Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Multi-Attribute guided Thermal Face Image Translation based on Latent Diffusion Model

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A latent diffusion model conditioned on facial attributes — gender, age, and skin tone — translates thermal face images into visible ones while preserving identity, reporting state-of-the-art scores on two benchmark datasets.

desk verdict Useful engineering combination with a plausible SOTA claim, but the identity evaluation is under-specified enough that the headline numbers should not be taken at face value. read the letter →

arxiv 2512.21032 v1 pith:2LVG7TUH submitted 2025-12-24 cs.CV

Mingshu Cai , Osamu Yoshie , Yuya Ieiri This is my paper · ORCID
classification cs.CV
keywords thermal-to-visiblefacetranslationlatentdiffusionmodelheterogeneousrecognitionmulti-attributeclassifierSelf-AttnMambaidentitypreservationskin-tonelabellingcross-modalsynthesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that thermal-to-visible (T2V) face translation is markedly improved by conditioning a latent diffusion model on explicit facial attributes extracted from the thermal image itself, rather than relying on the raw image alone. A multi-attribute classifier learns to read gender, age, and skin color from infrared input by mimicking a visible-light classifier on paired data, then feeds CLIP-encoded prompts into the denoising UNet. A Self-Attn Mamba module replaces standard attention for faster global modeling. On ARL-VTF and SpeakingFaces, the method reports the best image quality and identity verification scores, including Rank-1 accuracy of 81.83% versus 73.61% for the closest diffusion baseline. If correct, this offers a practical route to nighttime face recognition using thermal sensors feeding an existing visible-trained recognition system.

What carries the argument

The core mechanism is a two-part conditioning pipeline. First, a multi-attribute classifier: a frozen visible-light classifier (EfficientNet-based recognition with attribute heads) and a trainable thermal twin are trained together with an MSE consistency loss on paired thermal/visible images, so the thermal network learns to predict gender, age, and skin tone from infrared input. Second, those predicted labels are encoded into prompt embeddings by a fine-tuned CLIP encoder and injected into the latent diffusion UNet through cross-attention. The third component is Self-Attn Mamba, a bidirectional selective state-space module replacing multi-head self-attention in the UNet, which the paper arg

What would settle it

Train the method without the ID loss (or with a held-out recognition network different from F) and re-run Rank-1 and VR@FAR; if the large identity gains (e.g., Rank-1 81.83 vs 73.61) largely disappear, the reported identity preservation is an artifact of optimizing the same cosine-similarity objective that later scores the generated images.

Watch

Extended reading notes

Core claim

The central claim is that a latent diffusion model guided by attribute prompts — generated by a classifier that transfers visible-light attribute knowledge to thermal images — combined with a Mamba-based attention replacement, achieves state-of-the-art thermal-to-visible face synthesis. The attribute prompts close the modal gap in skin color, age, and gender; the identity loss pushes preservation of identity-specific structure; and the Self-Attn Mamba reduces computational cost while maintaining global feature modeling. The paper substantiates this with tables showing higher SSIM and PSNR, lower LPIPS and FID, and substantially higher Rank-1 and VR@FAR than five recent GAN and diffusion base

Load-bearing premise

The claim that identity preservation improves depends on the identity metrics being computed with a recognizer that is not the same network used to train the ID-loss objective, and the paper does not specify the matcher or gallery construction.

Editorial extensions

If this is right

  • Thermal surveillance images could be passed through this pipeline to produce visible-like faces that match existing visible-trained recognition systems, enabling night-time identification without retraining the recognizer.
  • The attribute-conditioning approach is a template for other cross-modal translation tasks where the domain gap is dominated by categorical attributes, such as NIR-to-visible or sketch-to-photo synthesis.
  • The Self-Attn Mamba swap suggests a general efficiency upgrade for conditional diffusion backbones, with reported single-image inference at 36 ms on a V100 GPU.
  • Identity verification scores in the 80–86% Rank-1 range on 100-identity test sets indicate the method is closing the gap to visible-light face recognition performance in the thermal-to-visible setting.

