Pith. sign in

REVIEW 4 major objections 6 minor 51 references

SSIMBaD: Sigma Scaling with SSIM-Guided Balanced Diffusion for AnimeFace Colorization

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 The paper claims that replacing the usual diffusion noise schedule with a squash function that makes visual structure degrade at a constant rate yields better anime-face colorization than prior reference-based methods.

desk verdict SSIMBaD is a useful empirical schedule tweak for anime colorization, but the causal link between SSIM linearity and the reported gains is not actually established. read the letter →

arxiv 2506.04283 v1 pith:DGHLJNCC submitted 2025-06-04 cs.GR cs.AIcs.CV

Junpyo Seo , Hanbin Koo , Jieun Yook , Byung-Ro Moon (Department of Computer Science , Seoul National University) This is my paper
classification cs.GRcs.AIcs.CV
keywords animefacecolorizationdiffusionmodelsnoiseschedulesigma-spacescalingstructuralsimilarity(SSIM)EDMtrajectoryrefinementreference-guided
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the noise schedule of a diffusion colorizer can be chosen so that visual degradation is spread evenly across the whole generation path, and that this choice is what improves anime face colorization. The proposed SSIMBaD framework selects a transformation $\phi^*(\sigma)=\sigma/(\sigma+0.3)$ by measuring how linearly the structural-similarity index (SSIM) falls as transformed noise grows, then uses that same transformation for training, the noise embedding, trajectory refinement, and inference. On a 31,696-image anime face dataset the authors report that this schedule, together with a trajectory-refinement objective, surpasses the SCFT and AnimeDiffusion baselines on PSNR, MS-SSIM, and FID in both same-reference and cross-reference settings. The significance is that a small, interpretable change to how noise levels are scheduled—not a larger network or more data—is claimed to carry most of the fidelity gain.

What carries the argument

The load-bearing object is the $\sigma$-space scaling function $\phi^*(\sigma)=\sigma/(\sigma+0.3)$, a bounded, monotone squash map that turns noise level $\sigma$ into a perceptual difficulty coordinate. Noise levels are sampled uniformly in $\phi^*$-space and mapped back through $(\phi^*)^{-1}$ to produce the schedule; the same value replaces the log-$\sigma$ embedding $c_{noise}$; and the same schedule defines both the training noise distribution and the Euler/Heun integration steps. The function's job is to equalize the SSIM drop per step, so no part of the trajectory is wasted on nearly-clean or nearly-pure-noise regions. The trajectory-refinement objective then trains the entire reverse path end-to-end against the ground truth.

What would settle it

Train SSIMBaD with a schedule that scores nearly as high on the same R2 criterion—for example $\sigma/(\sigma+0.5)$ or $\sigma/(\sigma+0.7)$, which the paper reports with $R^2$ 0.9275 and 0.8710—and compare PSNR, MS-SSIM, and FID on the same test set; if a clearly less-linear schedule matches or beats $\sigma/(\sigma+0.3)$, the claim that near-linear SSIM degradation is the operative cause fails. Re-running the R2 selection on a different 1% sample of training images and checking whether the top-ranked transformation changes would also test the stability of the schedule choice.

Watch

Extended reading notes

Core claim

The central discovery is that a single monotone squash function $\phi^*(\sigma)=\sigma/(\sigma+0.3)$ makes SSIM degradation nearly linear in the diffusion noise coordinate, and that scheduling noise by this function unifies training and inference. The paper searches a family of fourteen analytic transformations, computes the coefficient of determination ($R^2$) between transformed noise level and SSIM of a Gaussian-corrupted image, and finds that this bounded squash function has the highest linearity ($R^2=0.9793$ in the main table; $0.9949$ reported in the appendix). Using $\phi^*$ to build the $\sigma$ schedule, replacing the usual $\log(\sigma)$ noise embedding with $\phi^*(\sigma)$, and adding an objective that optimizes the whole Euler reverse trajectory yields PSNR 18.92 (same) and 15.84 (cross), MS-SSIM 0.8512 and 0.8207, and FID 34.98 and 37.10, beating both baselines. The authors frame this as aligning forward and reverse dynamics so that every denoising step faces equal perceptual difficulty.

