Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Advancing Facial Stylization through Semantic Preservation Constraint and Pseudo-Paired Supervision

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning StyleGAN with a semantic preservation constraint and pseudo-paired supervision keeps the source face's identity, pose, and expression visible while learning a target style, and the same generator then gives multimodal and…

desk verdict A clearly-specified StyleGAN fine-tuning method with consistent ablation gains; the main soft spot is the unvalidated alignment of pseudo-pairs, which is fixable. read the letter →

arxiv 2506.22022 v2 pith:4SLVUMCS submitted 2025-06-27 cs.CV

Zhanyi Lu , Yue Zhou This is my paper
classification cs.CV
keywords facialstylizationStyleGANinversionsemanticpreservationpseudo-pairedsupervisionperceptuallossreference-guidedmultimodal
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 argues that artifacts and content mismatches in StyleGAN-based facial stylization come from a semantic shift in the generator's latent space during fine-tuning: when the model adapts to a small style dataset, the latent distribution changes, and the meaning of a given latent code drifts away from the pretrained model's semantics. To stop that drift, the paper adds two losses to fine-tuning: a semantic preservation constraint that compares the pretrained and fine-tuned generators on the same noise, and pseudo-paired supervision that asks the fine-tuned generator to produce a stylized image from the latent code of a matching pseudo-real portrait. It also shows how to build multi-level pseudo-paired data from unpaired style images, and how the same fine-tuned generator can support multimodal and reference-guided stylization without extra networks. If the claim is right, StyleGAN stylization on small style collections can preserve identity and pose while still learning the target look.

What carries the argument

The load-bearing mechanism is a pair of losses applied while fine-tuning a pretrained StyleGAN. The semantic preservation constraint, $L_{\text{semantic}} = L_{\text{LPIPS}}(G(z), G'(z)) + \lambda_{\text{ID}} L_{\text{ID}}(G(z), G'(z))$, pins the fine-tuned generator $G'$ to the pretrained $G$ by penalizing perceptual and identity differences on the same random noise $z$; here LPIPS is a perceptual distance on deep image features and $L_{\text{ID}}$ is an identity-recognition loss. The pseudo-paired supervision, $L_{\text{paired}} = L_{\text{LPIPS}}(G'(w^{+}), S)$, pushes $G'$ to map the shared code $w^{+}$ to the stylized image instead of drifting to a different semantic. The pseudo-pairs themselves are built in three stages: a pSp encoder embeds the style image $S$ into $Z^{+}$ space to get a first-level code, that code is optimized against the semantic constraint in an unconstrained fine-tuned generator, and a $W^{+}$-space pSp encoder refines the result, so each style image yields a pseudo-real portrait $P$ and a shared code $w^{+}$. Together the two constraints keep the latent distribution from collapsing and keep the semantic directions of the fine-tuned generator aligned with those of the pretrained one.

What would settle it

Take a set of style images $S$, run the paper's three-step pipeline to get $w^{+}$ and $P = G(w^{+})$, and compute identity distance (ArcFace) and perceptual distance between $S$ and $P$; if for a given style these distances are no better than the distances to random real portraits, then the pseudo-pairs are not content-aligned and the paired-supervision term cannot be the mechanism producing the reported fidelity gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that residual artifacts and fidelity loss in StyleGAN-based facial stylization are caused by a semantic shift of the latent space during fine-tuning, and that this shift can be corrected by two constraints. The semantic preservation constraint, $L_{\text{semantic}} = L_{\text{LPIPS}}(G(z), G'(z)) + \lambda_{\text{ID}} L_{\text{ID}}(G(z), G'(z))$, keeps the fine-tuned generator $G'$ close to the pretrained generator $G$ on the same sampled noise; the pseudo-paired supervision, $L_{\text{paired}} = L_{\text{LPIPS}}(G'(w^{+}), S)$, teaches $G'$ to decode the shared latent code $w^{+}$ into the stylized image $S$ so that content correspondence is preserved between real and stylized domains. The paper further constructs multi-level pseudo-paired data by embedding the style image into $Z^{+}$ space, optimizing the code with semantic constraints, and refining it in $W^{+}$ space, so that stylized images can be paired with pseudo-real portraits even though no real paired dataset exists. Across five style datasets (cartoon, anime, fantasy, illustration, impasto), the method is shown to lower FID and perceptual distance relative to its baseline and to allow multimodal and reference-guided stylization from the same fine-tuned generator.

