Pith. sign in

REVIEW 3 major objections 6 minor 10 references

HyperGAN-CLIP: A Unified Framework for Domain Adaptation, Image Synthesis and Manipulation

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read One CLIP-conditioned hypernetwork can make a frozen StyleGAN2 generator adapt to new domains, transfer reference styles, and edit by text — with no text training.

desk verdict Solid unification paper with one unvalidated load-bearing assumption in the text-guided editing claim. read the letter →

arxiv 2411.12832 v1 pith:X7OQIC6E submitted 2024-11-19 cs.CV

classification cs.CV
keywords GANStyleGAN2CLIPhypernetworkdomainadaptationreference-guidedimagesynthesistext-guidedmanipulationone-shotlearning
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

HyperGAN-CLIP is an attempt to settle a practical question: can one architecture, trained once, make a pre-trained StyleGAN2 generator do three jobs that usually need separate models — one-shot adaptation to many new domains, reference-guided style transfer, and text-guided image editing? The paper's answer is yes, by conditioning a hypernetwork on Δ-CLIP embeddings, the difference between the CLIP embedding of the target (an image or a text prompt) and the source. Because the hypernetwork only predicts per-layer weight biases and scales that are injected as residual features into the frozen generator, the original model's identity and diversity are preserved while the new domain's characteristics appear. If the claim holds, practitioners can replace per-domain fine-tuned generators with one model and get text-guided editing without collecting any text-image pairs.

What carries the argument

The load-bearing object is the CLIP-conditioned hypernetwork module. For each layer i, two fully-connected heads read the Δ-CLIP embedding Δc = CLIP(conditioning input) − CLIP(source) and output a task-specific weight bias Δφ_i and channel-wise scale δ_i; these modulate a duplicated StyleGAN2 branch (θ*_i = δ_i · f(φ_i + Δφ_i, s_i)), whose features are scaled by η and added to the frozen branch's features. This residual feature injection is what lets the model add only the missing attributes while keeping the source distribution intact, and the Δ-CLIP centering is what lets image prompts and text prompts share one interface.

What would settle it

Measure the cosine similarity between Δ-CLIP text vectors and Δ-CLIP image vectors for a held-out set of attribute prompts (e.g., 'surprised', 'Elsa from Frozen') against image pairs illustrating those attributes; if the mean similarity is no better than chance for prompts the model has not seen, then the claimed zero-text-training editing pathway is not actually driven by semantic alignment.

Watch

Extended reading notes

Core claim

The central claim is that the weights of a pre-trained StyleGAN2 generator can be modulated dynamically by a hypernetwork conditioned on CLIP embeddings, so that a single trained model performs multi-domain one-shot adaptation, reference-guided image synthesis, and text-guided image manipulation with no retraining for each task. The modulation is not applied to the original weights; a duplicated generator branch produces domain-specific features using hypernetwork-predicted weight biases Δφ_i and scales δ_i, and these are injected into the frozen branch through F'_i = F_i + η F*_i. Training uses CLIP-based directional losses (across and within domains), a CLIP-conditioned projection discriminator, a contrastive domain-separating loss, and identity, L2, and LPIPS losses. The paper reports that this preserves source identity while capturing target characteristics, and demonstrates text-guided manipulation with no text training data by feeding the text's Δ-CLIP embedding into the same pathway learned from image pairs.

Load-bearing premise

The load-bearing assumption is that the Δ-CLIP vector computed from text (e.g., CLIP('blonde hair') − CLIP('face')) is semantically aligned with the Δ-CLIP vector of the corresponding image transformation learned from image pairs, so that a prompt can stand in for a reference image without any text training.

Editorial extensions

If this is right

  • One trained HyperGAN-CLIP model replaces per-domain fine-tuned generators: the paper demonstrates adaptation to 101 FFHQ-based domains and 52 AFHQ animal domains, where baselines such as StyleGAN-NADA and Mind-the-Gap need a separate model for each domain.
  • Text-guided manipulation is possible without any text-image training pairs: the Δ-CLIP vector of a prompt such as 'blonde hair' versus 'face' is fed into the same hypernetwork pathway learned from reference-image pairs, yielding edits on CelebA that are competitive with text-trained methods on attribute-accuracy and precision metrics.
  • The frozen base generator keeps its latent-space semantics, so existing latent-space editing directions, such as InterfaceGAN-style controls, continue to work inside adapted domains.
  • Because the residual features are scaled by η = 0.1 and the CLIP embedding can be scaled or interpolated at inference, users get a continuous control knob over how strongly the target style is applied.
  • The paper's ablation shows that adding the CLIP-conditioned discriminator and the residual-feature scheme together improves one-shot adaptation quality substantially over using only modulated target features.