Reading between the lines

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

  • The reported identity gains may be partly inflated if the recognition network used to compute Rank-1 and VR@FAR is the same network F used in the ID-loss objective (Eq. 4); the paper does not name the matcher or gallery protocol, so an independent evaluation with a held-out recognizer would clarify the true identity-preservation benefit.
  • The classifier training relies on paired thermal-visible images with consistent attributes, which may limit deployment to settings where such pairs exist; an extension could test whether the attribute classifier can be trained or fine-tuned with unpaired data or synthetic thermal images.
  • The 19-level PANTONE skin-tone labels anchored to visible images provide a strong inductive bias, but errors in skin-tone prediction from thermal input could propagate into the generated face; stress-testing on more ethnically diverse populations is a natural next step.
  • The 36 ms inference claim depends on the specific GPU and implementation; a useful benchmark would be measuring runtime and memory on edge hardware typical of surveillance deployments.
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 a latent diffusion model for thermal-to-visible (T2V) face translation. The method combines a VQ-VAE latent space, a multi-attribute classifier that extracts age/gender/skin-tone from thermal images, CLIP-encoded prompts, an identity-preservation loss, and a 'Self-Attn Mamba' module intended to replace multi-head self-attention in the diffusion UNet. Experiments on ARL-VTF and SpeakingFaces report state-of-the-art image quality (SSIM, LPIPS, FID, PSNR) and identity-preservation metrics (Rank-1, VR@FAR), with ablations attributing gains to each component and a speed comparison against T2V-DDPM.

Significance. If the results hold, the work addresses a practically important problem—nighttime thermal-to-visible face recognition—and its attribute-conditioned latent diffusion design is a plausible route to improving both image fidelity and identity retention. Strengths include the ablation study, which shows monotonic improvements as components are added, and the fact that the image-quality metrics (SSIM, PSNR, LPIPS, FID) are not directly optimized by the noise-prediction loss, giving the quality claim some independence from the training objective. However, the central identity-preservation claim is currently supported only by point estimates from an evaluation protocol that omits the identity matcher, no code or data are released, the inference-speed comparison is confounded, and recent identity-preserving baselines are missing. These gaps must be addressed before the SOTA claim can be accepted.

major comments (4)
  1. [§4.3 / §3.3] The identity metrics are at risk of circularity. Eq. (4) trains the generator against a pre-trained face recognition network F, while §4.3 defines Rank-1 and VR@FAR without specifying which matcher is used. If the evaluation matcher is the same F, Tables 2, 4, and 6 quantify overfitting to F's embedding geometry rather than general identity preservation, especially because the compared baselines are not trained with an ID loss. The authors must name the face recognizer used in Eq. (4), the matcher used for Rank-1/VR@FAR, and the gallery/probe protocol. They should also report identity metrics using at least one independent face recognizer not used in training. This is load-bearing because identity preservation is the paper's headline contribution.
  2. [Tables 1-4] The SOTA claim rests on single point estimates without error bars, multiple seeds, or statistical significance tests. For example, Table 3 shows FID of 21.16 for OURS versus 35.58 for BBDM—a large margin, but with no indication of variance across runs or sensitivity to hyperparameters. No code or data are released, so the reported numbers cannot be independently verified. The authors should report mean/std over at least three runs and, ideally, make code and model weights available.
  3. [Table 7 / §4.6] The inference-speed comparison is confounded. T2V-DDPM is a pixel-space DDPM with 1000 steps and 273.68M parameters, while OURS is a latent diffusion model with a VQ-VAE autoencoder and only 75.96M parameters. The table therefore compares two different architectures, not the effect of replacing self-attention with 'Self-Attn Mamba.' To support the speed claim, the authors should compare the same latent diffusion base with MHSA versus their Mamba module under identical timesteps, batch size, and hardware, and report per-step latency and FLOPs.
  4. [§2.4 / §4.4] The paper cites DiffTV and Pair-ID as recent identity-preserving T2V methods in §2.4 but does not compare against them in Tables 1-4 or the qualitative figures. Given the abstract and §4.4 claim superiority over 'recent GAN and diffusion-based methods,' omitting the two most relevant recent identity-preserving baselines weakens the SOTA claim. The authors should either include these methods in the comparison or explicitly justify their exclusion.
minor comments (6)
  1. [§4.4] The text mentions 'images generated by SR3,' but SR3 is not among the compared methods (CFSM, GP-UNIT, BBDM, DiffuseIT, T2V-DDPM). This appears to be a typo that should be corrected.
  2. [§3.2] The citation for Latent Diffusion Models is incorrect: reference [15] is the HFB face database paper, not Rombach et al. The LDM citation should be fixed.
  3. [§2.3] The related work paragraph cites T2V-DDPM as [20], but [20] is AT-DDPM; T2V-DDPM is reference [21]. Also, 'Guided diffusion models [2]' cites a physiology-based face recognition paper, which does not support the statement. Several reference numbers are mismatched and should be audited.
  4. [Eq. (5) / Figure 3] Equation (5) minimizes squared differences between 'classification results,' while Figure 3's caption says consistency is enforced between feature vectors f_vis and f_IR using MSE. Please clarify whether the consistency loss is applied to classifier logits, post-softmax probabilities, or intermediate features.
  5. [Table 4] The table title says 'SpeakingFace' whereas the dataset is 'SpeakingFaces.' Use consistent nomenclature throughout.
  6. [Figure 2 caption] The caption states 'pre-trained classifier (parameters frozen, as shown in Figure 3),' but Figure 3 includes a trainable infrared classifier copy. The caption should be reworded to avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the main claims rest on empirical comparisons, and the potential identity-metric confound is not established by the paper's own text.