Load-bearing premise

The entire pseudo-paired supervision scheme assumes that a stylized image $S$ and the pseudo-real portrait $P$ produced from a shared latent code $w^{+}$ actually show the same person, pose, and content; if the inversion of $S$ is imperfect, the paired loss trains the generator to map mismatched content and the reported fidelity gain collapses.

Editorial extensions

If this is right

  • Stylized outputs preserve facial structure, orientation, glasses, and identity while still applying the target style, because the same latent code now means the same face in both domains.
  • Fine-tuning remains feasible on small style datasets: the paper uses between 120 and 317 style images per style and reports convergence in roughly 1000 to 3000 iterations, about 0.5 hours per style at 1024 resolution.
  • Because the constraints keep the latent space semantically aligned, the same fine-tuned generator can be reused for general, multimodal, and reference-guided stylization by mixing content codes with random noise or reference embeddings at different style-mixing layers.
  • The constraints reduce the semantic distance between the pretrained and fine-tuned generators, which the paper measures as a lower FID and a smaller semantic separation on the style datasets.
  • The approach avoids additional network design or training for multimodal and reference-guided modes, since those modes reuse the same fine-tuned generator with different latent encodings.

Reading between the lines

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

  • The paper does not isolate how much of the gain comes from each pseudo-pair level; an extension would ablate the $W^{+}_1$, $W^{+}_2$, and $W^{+}_3$ supervision levels per style, since the appendix already shows style-dependent preferences.
  • The same two-loss recipe could plausibly transfer to other base generators or to non-face domains, such as full-body stylization or product rendering, whenever a pretrained generator and a small style collection exist; this is a testable extrapolation rather than a claim in the paper.
  • The pseudo-pair premise can be checked directly by measuring identity and perceptual consistency between $S$ and $P$ before training; if that consistency is low on a given style set, the paired loss may be learning a spurious mapping, and the paper's manual choice of supervision level could be replaced by a data-driven threshold.
  • Because the paper applies the constraints only to the generator, an end-to-end variant that also updates the encoder might reduce the local artifacts noted for anime style; the authors themselves flag this direction as future work.
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 a StyleGAN fine-tuning approach for facial stylization that augments the standard adversarial loss with two additional terms: a semantic preservation constraint (LPIPS and identity loss between the outputs of the pretrained generator G and the finetuned generator G' on the same noise input, Eq. 1) and a pseudo-paired supervision loss (LPIPS between G'(w+) and a stylized image S, Eq. 2), using pseudo-pairs (P,S) constructed through a three-level latent-space pipeline (embedding, optimization, refinement; Section 3.2). The authors claim that these constraints reduce semantic shift in the latent space during fine-tuning, yielding higher quality and more faithful stylization, and they further show that the framework supports multimodal and reference-guided stylization without additional network design or training. Experiments cover five styles (cartoon, anime, fantasy, illustration, impasto) and compare against GAN-based and diffusion-based baselines using FID, perceptual loss, semantic distance, user studies, and runtime metrics.

Significance. If the pseudo-pair alignment assumption is validated, the paper offers a simple and practical recipe for StyleGAN-based facial style transfer with notable FID improvements across multiple styles (Table 4), and the multi-level pseudo-pair generation idea is a useful contribution that goes beyond standard pair construction. The method's compatibility with multimodal and reference-guided stylization without extra training is attractive and well illustrated. However, the central novelty rests on two load-bearing elements that need stronger evidence: the reliability of pseudo-pair content alignment, and the independence of the semantic-preservation evaluation from the training objective. The paper does provide consistent ablations and a user study, which are strengths, but the fidelity claim is not uniformly supported by the reported metrics.

