noisy clean target; if that identification fails, the closed form is only an approximation.

Editorial extensions

If this is right

  • A pretrained DDPM can be reused for conditional generation without any modification, making the translator the only trainable component.
  • The training cost drops to hundreds of epochs on a single translator instead of thousands of epochs of conditional diffusion training, and inference uses roughly 10 DDIM steps from the intermediate timestep.
  • Because the translator is task-agnostic in form, any existing paired I2I network (Pix2Pix, TSIT, or future backbones) can be wrapped by DMT to gain the diffusion prior's realism.
  • The timestep $t^*$ chosen by the SSIM-intersection rule is near-optimal on four datasets, meaning the method can be deployed without per-task tuning.
  • Asymmetric timestep pairs $(s,t)$ do not beat the symmetric choice $s=t$, so the single-timestep design is not leaving performance on the table.

Reading between the lines

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

  • A natural extension is to replace the paired translator with a conditional model trained on unpaired data at the chosen timestep, which would extend DMT to settings where paired targets are unavailable while keeping the frozen DDPM.
  • The proof's reliance on $y_0$ being recoverable from $x_t$ suggests the closed form is exact only when the source-to-target mapping is deterministic; for stochastic or ambiguous mappings, a mixture-of-Gaussians translator might improve results.
  • One could test the timestep-selection heuristic analytically: if the SSIM curves are roughly linear near the intersection, $t^*$ should scale with the perceptual distance between domains, giving a closed-form prior for new tasks.
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 / 5 minor

Summary. The paper proposes DMT, a method that equips a pre-trained DDPM with a lightweight translator that maps noisy source xt to noisy target yt at a single timestep, then denoises with the frozen DDPM. It claims theoretical proof that one-step transfer is feasible and sufficient, and an automatic timestep selection strategy. Experiments on stylization, colorization, segmentation, and sketch show improved FID/LPIPS and speedups against Palette, Pix2Pix, and TSIT.

Significance. If the theoretical claim were valid, it would justify a simple and efficient recipe: train any I2I translator on noisy pairs at one timestep and reuse a frozen DDPM. The empirical results are encouraging: the method consistently improves over its GAN backbones and is 40-80x faster than Palette. The paper ships pseudo-code and extensive ablations. However, the central proof is incorrect and the timestep selection fails on one task; the contribution currently rests on heuristic validity.

major comments (3)
  1. [Appendix B, Theorem 2] The minimizer of the KL in Eq. (15) for a fixed xt is p*(yt|xt)=∫ q(y0|xt) q(yt|y0) dy0, a mixture of Gaussians, not the Gaussian N(yt;√¯αt y0,(1−¯αt)I). The proof's step from 'q(yt|y0) is Gaussian' to 'pθ(yt|xt) is Gaussian with mean √¯αt y0' is invalid unless q(y0|xt) is a point mass, which fails generically. Since Eq. (18) and the objective Eq. (19) depend on this identification, the theoretical claim of feasibility and sufficiency stated in the abstract and Section 3.2 is unsupported.
  2. [Section 4.4, Table 5] The text states that the automatically selected timestep t* achieves performance comparable to the optimal t in Table 5. The stylization column contradicts this: t=20 gives FID 11.01 and SSIM 0.760, while t*=50 gives FID 20.46 and SSIM 0.732. The selected timestep is far from optimal on this task, so the practical selection strategy is not validated as claimed.
  3. [Section 3.4] The 'theoretical analysis' of system complexity is not a proof. The decomposition C(S)=max{C(S1),C(S2),C(S3)} is asserted without defining complexity, the relations f(t)≈h(t) and monotonicity of f and g are assumptions rather than derived facts, and the conclusion that max{f,g} is minimized at the intersection of the two curves requires regularity conditions that are not stated. The SSIM-intersection rule may be a reasonable heuristic, but presenting it as theory overstates its status.
minor comments (5)
  1. [Figures 5 and 6] The captions contain the typo 'TSID-DMT'; this should be 'TSIT-DMT'.
  2. [Equation (34)] Equation (34) is malformed: the last three terms are written as λSSIM(... ) without a plus sign; the intended weighted sum should be typeset with explicit addition.
  3. [Section 3.4] The text refers to 'SSIM distance' but never defines it; the figures plot a quantity labeled 'SSIM distance' that appears to be 1−SSIM. Please define the metric.
  4. [Section 4.5, Eq. (35)] The negative log-likelihood should be −log pθ(x0|x0), not log pθ(x0|x0); in continuous spaces the pointwise likelihood is not a probability, so the displayed equality to 0 is not meaningful as written.
  5. [Tables 4-7] The quantitative comparisons report single runs without confidence intervals or significance tests; given the moderate gaps on some tasks (e.g., colorization FID 13.03 vs 13.28 at t=0 vs t*), please state how many seeds were used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DMT objective is a standard L2 regression on noisy source/target pairs, and no fitted constant is renamed as a prediction; the Theorem 2 proof contains a non-circular mathematical gap.