full rationale

The paper's derivation chain is a conventional trained generative pipeline: a latent diffusion loss (Eq. 1), an identity-consistency loss (Eq. 4), and a cross-modal classifier distillation loss (Eq. 5). None of the reported headline quantities (SSIM, LPIPS, FID, PSNR, Rank-1, VR@FAR) is defined as the training loss or as a fitted parameter of the model; they are measured on held-out test sets against independent baselines. The ID loss in Eq. 4 optimizes cosine similarity to a pre-trained face recognizer F, while the Rank-1/VR metrics in §4.3 describe a recognition protocol but do not name the matcher. If the matcher were exactly F, the Rank-1 gains would be partly forced by construction; however, the paper does not state that identity, so flagging it as circular would require speculation beyond the text. There are no self-citations used as load-bearing support, no uniqueness theorem imported from the authors' prior work, and no known result renamed as an organizing principle. The omission of the evaluation matcher is a reproducibility/correctness concern, not a demonstrated circularity. Therefore, under the hard rule requiring a quoted reduction, the appropriate finding is no significant circularity.

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

No new physical entities, forces, or conserved quantities are introduced. The 'Self-Attn Mamba module' and '19 PANTONE skin-tone labels' are model/taxonomy inventions, but they are software artifacts rather than entities with independent falsifiable handles. The key unvalidated inventions are the skin-tone label set and the attribute-conditioning path.

free parameters (4)
  • 19 PANTONE skin-tone labels = not reported
    The paper references the PANTONE SkinTone Guide to define 19 skin tone classes (§3.3), but does not specify how an image is mapped to a label; the classifier is trained to match a visible-light classifier's softmax outputs (§3.4), so the label set and the mapping are hand-chosen and affect all attribute-guided conditioning.
  • Prompt template wording = 'A face of {age} years old {gender} with {skin} skin.' (Fig. 2)
    The exact prompt text used to create CLIP embeddings is a hand-designed choice; small wording changes can shift CLIP embeddings and thus the conditioning, but no prompt-search or robustness analysis is provided.
  • ID-loss weight = not reported
    The total training objective must combine the LDM noise-prediction loss (§3.2, Eq. 1) with L_ID (§3.3, Eq. 4), but no weighting coefficient is given, making the balance between image quality and identity preservation a silent free parameter.
  • VQ-VAE hyperparameters = not reported
    The paper trains VQ-VAEs for visible and thermal images from scratch (§3.3), but does not report codebook size, latent downsampling factor f, or reconstruction quality, all of which bound the fidelity of the final generated images.
assumptions (4)
  • domain assumption The paired thermal-visible images in ARL-VTF and SpeakingFaces are geometrically aligned and show the same identity, expression, and lighting semantics.
    The classifier distillation (Eq. 5) and the generation pipeline (Eq. 3) assume pixel-level and semantic correspondence between I_th and I_vis. The paper relies on dataset synchronization but does not verify alignment in the chosen subsets.
  • domain assumption A classifier trained on visible images can serve as a reliable teacher for thermal images via feature-consistency distillation.
    The multi-attribute classifier is trained by minimizing the MSE between visible and thermal feature vectors (Eq. 5). This assumes cross-modal feature transferability, but no classification accuracy on thermal images is reported.
  • domain assumption The pre-trained face-recognition network F produces embeddings that are stable proxies for identity across the visible domain.
    The ID loss (Eq. 4) assumes F's cosine similarity between the generated visible image and a reference visible image is a valid training signal for identity preservation. No analysis of F's racial, age, or imaging-condition bias is included.
  • ad hoc to paper The Mamba state-space equations (Eqs. 6-8) are a drop-in replacement for multi-head self-attention in latent diffusion UNets with equivalent or better expressiveness.
    The paper asserts 'enhanced global feature modeling' and 'higher efficiency' for Self-Attn Mamba, but provides no theoretical analysis, only a confounded speed comparison (Table 7) and one ablation row (Variant A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Attribute guided Thermal Face Image Translation based on Latent Diffusion Model." pith.science (2026). https://pith.science/paper/2LVG7TUH