major comments (3)
  1. [Section 3.2, Eq. (2)] The pseudo-paired supervision assumes that P_i and S generated from a common latent code w+ share content (identity, pose, expression). The paper explicitly concedes that 'current real-domain encoders are imperfect in content embedding,' yet it provides no quantitative validation that the three constructed levels P1, P2, P3 actually preserve content correspondence with S. Without such a check, the loss L_paired = LPIPS(G'(w+), S) may train G' to map w+ to a stylized image of different content, so the reported FID/perceptual gains cannot be attributed to content preservation. Please add an alignment evaluation (e.g., ArcFace identity distance, facial landmark/pose distance, or a user study on pair correspondence) for each style and pseudo-pair level, and discuss how mismatched pairs are handled.
  2. [Section 4.3, Table 4, and Eq. (1)] The 'Dis.' metric used to demonstrate semantic preservation is the same family of perceptual/semantic distances that the semantic loss in Eq. (1) directly minimizes (LPIPS between G and G' outputs on the same noise). Therefore the reduction of Dis. in Table 4 is partly tautological; the independent evidence for quality is the FID column, which does improve. To support the claim that the constraints make G' semantically closer to G, please report a metric not used as a training objective (e.g., ArcFace identity distance on held-out z, or a latent semantic-direction consistency measure), or explicitly state that Dis. is a check on the training objective rather than an independent evaluation. This does not invalidate the FID gains, but it weakens the semantic-preservation conclusion as currently worded.
  3. [Section 4.2, Tables 1 and 2] The comparison mixes protocols: UI2I-style and DualStyleGAN are reference/exemplar-based methods, while U-gat-it, Toonify, NTC, InST, and StyleID are used in a non-reference or image-level setting. The paper reports both 'Ours' and 'Ours (ref)', but the headline claim 'our method ... surpasses previous methods' does not respect this protocol distinction; the fair comparison for reference-guided methods is Ours(ref) against UI2I-style/DualStyleGAN, and for non-reference methods Ours against the others. Moreover, the user fidelity scores in Table 2 show StyleID (4.3 for cartoon) and InST (4.0) at or above Ours (4.0), which conflicts with the abstract's claim of 'more faithful' stylization; the text's assertion that diffusion methods do not exhibit sufficient stylization effects is not backed by an objective stylization-strength metric. Please provide protocol-consistent tables and an explicit quality-fidelity tradeoff analysis.
minor comments (5)
  1. [Eq. (1) and Eq. (3)] The notation in Eq. (1) writes L_LPIPS(G(z) - G'(z)) and L_ID(G(z) - G'(z)), using a difference of images where a pair argument is intended; it should read L_LPIPS(G(z), G'(z)) and L_ID(G(z), G'(z)). Also, 'λsematic' in Eq. (3) is a typo for 'λsemantic'.
  2. [Section 3.2, Eq. (5)] The symbol Lsemantic in Eq. (5) is reused for a different quantity than in Eq. (1): here it compares G*(z+) with the style image S, not G(z) with G'(z). Please clarify notation to avoid confusion between the semantic preservation loss and the optimization objective used for pseudo-pair generation.
  3. [Section 3.2, Step 3 (Eq. (6))] The text says 'we refine P2 using a pSp encoder pretrained in W+ space', but Eq. (6) encodes the original style image S (E_w+(S)), not P2, and then generates P3 = G(w+3). This discrepancy between the textual description and the equation should be resolved.
  4. [Table 3] The row for DualStyleGAN cites reference [22], but the method is Pastiche Master, which is reference [11] in the bibliography; the citation likely should be [11].
  5. [Section 4.1 / Appendix 6.1] The paper reports per-style truncation values and per-style pseudo-pair latent levels (e.g., W+3 for cartoon/impasto, W+1 for anime) without describing how these were selected on a validation set. Since these are additional free parameters, please state the selection protocol (e.g., hold-out validation based on FID or visual inspection) to avoid test-set overfitting concerns.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the ablation's 'semantic distance' between G and G′ is the same output-distance objective that Eq. (1) minimizes; FID/user-study evidence keeps the central stylization claim independent.

  1. other [Section 3.1, Eq. (1); Section 4.3, Table 4]
    "Lsemantic(G(z), G′(z)) = LLPIPS(G(z) − G′(z)) + λIDLID(G(z) − G′(z)). (1) ... To further quantify the effects of proposed constraints, we computed the semantic distance [9] between G and G′, as well as the FID score to corresponding dataset. As shown in Table 4, the results not only demonstrate that our improvements enhance the stylization effect but also indicate that the proposed constraints make the fine-tuned models semantically closer to the pre-trained model."