Reading between the lines

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

  • If the Δ-CLIP modality transfer is as reliable as the paper assumes, the same hypernetwork should accept conditioning from other modalities that CLIP can encode or that can be mapped into CLIP space, such as sketches or audio descriptions, with no architectural change; this is my extension, not a paper claim.
  • The paper's stated limitation that only semantically similar, seen concepts work suggests a natural scaling test: training on a much wider and denser set of domains should broaden zero-shot transfer, and failure of that scaling would pinpoint where the cross-modal alignment breaks.
  • The residual-injection design implies that the hypernetwork contributes only a small parameter overhead on top of the frozen generator, so a single checkpoint could ship many styles and be updated by appending new domain embeddings rather than retraining weights; this is an implication the paper does not quantify.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes HyperGAN-CLIP, a unified framework built on a frozen StyleGAN2 generator in which a CLIP-conditioned hypernetwork predicts per-layer weight modulations and injects residual features into the frozen generator. The framework is trained with a multi-term loss and a CLIP-conditioned discriminator, and the same trained model family is applied to three tasks: multi-domain one-shot domain adaptation (FFHQ, AFHQ), reference-guided image synthesis (CelebA-HQ), and text-guided image manipulation trained without text-specific data. Quantitative results (FID, quality, diversity, identity similarity, CLIP similarity, AMA, CMP) and a user study are reported against several GAN- and diffusion-based baselines. Code and model links are provided.

Significance. If the claims hold, the framework is a practical step toward a single generator that can handle multiple adaptation and editing tasks without per-domain fine-tuning, and the text-guided editing without text-specific training data would be a notable capability. The paper presents quantitative results for all three tasks, comparisons to numerous baselines, a user study, and a public implementation, which are clear strengths. The architectural idea is a reasonable extension of prior hypernetwork-based adaptation work, though the incremental novelty relative to HyperDomainNet and DynaGAN should be more sharply delineated. The main risk is that the headline text-free editing claim rests on an unverified cross-modal alignment assumption, and the FID evaluation protocol has methodological weaknesses that temper the reported numbers.