full rationale

The paper is self-contained against external benchmarks: the headline comparisons are against Pix2Pix, TSIT, SPADE, Palette, and QMUPD on external datasets (AFHQ, CelebA-HQ, Edges2handbags), and the DMT training objective (Eq. 19) is a plain L2 regression on paired noisy images, not a fitted parameter disguised as a prediction. The timestep t* is pre-selected from SSIM distances before training (Section 3.4) and then evaluated with FID, so the reported metrics are not constructed from the selection criterion. The only self-citation is QMUPD [42], used to synthesize the handcrafted Portrait dataset and as a baseline; the central claim also rests on external data and methods, so this citation is not load-bearing. Section 4.5 states honest limitations (Markov assumption, paired data, degenerate x0=y0 case), none of which conceal a circular argument. The proof of Theorem 2 in Appendix B contains an unjustified step: from q(yt|y0) ~ N(sqrt(alpha_t)y0, ...) it concludes the optimal pθ(yt|xt) has mean sqrt(alpha_t)y0, whereas minimizing Eq. (15) over pθ(yt|xt) for a fixed xt gives a mixture E_{q(y0|xt)}[q(yt|y0)]. This is a mathematical correctness gap in the theoretical justification, not a circular reduction: the objective and the empirical pipeline are not equivalent to the claimed result by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central derivation rests on standard variational calculus plus two domain assumptions: paired data with shared noise, and a heuristic complexity model. The most fragile input is the claim that the optimal translator is a Gaussian with the paired target mean, which is not achievable from xt alone.

free parameters (2)
  • translation timestep t* = stylization 50, colorization 5, segmentation 200, sketch 20
    Selected by SSIM intersection heuristic before training; affects FID heavily (Table 5).
  • regularization weight lambda = 0.5
    Hand-chosen in Eq (34) for asymmetric timestep search.
assumptions (4)
  • standard math Variational bound and Bayes rule for Gaussian diffusion chains
    Used in Lemma 1 and Appendix B to derive the VLB upper bound.
  • domain assumption Paired source-target data (x0,y0) with a shared noise sample z_t in the forward processes
    Eq (17) requires the same z_t for both domains; this is a training-data construction, not a general property.
  • ad hoc to paper Complexity of the whole I2I system equals the max of three subsystem complexities; f(t) and g(t) are monotone; f(t) approximately h(t)
    Stated in Section 3.4 without validation; used to justify the timestep intersection heuristic.
  • ad hoc to paper The optimal pθ(yt|xt) is the Gaussian q(yt|y0) with mean sqrt(alpha_bar_t)y0, implying y0 is recoverable from xt
    Theorem 2 and Appendix B; this is the load-bearing gap in the theoretical claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Diffusion Model Translator for Efficient Image-to-Image Translation." pith.science (2026). https://pith.science/paper/22HFX6IG