    The 'semantic distance' Dis. reported in Table 4 is presented as evidence that the constraints make the finetuned model semantically closer to the pretrained model. But Eq. (1) directly trains the finetuned model to minimize a perceptual-plus-identity distance between the outputs of G and G′ on the same latent input. The text introducing Eq. (1) describes it as 'comparing the outputs of the pretrained and finetuned models using the same noise input,' which is precisely the quantity later reported as Dis. in the ablation. Thus the observed drop in Dis. from baseline to +SC is largely a consequence of optimizing that same objective, not an independent confirmation that semantics were preserved.

full rationale

The paper's central stylization claim is supported by FID scores, user studies, and qualitative comparisons against external baselines, none of which reduce to the training losses by construction. The pseudo-paired supervision pipeline (Section 3.2) does rely on an unvalidated alignment assumption between inverted codes and style images, but that is a correctness risk rather than a circular derivation: the paper never defines the pseudo-pairs in terms of the final evaluation metric, and the L_paired loss is a genuine training objective rather than a renamed measurement. There are no load-bearing self-citations: the authors cite external work (pSp, UI2I-Style, LPIPS, ArcFace) and do not invoke their own prior results as the sole justification for the approach. The one concrete circular step is in the ablation study, where the 'semantic distance' between G and G′ is used as evidence for semantic preservation even though Eq. (1) is explicitly trained to minimize that same family of distances. That step is partial and does not invalidate the independent FID and user-study evidence, hence the score of 4 rather than higher.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The ledger shows that the central claim rests on domain assumptions about latent-space semantics, encoder transfer, and metric validity rather than on derived guarantees. Six free parameters enter the reported results, and the per-style truncation and pseudo-pair latent level are selected against the reported FID, so the numbers include test-metric tuning. The 'semantic shift' hypothesis is the one invented construction; its measured support is partly circular because the measurement metric matches the training loss. The independent evidence for the central claim is the FID comparison and the user study, which are not part of the training objectives.

free parameters (6)
  • lambda_semantic = 1
    Weight of the semantic preservation loss; searched from 0.001 to 10 in 10x steps and set to the value judged optimal (Appendix 6.1).
  • lambda_paired = 1
    Weight of pseudo-paired supervision; searched from 0 to 5 in 0.5 steps and set to 1 (Appendix 6.1).
  • lambda_ID (identity loss ratio) = 0.1
    Fixed ratio of LPIPS to identity loss inside the semantic constraint (Appendix 6.1).
  • Per-style truncation value = 0.7 cartoon, 0.6 anime, 0.9 others
    Chosen per style to minimize artifacts or enhance fidelity; directly sets test-time outputs (Appendix 6.1).
  • Per-style pseudo-pair latent level = W+1 for anime, W+2 for other styles
    Selected per style based on the FID comparison in Appendix 6.4 (Table 5), i.e., tuned against the evaluation metric.
  • Finetuning iteration count = 1000 per style, 3000 for anime
    Convergence point chosen per style; controls the degree of stylization (Appendix 6.1).
assumptions (5)
  • domain assumption LPIPS and ArcFace identity distance capture the portrait semantics (identity, pose, glasses, hairstyle) that must be preserved during stylization.
    Eq. (1) defines the semantic preservation constraint through these two metrics; if they miss relevant semantics, the constraint preserves the wrong content.
  • domain assumption A common latent code w+ should generate content-corresponding real and stylized portraits; the real and stylized domains are aligned in latent space.
    Stated explicitly in Section 3.2 as the premise for pseudo-paired supervision; if false, the pseudo-pairs are misaligned and Eq. (2) teaches spurious mappings.
  • domain assumption The unconstrained finetuned generator G* can express stylized images well enough that optimizing z+ against the semantic loss yields a meaningful pseudo-real code.
    Step 2 of Section 3.2 optimizes inside G*; a poorly adapted G* would produce misleading codes for the pseudo-pairs.
  • domain assumption Pretrained pSp encoders, modified for W and Z+ spaces, transfer to out-of-domain stylized images S.
    Section 3.2 and Appendix 6.2 rely on encoders trained on FFHQ-like real faces to embed cartoon, anime, and impasto images; the paper concedes these encoders are imperfect.
  • ad hoc to paper Perceptual distance between G(z) and G'(z) over random z measures the latent 'semantic shift' and tracks stylization quality.
    Table 4 uses the semantic distance ('Dis.', from [9]) as evidence for the mechanistic story, but Eq. (1) directly minimizes this LPIPS family, so the evidence is partly by construction.
invented entities (1)
  • Latent 'semantic shift' of the StyleGAN W-space during finetuning
    purpose: Postulated cause of artifacts and fidelity loss in prior StyleGAN stylization methods; motivates both proposed constraints
    The only quantitative evidence (Table 4, 'Dis.') uses a perceptual-semantic distance in the same family as the loss Eq. (1) designed to reduce it; the paper does not measure latent directions or distributions directly, so independent evidence is weak.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Facial Stylization through Semantic Preservation Constraint and Pseudo-Paired Supervision." pith.science (2026). https://pith.science/paper/4SLVUMCS