major comments (3)
  1. [Sec. 3.2.1, Eq. (5), Sec. 4.4] The headline claim of text-guided image manipulation without text-specific training data relies on the assumption that the text-derived Δ-CLIP vector Δc_text = CLIP(t_target) − CLIP(t_source) occupies the same region of Δc space and induces the same weight modulation as the image-derived Δ-CLIP directions used during reference-guided training. This assumption is never directly measured. The quantitative evaluation in Section 4.4 uses only a fixed list of CelebA attributes for AMA/CMP; novel prompts such as 'Elsa from Frozen' are shown only qualitatively. The supplement's own Limitations section (Supplementary Section 9) states that the model 'can produce visually plausible results only for concepts encountered during training or those that are semantically similar,' which is exactly the expected failure mode if the cross-modal alignment is only partial. To support the central text-free claim, the authors should quantify the alignment (e.g., distribution overlap between text and image Δ-CLIP vectors for matched attributes, or a nearest-neighbor analysis) and report AMA/CMP on a held-out set of concepts with a range of semantic distances from the training distribution. Without such evidence, the abstract's claim that the method enables text-guided manipulation 'without the need for text-specific training data' is overstated.
  2. [Sec. 4.1, Eq. (6)] The total loss is a weighted sum of eight terms with hand-set coefficients (λ1 = 30 through λ8 = 12) and an additional scaling parameter η = 0.1. The paper provides no sensitivity analysis, no validation-based selection criterion, and no ablation over these coefficients; the supplement's ablation only removes entire loss terms rather than varying their weights. Since the quantitative comparisons in Tables 1–3 are all obtained under this single configuration, the robustness of the reported results to hyperparameter choice is unknown. A sensitivity study (e.g., perturbing each λ while holding the others fixed and reporting FID/AMA) or at least a statement of how the values were chosen should be added.
  3. [Supplementary Sec. 2 (Evaluation Details)] The FID protocol for the domain-adaptation experiments uses reference images generated by the NADA-expanded Domain Expansion model and represents the generated distribution with only 100 randomly sampled images per target domain. FID computed with a 100-image generated set and a 1K-image reference set will have high variance, and the paper does not report the number of latent samples, the seed, or confidence intervals. Moreover, because the same model that produced the one-shot training images also generates the reference distribution, the outcome may partly reflect how well the evaluated methods mimic that specific generator rather than adaptation to a real target distribution. The authors should report FID across multiple sampling seeds with mean and standard deviation and should discuss the implications of using a synthetic reference distribution.
minor comments (6)
  1. [Sec. 3.2.1] The notation for the anchor image xfixed is used both for the inverted target image in domain adaptation and for the average StyleGAN image in reference-guided synthesis; the main text should clarify which variant is used in Eqs. (8)–(9) for each application.
  2. [Sec. 4.2] The phrase '101 new domains introduced in the expanded version of StyleGAN-NADA' is imprecise; the authors should cite and describe the Domain Expansion model by Nitzan et al. more carefully, since StyleGAN-NADA originally introduced a smaller set of domains.
  3. [Table 3] The reported FID for HyperGAN-CLIP (87.851) is substantially higher than for DiffusionCLIP (29.280) and Plug-and-play (68.287); the statement that the model 'finds a good balance between the metrics' should be backed by a statistical comparison (e.g., confidence intervals) or by acknowledging the FID gap more explicitly.
  4. [Supplementary Table 1] The user study table does not define the ranking scale; in the domain-adaptation rows, Ours (1.65) and DynaGAN (1.58) are closer to 1 than HyperDomainNet (2.77), which suggests that 1 is best, but this should be stated explicitly in the text.
  5. [Supplementary Sec. 6 and Sec. 8] The supplement contains inconsistent naming: the framework is referred to as 'HyperCLIP-GAN' in the caption of Supplementary Figure 6 and in Section 8, while the main paper and the rest of the supplement use 'HyperGAN-CLIP'.
  6. [General] There are several typos and incomplete references: 'accross' (Supplementary Section 2), 'Multple' (Supplementary Section 8 heading), 'Ta rg et' (Supplementary Figure 5), and the Domain Expansion reference in Section 4.2 lists only a URL rather than a full citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: HyperGAN-CLIP trains a conditional hypernetwork on image CLIP deltas and evaluates against external baselines; the text-free editing result is an empirical generalization, not a consequence of fitting the test input.

full rationale

The paper's derivation chain is an empirical training pipeline, not a closed-form argument that reduces to its own inputs. In Sec. 3.1, the hypernetwork H_i maps a Δ-CLIP embedding to per-layer weight modulations (Eq. 5), and the training losses in Sec. 3.2.1 (L_CLIP, L_CLIP-Across, L_CLIP-Within) supervise that map using image-derived directional CLIP differences. The text-guided editing claim in Sec. 3.2.1 feeds Δc_text = CLIP(t_target) − CLIP(t_source) into the same trained H_i without any text supervision. This is a genuine zero-shot generalization claim: the text delta is not used during training, and the output is not fed back into the training objective, so the result is not statistically forced by construction. The paper's own supplement (Sec. 9) admits the method only works for concepts encountered during training or semantically similar ones, which confirms that the cross-modal alignment is an empirical, partially verified assumption rather than a tautology. The Δ-CLIP conditioning is adopted from DeltaEdit (Lyu et al. 2023), an external prior work, not from the authors' own prior results. The only overlapping-author citation, CLIPInverter (Baykal et al. 2023), is used for the AMA/CMP evaluation methodology and as a comparison baseline; it is not load-bearing for the central claims. The FID protocol for domain adaptation samples 'real' images from the Domain Expansion model that also produced the single training image; this is a standard test-distribution choice for one-shot adaptation, not a case of evaluating on the model's own training output. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no result is equivalent by definition to its inputs. The unverified text-image Δ-CLIP alignment is a correctness/robustness risk, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The ledger shows that the paper's contributions are architectural and procedural: two new modules (the CLIP hypernetwork and residual injection) plus a loss schedule. All central claims also depend on CLIP, StyleGAN2, e4e inversion, and the NADA-expanded model, which are pulled from prior work rather than derived. The hand-set constants and the evaluation sampling protocol are the most fragile choices, since no sensitivity analysis is reported.