@misc{pith2026251221032,
  author       = {Pith},
  title        = {Pith review of: Multi-Attribute guided Thermal Face Image Translation based on Latent Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LVG7TUH}},
  note         = {Machine review of arXiv:2512.21032}
}
read the original abstract

Modern surveillance systems increasingly rely on multi-wavelength sensors and deep neural networks to recognize faces in infrared images captured at night. However, most facial recognition models are trained on visible light datasets, leading to substantial performance degradation on infrared inputs due to significant domain shifts. Early feature-based methods for infrared face recognition proved ineffective, prompting researchers to adopt generative approaches that convert infrared images into visible light images for improved recognition. This paradigm, known as Heterogeneous Face Recognition (HFR), faces challenges such as model and modality discrepancies, leading to distortion and feature loss in generated images. To address these limitations, this paper introduces a novel latent diffusion-based model designed to generate high-quality visible face images from thermal inputs while preserving critical identity features. A multi-attribute classifier is incorporated to extract key facial attributes from visible images, mitigating feature loss during infrared-to-visible image restoration. Additionally, we propose the Self-attn Mamba module, which enhances global modeling of cross-modal features and significantly improves inference speed. Experimental results on two benchmark datasets demonstrate the superiority of our approach, achieving state-of-the-art performance in both image quality and identity preservation.

Figures

Figures reproduced from arXiv: 2512.21032 by the authors.

Figure 1
Figure 1. Limitations of mainstream methods: GAN-based mod [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pipeline of our method. Our model begins with training a VQ-VAE model from scratch, enabling the denoising process to operate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Classifier module. During classifier training, an image pair consisting of an infrared image and an RGB face image [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative results on the ARL-VTF dataset for translating facial images from thermal to visible. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on the SpeakingFaces dataset for translating facial images from thermal to visible. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of improvement obtained by different vari [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 4 linked inside Pith

  1. [1]

    Abdrakhmanova, A

    M. Abdrakhmanova, A. Kuzdeuov, S. Jarju, Y . Khassanov, M. Lewis, and H. A. Varol. Speakingfaces: A large-scale multimodal dataset of voice commands with visual and ther- mal video streams.Sensors, 21(10):3465, 2021

    2021

  2. [2]

    Buddharaju, I

    P. Buddharaju, I. T. Pavlidis, P. Tsiamyrtzis, and M. Baza- kos. Physiology-based face recognition in the thermal in- frared spectrum.IEEE transactions on pattern analysis and machine intelligence, 29(4):613–626, 2007

    2007

  3. [3]

    Chen and A

    C. Chen and A. Ross. Matching thermal to visible face im- ages using hidden factor analysis in a cascaded subspace learning framework.Pattern Recognition Letters, 72:25–32,

  4. [4]

    W. Chen, J. Zhang, J. Wu, H. Wu, X. Xiao, and L. Lin. Id- aligner: Enhancing identity-preserving text-to-image gener- ation with reward feedback learning, 2024

    2024

  5. [5]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  6. [6]

    X. Di, B. S. Riggan, S. Hu, N. J. Short, and V . M. Pa- tel. Polarimetric thermal to visible face verification via self- attention guided synthesis. In2019 International Conference on Biometrics (ICB), pages 1–8. IEEE, 2019

    2019

  7. [7]

    X. Di, B. S. Riggan, S. Hu, N. J. Short, and V . M. Patel. Multi-scale thermal to visible face verification via attribute guided synthesis.IEEE Transactions on Biometrics, Behav- ior, and Identity Science, 3(2):266–280, 2021

    2021

  8. [8]

    Y . Han, J. Zhu, K. He, X. Chen, Y . Ge, W. Li, X. Li, J. Zhang, C. Wang, and Y . Liu. Face adapter for pre-trained diffusion models with fine-grained id and attribute control, 2024

    2024

Show all 35 references
  1. [9]

    Heusel, H

    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.Advances in neural information processing systems, 30, 2017

    2017

  2. [10]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  3. [11]

    S. Hu, N. Short, B. S. Riggan, M. Chasse, and M. S. Sar- fraz. Heterogeneous face recognition: recent advances in infrared-to-visible matching. In2017 12th IEEE Interna- tional Conference on Automatic Face & Gesture Recognition (FG 2017), pages 883–890. IEEE, 2017

    2017

  4. [12]