@misc{pith2026250622022,
  author       = {Pith},
  title        = {Pith review of: Advancing Facial Stylization through Semantic Preservation Constraint and Pseudo-Paired Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4SLVUMCS}},
  note         = {Machine review of arXiv:2506.22022}
}
read the original abstract

Facial stylization aims to transform facial images into appealing, high-quality stylized portraits, with the critical challenge of accurately learning the target style while maintaining content consistency with the original image. Although previous StyleGAN-based methods have made significant advancements, the generated results still suffer from artifacts or insufficient fidelity to the source image. We argue that these issues stem from neglecting semantic shift of the generator during stylization. Therefore, we propose a facial stylization method that integrates semantic preservation constraint and pseudo-paired supervision to enhance the content correspondence and improve the stylization effect. Additionally, we develop a methodology for creating multi-level pseudo-paired datasets to implement supervisory constraint. Furthermore, building upon our facial stylization framework, we achieve more flexible multimodal and reference-guided stylization without complex network architecture designs or additional training. Experimental results demonstrate that our approach produces high-fidelity, aesthetically pleasing facial style transfer that surpasses previous methods.

Figures

Figures reproduced from arXiv: 2506.22022 by the authors.

Figure 1
Figure 1. We propose a facial stylization approach supporting general, multimodal, and reference-guided stylization. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Facial stylization framework: supporting reference-guided and multimodal facial stylization [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Changes in the latent space data distribution and semantics during finetuning. Top: Unconstrained; Bottom: [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (24 more)
Figure 4
Figure 4. Figure 4: Model finetuning with semantic preservation constraint and pseudo-paired supervision. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Results of random sampling images during finetuning with and without proposed constraints. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Pseudo-paired data generation in latent space. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Example of pseudo-paired data Through the aforementioned method, we obtain three levels of realistic-stylized portrait paired data as in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results of the comparative experiment, styles from top to bottom: Cartoon, Anime, Fantasy, [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Ablation study of semantic and pseudo-paired constraints. SC represents semantic constraint, and PC [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Multimodal stylization results. First row: UI2I-style method [ [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Reference-guided stylization results. Mixed encoding combinations: 3, 6, and 9. [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: Paired data obtained by sequentially operating on images in different latent spaces. [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 14
Figure 14. Figure 14: Stylization results of encoding the input image to different latent spaces. We use a modified pSp encoder to [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: Stylization results of encoding the reference image to different latent spaces. We adjust [6] to optimize the [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: Results from pseudo-paired data under supervision. Pairs of rows show examples in various styles. [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Portrait stylization results 16 [PITH_FULL_IMAGE:figures/full_fig_p016_17.png]
Figure 18
Figure 18. Figure 18: Portrait stylization results 17 [PITH_FULL_IMAGE:figures/full_fig_p017_18.png]
Figure 19
Figure 19. Figure 19: Multimodal portrait stylization results in cartoon style, with encoding combinations 3, 6, 9 and 12. [PITH_FULL_IMAGE:figures/full_fig_p018_19.png]
Figure 20
Figure 20. Figure 20: Multimodal portrait stylization results in anime style. [PITH_FULL_IMAGE:figures/full_fig_p019_20.png]
Figure 21
Figure 21. Figure 21: Multimodal portrait stylization results in fantasy style. [PITH_FULL_IMAGE:figures/full_fig_p020_21.png]
Figure 22
Figure 22. Figure 22: Multimodal portrait stylization results in illustration style. [PITH_FULL_IMAGE:figures/full_fig_p021_22.png]
Figure 23
Figure 23. Figure 23: Multimodal portrait stylization results in impasto style. [PITH_FULL_IMAGE:figures/full_fig_p022_23.png]
Figure 24
Figure 24. Figure 24: Reference-guided portrait stylization results, with encoding combinations 3, 6 and 9. [PITH_FULL_IMAGE:figures/full_fig_p023_24.png]
Figure 25
Figure 25. Figure 25: Reference-guided portrait stylization results. [PITH_FULL_IMAGE:figures/full_fig_p024_25.png]
Figure 26
Figure 26. Figure 26: Pseudo-paired dataset: fantasy and illustration style. [PITH_FULL_IMAGE:figures/full_fig_p025_26.png]
Figure 27
Figure 27. Figure 27: Pseudo-paired dataset: cartoon and impasto style. [PITH_FULL_IMAGE:figures/full_fig_p026_27.png]
Figure 28
Figure 28. Figure 28: Pseudo-paired dataset: anime style. References [1] Yang Zhao, Diya Ren, Yuan Chen, Wei Jia, Ronggang Wang, and Xiaoping Liu. Cartoon image processing: a survey. International Journal of Computer Vision, 130(11):2733–2769, 2022. [2] Jun-Yan Zhu, Taesung Park, Phillip I…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 35 canonical work pages