free parameters (4)
  • Scaling parameter eta = 0.1
    Set by hand in Sec 4.1 to control the magnitude of the residual features injected into the frozen generator. The paper says it is set to prevent a large distribution shift; no search or sensitivity analysis is given.
  • Loss weights lambda_1 through lambda_8 = 30, 1.5, 0.5, 0.2, 1.0, 3.0, 8.0, 12.0
    Chosen empirically in Sec 4.1. The paper gives no justification, grid search, or sensitivity analysis for these eight numbers. They affect the balance among CLIP alignment, identity preservation, reconstruction, and adversarial terms.
  • Alpha mixing weight for combined reference and text embeddings = 0.5
    Used in Sec 4.3 to combine a reference image CLIP embedding with a text attribute embedding, CLIP(x_target) + alpha CLIP(t_target), to refine editing. The value is described as a demonstration; no search is reported.
  • Number of sampled images per domain for FID evaluation = 100
    In the supplementary evaluation details, 100 generated images per domain are randomly sampled to represent the generated distribution. This is a methodological choice that affects the FID estimates, particularly with high variance due to one-shot settings.
assumptions (5)
  • domain assumption A pre-trained StyleGAN2 generator on FFHQ or AFHQ cats is available and can be treated as a frozen source encoder.
    All experiments depend on the existence and quality of StyleGAN2 pre-trained on the source domains, described in Sec 4.2 and used without modification throughout.
  • domain assumption The e4e encoder provides an adequate inversion of target images into the source latent space, and latent truncation keeps inversions in-distribution.
    Section 4.1 states e4e inversion with truncation is used to obtain x_fixed. If the target image cannot be inverted well, the directional CLIP losses and identity preservation are compromised.
  • domain assumption CLIP embeddings of images and text share a sufficiently aligned semantic space such that delta_CLIP(text) approximates delta_CLIP(image) for editing directions.
    The text-guided manipulation experiment in Sec 4.4 relies on this cross-modal equivalence without direct measurement. The supplementary limitations section partially acknowledges this fails for unseen concepts.
  • domain assumption The NADA-expanded Domain Expansion model generates realistic target domain images, and the single image sampled per domain is representative of that domain.
    Used in Sec 4.2 to generate training targets and FID reference sets. No independent verification of the representativeness of a single sampled image is provided.
  • standard math Standard PyTorch autograd, Adam convergence, and the StyleGAN2 codebase perform as expected.
    No formal verification. The paper relies on standard deep learning infrastructure without mentioning seeds, gradient checks, or hyperparameter search protocols.
invented entities (3)
  • CLIP-conditioned hypernetwork module H_i
    purpose: Predicts per-layer weight biases and channel-wise scale parameters from the Δ-CLIP embedding, producing modulated weights for a duplicated StyleGAN2 generator.
    A new architectural component introduced in this paper. Its effectiveness is only evidenced by the paper's own experiments; no external benchmark or formal check is provided.
  • Residual feature injection scheme, Eq. 1
    purpose: Adds scaled modulated features to the frozen generator's features to preserve source identity while incorporating target style.
    This is a new design choice introduced by the paper. No theoretical analysis or standalone experimental evidence outside the framework's ablations is given.
  • CLIP-conditioned discriminator
    purpose: A projection discriminator with a frozen CLIP backbone, trained head-only, to distinguish source and target domain distributions and improve image quality.
    An architectural modification of existing projection discriminators. Its benefit is supported only by the included ablation study (FID improvement from 43.43 to 33.76).

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperGAN-CLIP: A Unified Framework for Domain Adaptation, Image Synthesis and Manipulation." pith.science (2026). https://pith.science/paper/X7OQIC6E

@misc{pith2026241112832,
  author       = {Pith},
  title        = {Pith review of: HyperGAN-CLIP: A Unified Framework for Domain Adaptation, Image Synthesis and Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X7OQIC6E}},
  note         = {Machine review of arXiv:2411.12832}
}
read the original abstract

Generative Adversarial Networks (GANs), particularly StyleGAN and its variants, have demonstrated remarkable capabilities in generating highly realistic images. Despite their success, adapting these models to diverse tasks such as domain adaptation, reference-guided synthesis, and text-guided manipulation with limited training data remains challenging. Towards this end, in this study, we present a novel framework that significantly extends the capabilities of a pre-trained StyleGAN by integrating CLIP space via hypernetworks. This integration allows dynamic adaptation of StyleGAN to new domains defined by reference images or textual descriptions. Additionally, we introduce a CLIP-guided discriminator that enhances the alignment between generated images and target domains, ensuring superior image quality. Our approach demonstrates unprecedented flexibility, enabling text-guided image manipulation without the need for text-specific training data and facilitating seamless style transfer. Comprehensive qualitative and quantitative evaluations confirm the robustness and superior performance of our framework compared to existing methods.