Load-bearing premise

The load-bearing assumption is that making the structural-similarity score fall as evenly as possible across noise levels, measured on 1% of the training images, is what actually drives the reported quality gains; the paper does not show a causal chain from this linearity to the final PSNR, MS-SSIM, and FID numbers.

Editorial extensions

If this is right

  • Perceptually uniform scheduling lets a diffusion model trained for 300 epochs beat the GAN-based SCFT baseline and the pretrained and finetuned AnimeDiffusion variants on PSNR, MS-SSIM, and FID in both same- and cross-reference settings.
  • Using the same $\phi^*$ transformation for training, noise embedding, and inference removes the mismatch between EDM's log-normal training distribution and its power-law inference schedule, which the paper identifies as a source of perceptual misalignment.
  • Trajectory refinement improves all three reported metrics at once, whereas AnimeDiffusion's finetuning improves PSNR and MS-SSIM while worsening FID, suggesting the refinement objective acts differently from a simple finetuning step.
  • An appendix result shows that adding a rotation to the TPS reference warping shrinks the same-versus-cross PSNR gap from 9.1 dB to 3.08 dB, so the cross-reference robustness of the full pipeline depends on this alignment step as well as on the schedule.

Reading between the lines

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

  • An implication the authors leave implicit is that the same $R^2$-based schedule search could be applied to other conditional diffusion tasks where structural fidelity matters, such as super-resolution, sketch-to-image, or medical imaging; the paper's conclusion only gestures at this possibility.
  • Because the $R^2$ selection was computed on a random 1% of training images, the choice of the constant 0.3 is plausibly dataset-dependent; a natural test is to rerun the schedule search on a different dataset and check whether the optimal squash constant shifts.
  • The comparison between $\sigma/(\sigma+0.3)$ and the nearly-as-linear $\sigma/(\sigma+0.5)$ would isolate whether it is linearity itself, or the specific placement of intermediate noise levels, that drives the metric gains; the paper's current evidence cannot distinguish these.
  • The trajectory-refinement gain may transfer to other samplers even without a new schedule, since minimizing the full Euler path can be read as correcting per-step integration error; testing SSIMBaD's refinement objective under the default EDM schedule would separate schedule effects from refinement effects.
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 SSIMBaD, a modification of the EDM continuous-time diffusion framework for reference-guided anime sketch colorization. The core idea is to replace EDM's log-based noise schedule with a sigma-space transformation phi*(sigma) = sigma/(sigma + 0.3), selected by fitting the linearity (R^2) of SSIM degradation under Gaussian noise on 1% of training images (Section 4.1, Eq. 6). This transformation is then used for training-time sigma sampling (Eq. 7), noise embedding c_noise(phi*(sigma)), and inference scheduling (Eq. 10), and it is combined with a trajectory-refinement objective (Eq. 9). The paper reports state-of-the-art PSNR, MS-SSIM, and FID on the AnimeFace dataset under both same-reference and cross-reference settings (Table 2), with a cumulative ablation in Table 3.

Significance. If the central claim were established, the contribution would be notable: a conceptually simple, perceptually motivated noise schedule plus a reverse-trajectory refinement loss would improve structural fidelity and style transfer over strong baselines such as SCFT and AnimeDiffusion. The paper ships code and documents training hyperparameters in Appendix F, which aids reproducibility. The ablation study in Table 3 shows a clear direction of improvement, and the appendix (Appendix D) candidly reports the sensitivity of results to the TPS-rotation augmentation. However, the current evidence does not isolate whether SSIM linearity, the changed sigma distribution, the changed noise embedding, or the trajectory refinement is responsible for the reported gains; the benchmark numbers are single runs without confidence intervals; and the choice of rotation augmentation materially changes the headline metrics. The idea is worth pursuing, but the causal claim needs substantially stronger experimental support.