  1. [1]

    Cartoon image processing: a survey

    Yang Zhao, Diya Ren, Yuan Chen, Wei Jia, Ronggang Wang, and Xiaoping Liu. Cartoon image processing: a survey. International Journal of Computer Vision, 130(11):2733–2769, 2022

    work page 2022

  2. [2]

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

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223–2232, 2017

    work page 2017

  3. [3]

    U-gat-it: unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation

    J Kim. U-gat-it: unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830, 2019

    arXiv 1907

  4. [4]

    GANs N' Roses: Stable, Controllable, Diverse Image to Image Translation (works for videos too!)

    Min Jin Chong and David Forsyth. Gans n’roses: Stable, controllable, diverse image to image translation (works for videos too!). arXiv preprint arXiv:2106.06561, 2021

    work page Pith review arXiv 2021

  5. [5]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019

    2019

  6. [6]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020

    2020

  7. [7]

    Stylealign: Analysis and applications of aligned stylegan models

    Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021

    arXiv 2021

  8. [8]

    Resolution Dependent GAN Interpolation for Controllable Image Synthesis Between Domains

    Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020

    work page Pith review arXiv 2010

Show all 53 references
  1. [9]

    Unsupervised image-to-image translation via pre-trained stylegan2 network

    Jialu Huang, Jing Liao, and Sam Kwong. Unsupervised image-to-image translation via pre-trained stylegan2 network. IEEE Transactions on Multimedia, 24:1435–1448, 2021

    2021

  2. [10]

    Agilegan: stylizing portraits by inversion-consistent transfer learning

    Guoxian Song, Linjie Luo, Jing Liu, Wan-Chun Ma, Chunpong Lai, Chuanxia Zheng, and Tat-Jen Cham. Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG), 40(4):1–13, 2021. 27 A PREPRINT - SEPTEMBER 1, 2025

    2021

  3. [11]

    Pastiche master: Exemplar-based high-resolution portrait style transfer

    Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693–7702, 2022

    2022

  4. [12]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021

    2021 arXiv

  5. [13]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Jonas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  6. [14]

    Diffusion in style

    Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Süsstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251–2261, 2023

    2023

  7. [15]

    Stylediffusion: Controllable disentangled style transfer via diffusion models

    Zhizhong Wang, Lei Zhao, and Wei Xing. Stylediffusion: Controllable disentangled style transfer via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677–7689, 2023

    2023

  8. [16]

    Inversion-based style transfer with diffusion models

    Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146–10156, 2023

    2023

  9. [17]

    Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer

    Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795–8805, 2024

    2024

  10. [18]