Figures

Figures reproduced from arXiv: 2411.12832 by the authors.

Figure 1
Figure 1. HyperGAN-CLIP and its Applications. Introducing HyperGAN-CLIP, a flexible framework that enhances the capabilities of a pre-trained StyleGAN model for a multitude of tasks, including multiple domain one-shot adaptation, reference-guided image synthesis and text-guided image manipulation. Our method pushes the boundaries of image synthesis and editing, enabling users to create diverse and high-quality images with rem… view at source ↗
Figure 2
Figure 2. Overview of HyperGAN-CLIP. This framework employs hypernetwork modules to adjust StyleGAN generator weights based on images or text prompts. These inputs facilitate domain adaptation, attribute transfer, or image editing. The modulated weights blend with original features to produce images that align with specified domains or tasks like reference-guided synthesis and text-guided manipulation, while maintaining sourc… view at source ↗
Figure 3
Figure 3. Comparison against the state-of-the-art few-shot domain adaptation methods. Our proposed HyperGAN-CLIP model outperforms competing methods in accurately capturing the visual characteristics of the target domains. Domain 1 Domain 2 Hybrid (a) Domain mixing. Our approach can fuse multiple domains to create novel compositions. By averaging and re-scaling the CLIP embeddings of two target domains, we can generate images… view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: Capabilities of HyperGAN-CLIP in blending domains and performing semantic edits within adapted domains. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison with state-of-the-art reference-guided image synthesis approaches. Our approach effectively transfers the style of the target image to the source image while effectively preserving identity compared to competing methods [PITH_FULL_IMAGE:figures/full_fig_p00…
Figure 6
Figure 6. Figure 6: Reference-guided image synthesis with mixed embeddings. Each row shows the input image, the initial result with the CLIP image embedding, the refined result with a mixed embedding that incorporates the target attribute with 𝛼 = 0.5, and the reference image, respectivel…
Figure 7
Figure 7. Figure 7: Comparisons with state-of-the-art text-guided image manipulation methods. Our model shows remarkable versality in manipulating images across a diverse range of textual descriptions. The results vividly illustrate our model’s ability to accurately apply changes based on…
Figure 1
Figure 1. Figure 1: Visualization of the directional CLIP losses. (a) for domain adaption. (b) for reference-guided image synthesis. with the source image to be high and ID similarity with the target image to be low as we want to preserve the identity of the source image while only transf…
Figure 2
Figure 2. Figure 2: A sample question from the user study. The participants rank the options from best to worst. The human evaluation comprises three sections, each dedicated to one of our tasks, with 25 questions per section. Within each part, users are shown a random source image alongs…
Figure 4
Figure 4. Figure 4: Scaling residual features. The top row shows the results obtained with our approach, whereas the bottom on corresponds to the results by the baseline model with the discriminator. Several artifacts instantly start to appear in the baseline results when scaling beyond t…
Figure 5
Figure 5. Figure 5: Zero-Shot Domain Adaptation. Our model can perform domain adaption quite reasonably well on target domains not seen during training. Here we provide results on target dog breed images from the AFHQ dog dataset not used in the training [PITH_FULL_IMAGE:figures/full_fig…
Figure 3
Figure 3. Figure 3: Controllable Manipulation. In our approach, we can vary the amount of residual features injected as well as the amount of target style latent, which gives users the ability to control degree of adaptation with respect to style consistency vs data fidelity. 0 β 3.5 [PI…
Figure 6
Figure 6. Figure 6: Qualitative results for the ablation study. Baseline network does not preserve the facial identity of the source image, giving an outcome closely resembling to the target image. When CLIP-conditioned discriminator is incorporated to the baseline, the image quality is i…
Figure 7
Figure 7. Figure 7: Impact of Δ-CLIP Embeddings. Our model equipped with Δ-CLIP embeddings performs semantic edits that are better aligned with the provided textual descriptions as compared to the version of our model that employs original CLIP embeddings. Overall, the results show that o…
Figure 8
Figure 8. Figure 8: Impact of Δ-CLIP Embeddings. As observed, we obtain much accurate manipulations while preserving the quality and fidelity when Δ-CLIP embeddings are used. Mingcong Liu, Qiang Li, Zekui Qin, Guoxin Zhang, Pengfei Wan, and Wen Zheng. 2021. BlendGAN: Implicitly GAN Blendi…
Figure 9
Figure 9. Figure 9: Additional qualitative comparison against the state-of-the-art few-shot domain adaptation methods on AFHQ dataset. Our proposed HyperGAN-CLIP model outperforms competing methods in accurately capturing the visual characteristics of the target domains. The synthesized i…
Figure 10
Figure 10. Figure 10: Additional qualitative comparison against the state-of-the-art few-shot domain adaptation methods on AFHQ dataset. Our proposed HyperGAN-CLIP model outperforms competing methods in accurately capturing the visual characteristics of the target domains. The synthesized …
Figure 11
Figure 11. Figure 11: Additional qualitative comparison with state-of-the-art reference-guided image synthesis approaches. Our approach effectively transfers the style of the target image to the source image while effectively preserving identity compared to competing methods. ACM Trans. Gr…
Figure 12
Figure 12. Figure 12: Additional qualitative comparisons with state-of-the-art text-guided image manipulation methods. Our model shows remarkable versality in manipulating images across a diverse range of textual descriptions. The results vividly illustrate our model’s ability to accuratel…
Figure 13
Figure 13. Figure 13: Text-guided editing results for the birds dataset. Our approach generalizes to other domains, such as the bird images. We demonstrate zero-shot text-guided image editing results. Source Target Output Source Target Output [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]
Figure 14
Figure 14. Figure 14: Reference-guided editing results for the birds dataset. Our reference-guided synthesis generalizes to the birds domain, illustrated by the various targets we provide. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: November 2024 [PITH_FULL_IMAGE:figures…
Figure 15
Figure 15. Figure 15: Additional qualitative comparison with HyperDomainNet. The comparisons on a smaller set of domains shows that our proposed HyperGAN￾CLIP model performs comparably or better than HyperDomainNet. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: November 202…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 6 canonical work pages