major comments (4)
  1. [4.1, Appendix B, Eq. (6)-(10)] The paper's load-bearing claim is that maximizing R^2 of SSIM-degradation linearity yields a schedule that improves final colorization quality. This causal link is not demonstrated. Switching from EDM's log-normal sigma sampling to uniform sampling in phi-space simultaneously changes (i) the training distribution over sigma, (ii) the noise embedding c_noise from log(sigma) to phi*(sigma), and (iii) the inference schedule in Eq. (10), while trajectory refinement (Eq. 9) is an additional change. The R^2 fit is computed on only 1% of training images with no confidence intervals, and no alternative transformation (e.g., log(sigma) or arcsinh(sigma)) is trained with the same pipeline to test whether the improvement is caused by SSIM linearity or merely by the new sigma distribution. Without such a control, the attribution of the Table 2/Table 3 gains to SSIM-aligned perceptual scaling is not established.
  2. [Table 3, Section 5.4] The cumulative ablation does not isolate the contribution of individual components. The row labeled '+ SSIM-aligned sigma-space scaling' changes both the training distribution and the noise embedding relative to the EDM row, and the following row adds trajectory refinement; there is no row that adds trajectory refinement to the default EDM schedule without SSIMBaD scaling, and no row that uses SSIMBaD scaling without the EDM backbone. Moreover, the cumulative progression is not monotone: adding EDM worsens MS-SSIM and FID, and adding SSIM-aligned scaling worsens FID further (from 52.18/53.60 to 53.33/55.18) before trajectory refinement improves all metrics. The statement that 'each component contributes' is therefore not supported by the presented data; single-factor ablations are needed.
  3. [Appendix D, Table 4 vs. Table 2] The headline comparison is heavily affected by an auxiliary augmentation choice that is not part of the SSIM-linearity argument. Adding rotation to the TPS reference changes same-reference PSNR from 23.10 to 18.92, cross-reference PSNR from 14.00 to 15.84, same-reference FID from 24.35 to 34.98, and cross-reference FID from 40.73 to 37.10. Thus the claimed superior cross-reference generalization comes at a substantial same-reference cost, and the reported FID improvement on the same-reference setting actually reverses. The paper should report both settings (with and without rotation) in the main text, justify the choice of rotation as part of the method rather than a test-protocol artifact, and avoid claiming state-of-the-art performance on the basis of a single configuration of this augmentation.
  4. [Tables 2 and 3] All quantitative results are reported from a single run, with no confidence intervals, standard deviations, or significance tests. For a method whose central advantage is described as a 3-5 dB PSNR improvement and a large FID reduction, the absence of repeated-run statistics makes it impossible to assess whether the differences are stable or within training variance. At minimum, the authors should report means and standard deviations over multiple seeds, or bootstrap intervals over the 579 test images.
minor comments (6)
  1. [Appendix B vs. Table 1] Appendix B states that phi*(sigma) = sigma/(sigma+0.3) achieves R^2 = 0.9949, while Table 1 reports R^2 = 0.9793 for the same function; the discrepancy should be reconciled.
  2. [Table 1] The table header contains malformed expressions ('log 1 p(σ)', 'σ p /σ p +1') that appear to be a LaTeX rendering error; the definitions of all candidate transformations should be readable and unambiguous.
  3. [Appendix A, Eq. (15)] The EDM noise embedding is written as c_noise = 1/4 ln σ in the appendix, while the main text repeatedly refers to the 'conventional log(σ) noise embedding'; please clarify which definition is used and whether c_noise is the raw embedding or a scaled version.
  4. [Section 5.3.2] There is a duplicated word in 'improving reconstruction reconstruction fidelity' and the text uses both 'SSIMBaD' and 'SSIMBad'; please standardize the spelling.
  5. [Abstract and Introduction] The phrase 'large-scale Anime face dataset' is only qualified in Section 5.1 as a 31,696-image split from AnimeDiffusion; please state the dataset provenance and split clearly in the abstract or introduction for readers who do not reach Section 5.
  6. [Section 5.3.3 and Figure 3] The claim that the proposed schedule yields 'smooth and balanced noise at every step' is illustrated on a single image with N=25 or N=50 steps; a quantitative distribution of SSIM slopes across many images would better support the uniformity claim.