    Generative adversarial nets

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

    2014

  11. [19]

    Unsupervised representation learning with deep convolutional generative adversarial networks

    Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015

    2015 arXiv

  12. [20]

    Image-to-image translation with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125–1134, 2017

    2017

  13. [21]

    Conditional generative adversarial nets

    Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014

    2014 arXiv

  14. [22]

    Dualgan: Unsupervised dual learning for image-to-image translation

    Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE international conference on computer vision, pages 2849–2857, 2017

    2017

  15. [23]

    Learning to discover cross-domain relations with generative adversarial networks

    Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. InInternational conference on machine learning, pages 1857–1865. PMLR, 2017

    2017

  16. [24]

    Image style transfer using convolutional neural networks

    Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414–2423, 2016

    2016

  17. [25]

    Texture networks: Feed-forward synthesis of textures and stylized images

    Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. arXiv preprint arXiv:1603.03417, 2016

    2016 arXiv

  18. [27]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501–1510, 2017

    2017

  19. [28]

    Arbitrary style transfer with style-attentional networks

    Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019

    2019

  20. [29]

    Dynamic instance normalization for arbitrary style transfer

    Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 4369–4376, 2020

    2020

  21. [30]

    Artflow: Unbiased image style transfer via reversible neural flows

    Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862–871, 2021

    2021

  22. [31]

    Styleformer: Real-time arbitrary style transfer via parametric style composition

    Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 14618–14627, 2021. 28 A PREPRINT - SEPTEMBER 1, 2025

    2021

  23. [32]

    Stytr2: Image style transfer with transformers

    Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326–11336, 2022

    2022

  24. [33]

    General image-to-image translation with one-shot image guidance

    Bin Cheng, Zuhao Liu, Yunbo Peng, and Yue Lin. General image-to-image translation with one-shot image guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22736–22746, 2023

    2023

  25. [34]

    Null-text guidance in diffusion models is secretly a cartoon-style creator

    Jing Zhao, Heliang Zheng, Chaoyue Wang, Long Lan, Wanrong Huang, and Wenjing Yang. Null-text guidance in diffusion models is secretly a cartoon-style creator. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5143–5152, 2023

    2023

  26. [35]

    Gan inversion: A survey

    Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(3):3121–3138, 2022

    2022

  27. [36]

    Interpreting the latent space of gans for semantic face editing

    Yujun Shen, Jinjin Gu, Xiaoou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243–9252, 2020

    2020

  28. [37]

    Ganspace: Discovering interpretable gan controls

    Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841–9850, 2020

    2020

  29. [38]

    Stylespace analysis: Disentangled controls for stylegan image generation

    Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12863–12872, 2021

    2021

  30. [39]

    Styleclip: Text-driven manipulation of stylegan imagery

    Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021

    2021

  31. [40]

    Closed-form factorization of latent semantics in gans

    Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1532–1540, 2021

    2021

  32. [41]

    Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF international conference on computer vision, pages 4432–4441, 2019

    Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF international conference on computer vision, pages 4432–4441, 2019

    2019

  33. [42]

    Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296–8305, 2020

    Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296–8305, 2020

    2020

  34. [43]

    Encoding in style: a stylegan encoder for image-to-image translation

    Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021

    2021

  35. [44]

    Designing an encoder for stylegan image manipulation

    Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1–14, 2021

    2021

  36. [45]

    Restyle: A residual-based stylegan encoder via iterative refinement

    Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6711–6720, 2021

    2021

  37. [46]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694–711. Springer, 2016

    2016

  38. [47]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2019

    2019

  39. [48]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  40. [49]

    Danbooru2019 portraits: A large-scale anime head illustration dataset

    Gwern Branwen, Anonymous, and Danbooru Community. Danbooru2019 portraits: A large-scale anime head illustration dataset. https://gwern.net/crop#danbooru2019-portraits, March 2019. Accessed: DATE

    2019

  41. [50]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  42. [51]

    A method for stochastic optimization

    D Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), volume 5, page 6. San Diego, California;, 2015. 29 A PREPRINT - SEPTEMBER 1, 2025

    2015

  43. [52]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  44. [53]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, volume 25, pages 1097–1105, 2012

    2012

  45. [54]

    Rethinking fid: Towards a better evaluation metric for image generation

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307–9...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.