@misc{pith2026250200307,
  author       = {Pith},
  title        = {Pith review of: A Diffusion Model Translator for Efficient Image-to-Image Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22HFX6IG}},
  note         = {Machine review of arXiv:2502.00307}
}
read the original abstract

Applying diffusion models to image-to-image translation (I2I) has recently received increasing attention due to its practical applications. Previous attempts inject information from the source image into each denoising step for an iterative refinement, thus resulting in a time-consuming implementation. We propose an efficient method that equips a diffusion model with a lightweight translator, dubbed a Diffusion Model Translator (DMT), to accomplish I2I. Specifically, we first offer theoretical justification that in employing the pioneering DDPM work for the I2I task, it is both feasible and sufficient to transfer the distribution from one domain to another only at some intermediate step. We further observe that the translation performance highly depends on the chosen timestep for domain transfer, and therefore propose a practical strategy to automatically select an appropriate timestep for a given task. We evaluate our approach on a range of I2I applications, including image stylization, image colorization, segmentation to image, and sketch to image, to validate its efficacy and general utility. The comparisons show that our DMT surpasses existing methods in both quality and efficiency. Code will be made publicly available.

Figures

Figures reproduced from arXiv: 2502.00307 by the authors.

Figure 1
Figure 1. Conceptual comparison between (a) existing methods [9]–[11] and (b) our DMT. {xt} T t=0 represent different states of the input from the source domain, while yT → y0 stands for the denoising process of DDPM. Here, T denotes the total number of noise-adding steps in the diffusion process. Instead of using the information ft(x) from the source domain (which can be the original or noisy image) for an iterative refineme… view at source ↗
Figure 2
Figure 2. Qualitative results of our proposed DMT on four I2I tasks: image stylization, image colorization, segmentation to image, and sketch to image. Here we equip a pre-trained DDPM with an efficient translation module. Our approach makes adequate use of the content information from the input condition as well as the domain knowledge contained in the learned denoising process. q(y0|x0) using the learned intermediate transl… view at source ↗
Figure 3
Figure 3. Analysis on the preset timestep, t. Our DMT needs a pre-defined timestep to learn and perform the distribution shift. We plot the distance between (xt, yt) and (x0, xt) at different timesteps, which are shown in red and blue curves, respectively. When t increases, d(xt, yt) decreases so that the distribution is easier to shift from xt to yt, while d(x0, xt) increases so that the input condition signal is becoming le… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Conceptual comparison for (a) multi-step DMT and (b) asymmetric DMT. {xt} T t=0 represent different states of the input from the source domain, while yT → y0 stands for the denoising process of DDPM. Here, T denotes the total number of noise-adding steps in the diffusi…
Figure 5
Figure 5. Figure 5: Qualitative comparison between DMT and SPADE [14] on segmentation-to-image task. Our proposed DMT achieves better image quality and content consisitency compared with SPADE. TABLE 3 Quantitative comparison between DMT and QMUPD [42] on image stylization task. FID, SSIM…
Figure 6
Figure 6. Figure 6: Qualitative comparison between DMT and QMUPD [42] on image stylization task. Our proposed DMT achieves better image quality and content consisitency compared with QMUPD. 2) By obtaining the mean by the functional approxi￾mator fθ according to Equation (18), we infer th…
Figure 7
Figure 7. Figure 7: Qualitative results for ablation study of the preset timestep t in our proposed DMT on the four I2I tasks. We observe that a smaller t helps in better retaining the content information from the input source, but suffers from a larger gap between the target domain and t…
Figure 8
Figure 8. Figure 8: Qualitative comparison. Our DMT achieves on par or better results than the three baseline methods Pix2Pix [7], Palette [9], TSIT [13] on the four I2I tasks, which are image stylization, image colorization, segmentation to image, and sketch to image. Significant differe…
Figure 9
Figure 9. Figure 9: Qualitative comparison when translating human face images to portraits, using our handcrafted Portrait dataset [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison when translating greyscale images to colorized ones, using AFHQ dataset [43] [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison when translating segmentation maps to images, using CelebA-HQ dataset [44] [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison when translating sketches to images, using Edges2handbags dataset [45], [46] [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages

  1. [1]

    Deep unsupervised learning using nonequilibrium thermody- namics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermody- namics,” in Int. Conf. Mach. Learn., 2015, pp. 2256–2265. 1, 2 IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE, VOL. 46, NO. 12, DECEMBER 2024 11

    work page 2015

  2. [2]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,” in Adv. Neural Inform. Process. Syst., 2020, pp. 6840–6851. 1, 2, 4, 7

    work page 2020

  3. [3]

    Score-based generative modeling through stochastic differential equations,

    Y. Song, J. Sohl-Dickstein, D. P . Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in Int. Conf. Learn. Represent., 2021. 1

    work page 2021

  4. [4]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in Int. Conf. Learn. Represent., 2021. 1, 7

    work page 2021

  5. [5]

    Improved denoising diffusion probabilistic models,

    A. Q. Nichol and P . Dhariwal, “Improved denoising diffusion probabilistic models,” in Int. Conf. Mach. Learn. , 2021, pp. 8162–

    work page 2021

  6. [6]

    Diffusion models beat gans on image synthesis,

    P . Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” in Adv. Neural Inform. Process. Syst., 2021. 1, 7

    work page 2021

  7. [7]

    Image-to-image translation with conditional adversarial networks,

    P . Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in IEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 1125–1134. 1, 2, 7, 8, 9, 10, 15

    work page 2017

  8. [8]

    Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models,

    H. Sasaki, C. G. Willcocks, and T. P . Breckon, “Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models,” arXiv preprint arXiv:2104.05358, 2021. 1, 2

    arXiv 2021

Show all 49 references
  1. [9]

    Palette: Image-to-image diffusion mod- els,

    C. Saharia, W. Chan, H. Chang, C. A. Lee, J. Ho, T. Salimans, D. J. Fleet, and M. Norouzi, “Palette: Image-to-image diffusion mod- els,” in SIGGRAPH ’22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 20...

    2022

  2. [10]

    ILVR: condition- ing method for denoising diffusion probabilistic models,

    J. Choi, S. Kim, Y. Jeong, Y. Gwon, and S. Yoon, “ILVR: condition- ing method for denoising diffusion probabilistic models,” in Int. Conf. Comput. Vis. IEEE, 2021, pp. 14 347–14 356. 1, 2

    2021

  3. [11]

    More control for free! image synthesis with semantic diffusion guidance,

    X. Liu, D. H. Park, S. Azadi, G. Zhang, A. Chopikyan, Y. Hu, H. Shi, A. Rohrbach, and T. Darrell, “More control for free! image synthesis with semantic diffusion guidance,” arXiv preprint arXiv:2112.05744, 2021. 1, 2

    2021 arXiv

  4. [12]

    Pretraining is all you need for image-to-image translation,

    T. Wang, T. Zhang, B. Zhang, H. Ouyang, D. Chen, Q. Chen, and F. Wen, “Pretraining is all you need for image-to-image translation,” in arXiv, 2022. 1, 2

    2022

  5. [13]

    TSIT: A simple and versatile framework for image-to-image translation,

    L. Jiang, C. Zhang, M. Huang, C. Liu, J. Shi, and C. C. Loy, “TSIT: A simple and versatile framework for image-to-image translation,” in ECCV, 2020. 1, 2, 7, 8, 9, 10, 15

    2020

  6. [14]

    Semantic image synthesis with spatially-adaptive normalization,

    T. Park, M.-Y. Liu, T.-C. Wang, and J.-Y. Zhu, “Semantic image synthesis with spatially-adaptive normalization,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition ,

  7. [15]