Circularity Check

1 steps flagged · score 6.0 of 10

The central 'near-linear SSIM degradation' property of phi* is the exact objective used to select phi* in Eq. (6), so the paper's perceptual-linearity claim is circular; the downstream colorization metrics remain independent evidence.

  1. fitted input called prediction [Section 4.1, Eq. (6), echoed in Abstract and Section 5.3.1/Table 1]
    "Our empirical search reveals that ϕ∗(σ) = σ σ+0.3 yields the highest R2 and near-linear SSIM degradation. We adopt this transformation consistently in both training and inference, unifying the sampling dynamics across the diffusion process."

    Equation (6) defines ϕ* as the arg max over Φ of the expected R² between the ϕ-scheduled noise levels σϕ_i and SSIM(Igt+σϕ_i·n, Igt). The immediately following sentence reports that this ϕ* 'yields the highest R2 and near-linear SSIM degradation'; that is the same criterion used to select it, so the property is guaranteed (among the candidate set) by construction rather than independently predicted. The Abstract converts this fitted property into the paper's core claim: 'This scaling ensures uniform visual difficulty across timesteps.' Since the R² search was run on 1% of training images, the near-linear degradation claim is a fit to that subset, not an external prediction. The Table 2 PSNR/MS-SSIM/FID numbers are not forced by the R² criterion, so the circularity is partial, not total.

full rationale

The paper's derivation chain has one load-bearing circular link: the schedule's headline property, linear perceptual degradation, is not a discovered consequence but the selection objective. Eq. (6) chooses ϕ* by maximizing R² of SSIM degradation linearity on 1% of training images; the text then reports that ϕ*(σ)=σ/(σ+0.3) has the highest R² and near-linear behavior, and the abstract states the scaling 'ensures uniform visual difficulty across timesteps.' This is fitted-input-called-prediction: the 'prediction' is the fit criterion restated. However, the paper's SOTA claims rest mainly on Table 2 and Table 3 metrics, which are external benchmarks and not implied by the R² maximization; a lower-R² schedule could in principle still perform well, so those numbers are independent evidence. No load-bearing self-citation or imported-uniqueness pattern was found; the paper does not cite itself for its central premise. The absence of a control schedule with lower R² under otherwise identical training and refinement is a confound that weakens the causal attribution to SSIM linearity, but that is a correctness limitation rather than circularity. Overall score 6: one central perceptual claim reduces by construction, while the empirical colorization gains are not circular.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one fitted scalar and on the assumption that SSIM linearity is the right perceptual objective. No new physical or architectural entities are introduced.

free parameters (1)
  • c in phi*(sigma) = sigma/(sigma + c) = 0.3
    Chosen by empirical search over candidate transformations to maximize R2 of SSIM degradation linearity on a 1% sample of training images (Section 4.1, Eq. 6; Appendix B). The schedule used for training, refinement, and inference depends on this value.
assumptions (4)
  • domain assumption SSIM is a valid and sufficient proxy for perceptual degradation, and linear SSIM degradation across timesteps is the property that improves colorization.
    The schedule is selected by Eq. 6 to maximize SSIM linearity, but no independent evidence links this property to PSNR, MS-SSIM, or FID improvements.
  • ad hoc to paper The hand-picked candidate set Phi of 13 transformations contains the optimal schedule.
    Section 4.1 restricts the search to listed analytic and squash functions with no argument that the true optimum lies in this family.
  • domain assumption EDM's preconditioning coefficients and constants such as sigma_data=0.5 transfer unchanged from unconditional image synthesis to reference-guided colorization.
    Appendix A adopts EDM constants directly from Karras et al. [11] without task-specific rederivation.
  • domain assumption Sampling sigma so that phi*(sigma) is uniform during training is the correct way to balance noise levels.
    Section 4.2 replaces EDM's log-normal sampling with uniform phi-space sampling and uses the same schedule in trajectory refinement; the benefit is shown only empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SSIMBaD: Sigma Scaling with SSIM-Guided Balanced Diffusion for AnimeFace Colorization." pith.science (2026). https://pith.science/paper/DGHLJNCC