    Karkkainen and J

    K. Karkkainen and J. Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. InProceedings of the IEEE/CVF winter confer- ence on applications of computer vision, pages 1548–1558, 2021

    2021

  5. [13]

    Kwon and J

    G. Kwon and J. C. Ye. Diffusion-based image translation using disentangled style and content representation.arXiv preprint arXiv:2209.15264, 2022

    2022 arXiv

  6. [14]

    B. Li, K. Xue, B. Liu, and Y .-K. Lai. Bbdm: Image-to-image translation with brownian bridge diffusion models. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 1952–1961, 2023

    1952

  7. [15]

    S. Z. Li, Z. Lei, and M. Ao. The hfb face database for het- erogeneous face biometrics research. In2009 Proc. CVPRW, pages 1–8. IEEE, 2009

    2009

  8. [16]

    J. Lin, Y . Wu, Z. Wang, X. Liu, and Y . Guo. Pair-id: A dual modal framework for identity preserving image generation. IEEE Signal Processing Letters, 2024

    2024

  9. [17]

    J. Lin, G. Zhao, J. Xu, G. Wang, Z. Wang, A. Dantcheva, L. Du, and C. Chen. Difftv: Identity-preserved thermal-to- visible face translation via feature alignment and dual-stage conditions. InProc. ACM MM 2024, pages 10930–10938, 2024

    2024

  10. [18]

    F. Liu, M. Kim, A. Jain, and X. Liu. Controllable and guided face synthesis for unconstrained face recognition. In European Conference on Computer Vision, pages 701–719. Springer, 2022

    2022

  11. [19]

    K. Mei, Y . Mei, and V . M. Patel. Thermal to visible image synthesis under atmospheric turbulence. In2022 IEEE In- ternational Conference on Image Processing (ICIP), pages 2051–2055, 2022

  12. [20]

    N. G. Nair, K. Mei, and V . M. Patel. At-ddpm: Restor- ing faces degraded by atmospheric turbulence using denois- ing diffusion probabilistic models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3434–3443, 2023

    2023

  13. [21]

    N. G. Nair and V . M. Patel. T2v-ddpm: Thermal to visible face translation using denoising diffusion probabilistic mod- els. In2023 Proc. FG, pages 1–7. IEEE, 2023

    2023

  14. [22]

    Poster, M

    D. Poster, M. Thielke, R. Nguyen, S. Rajaraman, X. Di, C. N. Fondje, V . M. Patel, N. J. Short, B. S. Riggan, N. M. Nasrabadi, et al. A large-scale, time-synchronized visible and thermal face dataset. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer ...

    2021

  15. [23]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021

    2021

  16. [24]

    B. S. Riggan, N. J. Short, and S. Hu. Thermal to visible synthesis of face images using multiple regions. In2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 30–38. IEEE, 2018

    2018

  17. [25]

    B. S. Riggan, N. J. Short, S. Hu, and H. Kwon. Estimation of visible spectrum faces from polarimetric thermal faces. In2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–7, 2016

    2016

  18. [26]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolu- tional networks for biomedical image segmentation. InMed- ical image computing and computer-assisted intervention– MICCAI 2015: 18th international conference, Munich, Ger- many, October 5-9, 2015, proceedings, part III 1...

    2015

  19. [27]

    Schultz Santos

    R. Schultz Santos. Pantone: Identity formation through colours. 2014

    2014

  20. [28]

    M. Tan. Efficientnet: Rethinking model scaling for convo- lutional neural networks.arXiv preprint arXiv:1905.11946, 2019

    1905 arXiv

  21. [29]

    Van Den Oord, O

    A. Van Den Oord, O. Vinyals, et al. Neural discrete represen- tation learning.Advances in neural information processing systems, 30, 2017

    2017

  22. [30]

    S. Yang, L. Jiang, Z. Liu, and C. C. Loy. Gp-unit: Generative prior for versatile unsupervised image-to-image translation. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 45(10):11869–11883, 2023

    2023

  23. [31]

    H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image dif- fusion models.arXiv preprint arXiv:2308.06721, 2023

    2023 arXiv

  24. [32]

    Zhang, A

    L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text-to-image diffusion models, 2023

    2023

  25. [33]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a percep- tual metric. InProceedings of the IEEE conference on com- puter vision and pattern recognition, pages 586–595, 2018

    2018

  26. [34]

    Zhang, Y

    Z. Zhang, Y . Song, and H. Qi. Age progression/regression by conditional adversarial autoencoder. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017

    2017

  27. [2016]

    Special Issue on ICPR 2014 Awarded Papers

    2014

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.