    Sean: Image synthe- sis with semantic region-adaptive normalization,

    P . Zhu, R. Abdal, Y. Qin, and P . Wonka, “Sean: Image synthe- sis with semantic region-adaptive normalization,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June

  8. [16]

    Wavegrad: Estimating gradients for waveform generation,

    N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “Wavegrad: Estimating gradients for waveform generation,”arXiv preprint arXiv:2009.00713, 2020. 2

    2009 arXiv

  9. [17]

    Diffwave: A versatile diffusion model for audio synthesis,

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Diffwave: A versatile diffusion model for audio synthesis,” arXiv preprint arXiv:2009.09761, 2020. 2

    2009 arXiv

  10. [18]

    Video diffusion models,

    J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet, “Video diffusion models,” arXiv preprint arXiv:2204.03458 ,

    arXiv

  11. [19]

    Imagen video: High definition video generation with diffusion models,

    J. Ho, W. Chan, C. Saharia, J. Whang, R. Gao, A. Gritsenko, D. P . Kingma, B. Poole, M. Norouzi, D. J. Fleet et al. , “Imagen video: High definition video generation with diffusion models,” arXiv preprint arXiv:2210.02303, 2022. 2

    arXiv

  12. [20]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,” arXiv preprint arXiv:2104.07636, 2021. 2

    2021 arXiv

  13. [21]

    Srdiff: Single image super-resolution with diffusion probabilistic models,

    H. Li, Y. Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y. Chen, “Srdiff: Single image super-resolution with diffusion probabilistic models,” Neurocomputing, 2022. 2

    2022

  14. [22]

    Learning to efficiently sample from diffusion probabilistic models,

    D. Watson, J. Ho, M. Norouzi, and W. Chan, “Learning to efficiently sample from diffusion probabilistic models,” arXiv preprint arXiv:2106.03802, 2021. 2

    2021 arXiv

  15. [23]

    Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models,

    F. Bao, C. Li, J. Zhu, and B. Zhang, “Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models,” in Int. Conf. Learn. Represent., 2022. 2

    2022

  16. [24]

    Gotta go fast when generating data with score- based models,

    A. Jolicoeur-Martineau, K. Li, R. Pich ´e-Taillefer, T. Kachman, and I. Mitliagkas, “Gotta go fast when generating data with score- based models,” arXiv preprint arXiv:2105.14080, 2021. 2

    2021 arXiv

  17. [25]

    Pseudo numerical methods for diffusion models on manifolds,

    L. Liu, Y. Ren, Z. Lin, and Z. Zhao, “Pseudo numerical methods for diffusion models on manifolds,” in Int. Conf. Learn. Represent. OpenReview.net, 2022. 2

    2022

  18. [26]

    Diffusion-based voice conversion with fast maximum likelihood sampling scheme,

    V . Popov, I. Vovk, V . Gogoryan, T. Sadekova, M. S. Kudinov, and J. Wei, “Diffusion-based voice conversion with fast maximum likelihood sampling scheme,” in Int. Conf. Learn. Represent. Open- Review.net, 2022. 2

    2022

  19. [27]

    Itˆo-taylor sampling scheme for denoising diffusion probabilistic models using ideal derivatives,

    H. Tachibana, M. Go, M. Inahara, Y. Katayama, and Y. Watanabe, “Itˆo-taylor sampling scheme for denoising diffusion probabilistic models using ideal derivatives,” arXiv preprint arXiv:2112.13339 ,

    arXiv

  20. [28]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps,

    C. Lu, Y. Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps,” NIPS, 2022. 2

    2022

  21. [29]

    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 Int. Conf. Learn. Represent., 2022. 2

    2022

  22. [30]

    Conditional generative adversarial nets,

    M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv preprint arXiv:1411.1784, 2014. 2

    2014 arXiv

  23. [31]

    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 Adv. Neural Inform. Process. Syst., 2014. 2

    2014

  24. [32]