@misc{pith2026250604283,
  author       = {Pith},
  title        = {Pith review of: SSIMBaD: Sigma Scaling with SSIM-Guided Balanced Diffusion for AnimeFace Colorization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DGHLJNCC}},
  note         = {Machine review of arXiv:2506.04283}
}
read the original abstract

We propose a novel diffusion-based framework for automatic colorization of Anime-style facial sketches. Our method preserves the structural fidelity of the input sketch while effectively transferring stylistic attributes from a reference image. Unlike traditional approaches that rely on predefined noise schedules - which often compromise perceptual consistency -- our framework builds on continuous-time diffusion models and introduces SSIMBaD (Sigma Scaling with SSIM-Guided Balanced Diffusion). SSIMBaD applies a sigma-space transformation that aligns perceptual degradation, as measured by structural similarity (SSIM), in a linear manner. This scaling ensures uniform visual difficulty across timesteps, enabling more balanced and faithful reconstructions. Experiments on a large-scale Anime face dataset demonstrate that our method outperforms state-of-the-art models in both pixel accuracy and perceptual quality, while generalizing to diverse styles. Code is available at github.com/Giventicket/SSIMBaD-Sigma-Scaling-with-SSIM-Guided-Balanced-Diffusion-for-AnimeFace-Colorization

Figures

Figures reproduced from arXiv: 2506.04283 by the authors.

Figure 1
Figure 1. Qualitative comparison of colorization results under the same-reference scenario. From left to right: (a) Sketch input. (b) Reference image. (c) SCFT [29]. (d) AnimeDiffusion [35] (pretrained). (e) AnimeDiffusion [35] (finetuned). (f) AnimeDiffusion (EDM backbone, default σ-schedule). (g) Our model (w/o trajectory refinement). (h) Our model (w/ trajectory refinement). (a) (b) (c) (d) (e) (f) (g) (h) [PITH_FULL_IMAG… view at source ↗
Figure 2
Figure 2. Qualitative comparison of colorization results under the cross-reference scenario. (a) Sketch input. (b) Reference image. (c) SCFT [29]. (d) AnimeDiffusion [35] (pretrained). (e) AnimeDiffusion [35] (finetuned). (f) AnimeDiffusion (EDM backbone, default σ-schedule). (g) Our model (w/o trajectory refinement). (h) Our model (w/ trajectory refinement) [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison of forward diffusion schedules. Top: SSIM curves for DDPM (a), EDM (b), and our schedule ϕ ∗ (σ) (c). Bottom: 5 × 5 corrupted grids (d)–(f) show each schedule’s visual effect. Our method yields perceptually uniform degradation across timesteps. 6 Conclusion This study proposes a novel framework, SSIMBaD, to address the perceptual inconsistency that has been overlooked in conventional diffusion-based model… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: SSIM degradation across various transformations ϕ(σ). Each curve shows the SSIM between the clean image and its noisy counterpart as the noise level σ increases, under a specific transformation ϕ. The transformations are ordered by increasing linearity (R2 ). Among the…
Figure 5
Figure 5. Figure 5: Qualitative comparison under the same-reference scenario. (a) Sketch input. (b) Reference image. (c) SCFT [29]. (d) AnimeDiffusion [35] (pretrained). (e) AnimeDiffusion [35] (finetuned). (f) AnimeDiffusion (EDM backbone, default σ-schedule). (g) Our model (w/o trajecto…
Figure 6
Figure 6. Figure 6: Qualitative comparison under the cross-reference scenario. (a) Sketch input. (b) Reference image. (c) SCFT [29]. (d) AnimeDiffusion [35] (pretrained). (e) AnimeDiffusion [35] (finetuned). (f) AnimeDiffusion (EDM backbone, default σ-schedule). (g) Our model (w/o traject…
Figure 7
Figure 7. Figure 7: SSIM vs ϕ(σ) curves for the same input image under forward (corruption, red) and reverse (generation, blue) processes. Finetuning improves perceptual linearity in certain regions, but quickly saturates due to existing generation dynamics. The model nonetheless maintain…
Figure 8
Figure 8. Figure 8: Comparison under same- and cross-reference scenarios without TPS rotation. (a) Sketch input. (b) Reference image (same style). (c) Reference image (cross style). (d–e) Our model under same-reference scenario (w/o and w/ trajectory refinement, no TPS rotation). (f–g) Ou…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 48 canonical work pages