  1. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44, 10 (oct 2022), 5962–5979

    ArcFace: Additive Angular Margin Loss for Deep Face Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence44, 10 (oct 2022), 5962–5979. https://doi.org/10.1109/tpami.2021.3087709 Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or

  2. [5]

    ACM Trans

    StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators. ACM Trans. Graph. 41, 4, Article 141 (jul 2022), 13 pages. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun

  3. [8]

    arXiv preprint arXiv:2211.12572 (2022)

    Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation. arXiv preprint arXiv:2211.12572 (2022). C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. 2011.The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001. California Institute of Technology. Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, ...

  4. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    Hairclip: Design your hair by text and reference image. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022). Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu

  5. [15]

    The comparisons on a smaller set of domains shows that our proposed HyperGAN- CLIP model performs comparably or better than HyperDomainNet

    Additional qualitative comparison with HyperDomainNet. The comparisons on a smaller set of domains shows that our proposed HyperGAN- CLIP model performs comparably or better than HyperDomainNet. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: November 2024

  6. [2015]

    arXiv:1512.03385 [cs.CV] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter

    Deep Residual Learning for Image Recognition. arXiv:1512.03385 [cs.CV] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter

  7. [2020]

    arXiv:2003.04297 [cs.CV] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou

    Improved Baselines with Momentum Contrastive Learning. arXiv:2003.04297 [cs.CV] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou

  8. [2021]

    InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 2085–2094. Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel

Show all 10 references
  1. [2022]

    In Computer Vision – ECCV 2022: 17th European Conference, Tel A viv, Israel, October 23–27, 2022, Proceedings, Part XIII (Tel Aviv, Israel)

    Image-Based CLIP-Guided Essence Transfer. In Computer Vision – ECCV 2022: 17th European Conference, Tel A viv, Israel, October 23–27, 2022, Proceedings, Part XIII (Tel Aviv, Israel). 695–711. Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He

  2. [2023]

    ACM Trans

    CLIP-Guided StyleGAN Inversion for Text-Driven Real Image Editing. ACM Trans. Graph. 42, 5, Article 172 (aug 2023), 18 pages. Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.