    Semantic image synthesis via adversarial learning,

    H. Dong, S. Yu, C. Wu, and Y. Guo, “Semantic image synthesis via adversarial learning,” in Int. Conf. Comput. Vis. , 2017, pp. 5706–

    2017

  25. [33]

    Generative attribute controller with conditional filtered generative adversarial net- works,

    T. Kaneko, K. Hiramatsu, and K. Kashino, “Generative attribute controller with conditional filtered generative adversarial net- works,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2017, pp. 6089–

    2017

  26. [34]

    Learning to generate images of outdoor scenes from attributes and semantic layouts,

    L. Karacan, Z. Akata, A. Erdem, and E. Erdem, “Learning to generate images of outdoor scenes from attributes and semantic layouts,” arXiv preprint arXiv:1612.00215, 2016. 2

    2016 arXiv

  27. [35]

    Photo- realistic single image super-resolution using a generative adver- sarial network,

    C. Ledig, L. Theis, F. Husz ´ar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang et al. , “Photo- realistic single image super-resolution using a generative adver- sarial network,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2017, pp. 4681–4690. 2

    2017

  28. [36]

    Scribbler: Controlling deep image synthesis with sketch and color,

    P . Sangkloy, J. Lu, C. Fang, F. Yu, and J. Hays, “Scribbler: Controlling deep image synthesis with sketch and color,” in IEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 5400–5409. 2

    2017

  29. [37]

    Generative image modeling using style and structure adversarial networks,

    X. Wang and A. Gupta, “Generative image modeling using style and structure adversarial networks,” in Eur. Conf. Comput. Vis. , 2016, pp. 318–335. 2

    2016

  30. [38]

    Age progression/regression by conditional adversarial autoencoder,

    Z. Zhang, Y. Song, and H. Qi, “Age progression/regression by conditional adversarial autoencoder,” in IEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 5810–5818. 2

    2017

  31. [39]

    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, 2018, pp. 586–595. 4, 7

    2018

  32. [40]

    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. Hochre- iter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in Adv. Neural Inform. Process. Syst., 2017. 4, 7

    2017

  33. [41]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P . Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Trans. Image Process., pp. 600–612, 2004. 4, 7

    2004

  34. [42]

    Quality metric guided portrait line drawing generation from unpaired training data,

    R. Yi, Y.-J. Liu, Y.-K. Lai, and P . L. Rosin, “Quality metric guided portrait line drawing generation from unpaired training data,” IEEE Trans. Pattern Anal. Mach. Intell. , DOI (identifier) 10.1109/TPAMI.2022.3147570, 2022. 6, 7

    2022

  35. [43]

    Stargan v2: Diverse image synthesis for multiple domains,

    Y. Choi, Y. Uh, J. Yoo, and J.-W. Ha, “Stargan v2: Diverse image synthesis for multiple domains,” in IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 8188–8197. 6, 7, 15, 18

    2020

  36. [44]

    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 Int. Conf. Learn. Represent., 2018. 6, 7, 15, 19

    2018

  37. [45]

    Gener- ative visual manipulation on the natural image manifold,

    J.-Y. Zhu, P . Kr ¨ahenb ¨uhl, E. Shechtman, and A. A. Efros, “Gener- ative visual manipulation on the natural image manifold,” in Eur. Conf. Comput. Vis., 2016. 6, 7, 15, 20

    2016

  38. [46]

    Holistically-nested edge detection,

    S. Xie and Z. Tu, “Holistically-nested edge detection,” in Int. Conf. Comput. Vis., 2015, pp. 1395–1403. 6, 7, 15, 20

    2015

  39. [47]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” in Adv. Neural Inform. Process. Syst., 2019. 7

    2019

  40. [48]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted inter- vention, 2015, pp. 234–241. 7 IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INT...

    2015

  41. [49]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Adv. Neural Inform. Process. Syst., 2017. 7 Mengfei Xia received the B.S. degree in 2020 from the Department of Mathematical Science, Tsinghua ...

    2017

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.