  1. [1]

    XDoG: An extended difference-of- Gaussians compendium including advanced image stylization,

    H. Winnemöller, J. E. Kyprianidis, and S. C. Olsen, “XDoG: An extended difference-of- Gaussians compendium including advanced image stylization,” Computers & Graphics, vol. 36, no. 6, pp. 740–753, 2012

    work page 2012

  2. [2]

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics,

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep Unsupervised Learning using Nonequilibrium Thermodynamics,” in Proceedings of the 32nd International Conference on Machine Learning (ICML) , 2015

    work page 2015

  3. [3]

    Denoising Diffusion Probabilistic Models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising Diffusion Probabilistic Models,” in Advances in Neural Information Processing Systems (NeurIPS) , 2020

    work page 2020

  4. [4]

    Generative Modeling by Estimating Gradients of the Data Distribution,

    Y . Song and S. Ermon, “Generative Modeling by Estimating Gradients of the Data Distribution,” in Advances in Neural Information Processing Systems (NeurIPS) , 2019

    work page 2019

  5. [5]

    Score-Based Gen- erative Modeling through Stochastic Differential Equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-Based Gen- erative Modeling through Stochastic Differential Equations,” inProceedings of the International Conference on Learning Representations (ICLR) , 2021

    work page 2021

  6. [6]

    Denoising Diffusion Implicit Models,

    J. Song, C. Meng, and S. Ermon, “Denoising Diffusion Implicit Models,” in Proceedings of the International Conference on Learning Representations (ICLR) , 2021

    work page 2021

  7. [7]

    Improved Denoising Diffusion Probabilistic Models,

    A. Q. Nichol and P. Dhariwal, “Improved Denoising Diffusion Probabilistic Models,” CoRR, vol. abs/2102.09672, 2021

    arXiv 2021

  8. [8]

    Diffusion Models Beat GANs on Image Synthesis,

    P. Dhariwal and A. Nichol, “Diffusion Models Beat GANs on Image Synthesis,” in Advances in Neural Information Processing Systems (NeurIPS) , 2021

    work page 2021

Show all 51 references
  1. [9]

    Variational Diffusion Models,

    D. P. Kingma, T. Salimans, B. Poole, and J. Ho, “Variational Diffusion Models,” inAdvances in Neural Information Processing Systems (NeurIPS) , 2021

    2021

  2. [10]

    Improved Techniques for Training Score-Based Generative Models,

    Y . Song and S. Ermon, “Improved Techniques for Training Score-Based Generative Models,” in Advances in Neural Information Processing Systems (NeurIPS) , 2020

    2020

  3. [11]

    Elucidating the Design Space of Diffusion-Based Generative Models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the Design Space of Diffusion-Based Generative Models,” in Advances in Neural Information Processing Systems , vol. 35, pp. 26565–26577, 2022

    2022

  4. [12]

    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 Recognition (CVPR), 2022

    2022

  5. [13]

    Progressive Distillation for Fast Sampling of Diffusion Models,

    T. Salimans and J. Ho, “Progressive Distillation for Fast Sampling of Diffusion Models,” in Proceedings of the International Conference on Learning Representations (ICLR) , 2022

    2022

  6. [14]

    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 Proceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  7. [15]

    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,” inAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  8. [17]

    GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models,

    A. Q. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. McGrew, I. Sutskever, and M. Chen, “GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models,” in Proceedings of the 39th International Conference on Machine Learning (ICML), 2022. 10

    2022

  9. [18]

    SDEdit: Guided Image Syn- thesis and Editing with Stochastic Differential Equations,

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.-Y . Zhu, and S. Ermon, “SDEdit: Guided Image Syn- thesis and Editing with Stochastic Differential Equations,” in Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  10. [19]

    Paint by Example: Exemplar-Based Image Editing with Diffusion Models,

    B. Yang, S. Gu, B. Zhang, T. Zhang, X. Chen, X. Sun, D. Chen, and F. Wen, “Paint by Example: Exemplar-Based Image Editing with Diffusion Models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2023

    2023

  11. [21]

    T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models,

    C. Mou, X. Wang, L. Xie, J. Zhang, Z. Qi, Y . Shan, and X. Qie, “T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models,” CoRR, vol. abs/2302.08453, 2023

    2023 arXiv

  12. [22]

    The Perception-Distortion Tradeoff in Image Restoration,

    Y . Blau and T. Michaeli, “The Perception-Distortion Tradeoff in Image Restoration,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  13. [23]

    Perceptual Losses for Real-Time Style Transfer and Super-Resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual Losses for Real-Time Style Transfer and Super-Resolution,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2016

    2016

  14. [24]

    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/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  15. [25]

    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. on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  16. [26]

    User-Guided Deep Anime Line Art Colorization with Conditional Adversarial Networks,

    Y . Ci, X. Ma, Z. Wang, H. Li, and Z. Luo, “User-Guided Deep Anime Line Art Colorization with Conditional Adversarial Networks,” in Proceedings of the 26th ACM International Conference on Multimedia (ACM MM), 2018

    2018

  17. [27]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  18. [28]

    Comicolorization: Semi-Automatic Manga Colorization,

    C. Furusawa, K. Hiroshiba, K. Ogaki, and Y . Odagiri, “Comicolorization: Semi-Automatic Manga Colorization,” in ACM SIGGRAPH Asia Technical Briefs, 2017

    2017

  19. [29]

    Reference-Based Sketch Image Colorization Using Augmented-Self Reference and Dense Semantic Correspondence,

    J. Lee, E. Kim, Y . Lee, D. Kim, J. Chang, and J. Choo, “Reference-Based Sketch Image Colorization Using Augmented-Self Reference and Dense Semantic Correspondence,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  20. [30]

    Tag2Pix: Line Art Colorization Using Text Tag with SECat and Changing Loss,

    H. Kim, H. Y . Jhoo, E. Park, and S. Yoo, “Tag2Pix: Line Art Colorization Using Text Tag with SECat and Changing Loss,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019

    2019

  21. [31]

    Two-Stage Sketch Colorization,

    L. Zhang, C. Li, T.-T. Wong, Y . Ji, and C. Liu, “Two-Stage Sketch Colorization,”ACM Trans. Graph. (Proc. SIGGRAPH Asia), vol. 37, no. 6, pp. 261:1–261:14, 2018

    2018

  22. [32]

    User-Guided Line Art Flat Filling with Split Filling Mechanism,

    L. Zhang, C. Li, E. Simo-Serra, Y . Ji, and C. Liu, “User-Guided Line Art Flat Filling with Split Filling Mechanism,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  23. [33]

    Eliminating Gradient Conflict in Reference- Based Line-Art Colorization,

    Z. Li, Z. Geng, Z. Kang, W. Chen, and Y . Yang, “Eliminating Gradient Conflict in Reference- Based Line-Art Colorization,” in Proceedings of the European Conference on Computer Vision (ECCV), 2022

    2022

  24. [34]

    Diffusart: Enhancing Line Art Colorization with Conditional Diffusion Models,

    H. Carrillo, M. Clément, A. Bugeau, and E. Simo-Serra, “Diffusart: Enhancing Line Art Colorization with Conditional Diffusion Models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 11

    2023

  25. [35]

    AnimeDiffusion: Anime Face Line Drawing Colorization via Diffusion Models,

    Y . Cao, X. Meng, P. Y . Mok, X. Liu, T.-Y . Lee, and P. Li, “AnimeDiffusion: Anime Face Line Drawing Colorization via Diffusion Models,” IEEE Trans. Vis. Comput. Graphics, 2024

    2024

  26. [36]

    Style-Structure Disentangled Features and Normalizing Flows for Diverse Icon Colorization,

    Y .-K. Li, Y .-H. Lien, and Y .-S. Wang, “Style-Structure Disentangled Features and Normalizing Flows for Diverse Icon Colorization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  27. [37]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  28. [38]

    Multimodal Unsupervised Image-to-Image Translation,

    X. Huang, M.-Y . Liu, S. Belongie, and J. Kautz, “Multimodal Unsupervised Image-to-Image Translation,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2018

    2018

  29. [39]

    Diverse Image-to-Image Translation via Disentangled Representations,

    H.-Y . Lee, H.-Y . Tseng, J.-B. Huang, M. K. Singh, and M.-H. Yang, “Diverse Image-to-Image Translation via Disentangled Representations,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018

    2018

  30. [40]

    Deep Exemplar-Based Colorization,

    M. He, D. Chen, J. Liao, P. V . Sander, and L. Yuan, “Deep Exemplar-Based Colorization,”ACM Trans. Graph., vol. 37, no. 4, pp. 47:1–47:16, 2018

    2018

  31. [41]

    Stylization-Based Architecture for Fast Deep Exemplar Colorization,

    Z. Xu, T. Wang, F. Fang, Y . Sheng, and G. Zhang, “Stylization-Based Architecture for Fast Deep Exemplar Colorization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  32. [42]

    Image Style Transfer Using Convolutional Neural Networks,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “Image Style Transfer Using Convolutional Neural Networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  33. [43]

    Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization,

    X. Huang and S. Belongie, “Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017

    2017

  34. [44]

    Consistency Models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency Models,” inProceedings of the 40th International Conference on Machine Learning , vol. 202, pp. 32211–32252, 2023

    2023

  35. [45]

    Instance-Aware Image Colorization,

    J.-W. Su, H.-K. Chu, and J.-B. Huang, “Instance-Aware Image Colorization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

    2020

  36. [46]

    RePaint: Inpainting using Denoising Diffusion Probabilistic Models,

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, “RePaint: Inpainting using Denoising Diffusion Probabilistic Models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022

    2022

  37. [47]

    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,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  38. [48]

    Multi-scale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multi-scale structural similarity for image quality assessment,” in The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers , vol. 2, pp. 1398–1402, 2003

    2003

  39. [49]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4401–4410, 2019

    2019

  40. [50]

    ImageNet: A large-scale hierarchical image database,

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

    2009

  41. [51]

    Diffusart: Enhancing Line Art Colorization with Conditional Diffusion Models,

    H. Carrillo, M. Clément, A. Bugeau, and E. Simo-Serra, “Diffusart: Enhancing Line Art Colorization with Conditional Diffusion Models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 12

    2023

  42. [52]

    ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models,

    J. Choi, S. Kim, Y . Jeong, Y . Gwon, and S. Yoon, “ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    2021

  43. [53]

    Adding Conditional Control to Pretrained Diffusion Models,

    L. Zhang and M. Agrawala, “Adding Conditional Control to Pretrained Diffusion Models,” CoRR, vol. abs/2302.05543, 2023. 13 Appendix A Details of the Proposed Framework A.1 Conditional Input Construction Let Igt ∈ RH×W ×3 denote the RGB ground-truth anime image, where H and W a...

    2023 arXiv

Pith tools

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