Pith. sign in

REVIEW 4 major objections 7 minor 38 references

Diff-ID keeps generated faces identity-consistent and more photorealistic than compared methods, winning on a joint identity–realism score rather than raw face similarity alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 01:54 UTC pith:CGJMAV4C

load-bearing objection Solid incremental face-ID diffusion system with honest FS reporting, but the headline FID/FIQ win is partly confounded by in-domain fine-tuning against off-the-shelf baselines. the 4 major comments →

arxiv 2607.25078 v1 pith:CGJMAV4C submitted 2026-07-27 cs.CV

Diff-ID: Identity Consistent Facial Image Generation and Morphing via Diffusion Models

Taimoor Rizwan , Sara Atito , Muhammad Awais , Zhenhua Feng , Josef Kittler This is my paper
classification cs.CV
keywords Diffusion modelsIdentity-preserving generationFace morphingArcFaceCLIPStable DiffusionFace Image QualityDDIM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Diffusion models can paint realistic faces, but they often drift away from a specific person’s identity—an unacceptable failure for biometrics, privacy-safe synthetic data, and security testing. Diff-ID fine-tunes a Stable Diffusion backbone with a lightweight dual cross-attention adapter that fuses ArcFace identity embeddings and CLIP semantic embeddings, trained on a custom 210K captioned face corpus. An ArcFace cosine-similarity “pseudo discriminator” loss, weighted more heavily as noise falls during denoising, further locks identity. On held-out and unseen faces the method does not beat InstantID on raw Face Similarity, but it posts the lowest FID and the highest FIQ (similarity divided by FID) among the compared systems, and it can morph between two faces via DDIM without per-person retraining. The authors argue that identity and realism must be scored together, because high similarity alone can reward unrealistic or stylized faces.

Core claim

Among the evaluated identity-guided diffusion methods, Diff-ID does not lead in raw ArcFace Face Similarity, but it achieves substantially lower FID and the highest FIQ-based identity–realism trade-off on both validation and unseen splits, while supporting qualitative face morphing in one unified model without per-identity fine-tuning.

What carries the argument

A dual cross-attention adapter that fuses projected ArcFace and CLIP embeddings inside a fine-tuned Stable Diffusion UNet, trained with an ArcFace cosine-similarity identity loss under exponential timestep weighting; FIQ = 100 · FS / FID is reported as a secondary joint score.

Load-bearing premise

The ranking depends on treating the evaluation splits as fair and identity-disjoint, even though the validation set comes from the same training pool and the “unseen” set is only assumed disjoint without a verified identity audit.

What would settle it

Re-run the same caption-conditioned generation protocol on a strictly identity-audited, non-overlapping face set: if Diff-ID no longer has the lowest FID and highest FIQ among the same baselines while InstantID still leads raw FS, the claimed trade-off advantage fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic training and test faces for recognition systems can be generated with stronger measured photorealism without sacrificing competitive identity retention.
  • Privacy-preserving biometric pipelines can substitute real subject images with Diff-ID outputs when identity geometry and realism both matter.
  • Smooth identity morphs for robustness or morph-attack study become available from a single checkpoint without DreamBooth-style per-person tuning.
  • Reporting FS and FID separately, with FIQ only as a secondary ratio, becomes a clearer standard than optimizing identity similarity alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If FIQ-style joint scoring were adopted more widely, methods that game ArcFace with stylized or over-sharpened faces would lose leaderboard advantage even when raw FS looks strong.
  • The same adapter-plus-weighted-identity-loss pattern could transfer to transformer diffusion backbones (e.g., DiT) with only injection-point changes, as the paper itself notes architecturally.
  • Without verified demographic and consent audits on the 210K mix, deployment for biometric synthetic data still carries bias and licensing risk the metrics do not capture.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper presents Diff-ID, an identity-conditioned face generation framework built on Stable Diffusion 1.5. A dual cross-attention adapter fuses ArcFace identity embeddings with CLIP image/text embeddings (without softmax normalization), refined by a Fusion MLP and injected into a fine-tuned UNet; training adds an ArcFace cosine-similarity "pseudo-discriminator" loss with exponential timestep weighting, on a new 210K BLIP-captioned corpus drawn from CelebA-HQ/FFHQ/LAION-Face. The headline empirical claim (Table 1) is deliberately modest: Diff-ID does not beat InstantID on raw ArcFace Face Similarity, but achieves the lowest FID and the highest value of a proposed composite FIQ = 100·FS/FID on both an in-domain validation split and an LFW "unseen" split, i.e., the best measured identity–realism trade-off. A DDIM-inversion morphing pipeline with joint latent/embedding interpolation is presented qualitatively, without per-identity fine-tuning.

Significance. If the evaluation holds up, the paper makes a useful, honestly framed contribution: it explicitly concedes the raw-FS lead to InstantID, argues for joint identity–realism evaluation, and backs the composite-metric conclusion with a sensitivity analysis over six alternative formulations (Table 2), which is more rigor than most composite-score proposals receive. The adapter ablation (Table 3) isolating ArcFace-only, CLIP-only, Joint, and Joint+Fusion-MLP variants is a genuine strength, as is the qualitative morphing pipeline that avoids per-identity checkpoints. The 210K captioned corpus could be useful if released. Offsetting this: no code, data, or model release is mentioned; FIQ is unvalidated against human judgment (the authors say so); and the central FID result is exposed to a train/eval distribution-match confound detailed below. The result is incrementally significant for the identity-preserving generation subfield rather than a conceptual advance.

major comments (4)
  1. [§4.2, Table 1] The central claim (lowest FID, hence highest FIQ, on both splits) is confounded by a train/eval distribution match. §4.2 states the validation split is sampled from the same curated CelebA-HQ/FFHQ/LAION-Face pool Diff-ID was fine-tuned on for 1M steps, while all baselines are off-the-shelf models trained on other data. FID measures distance to a reference distribution; a model fine-tuned on (a superset of) the reference pool will score lower almost by construction, independent of perceptual realism. The FID protocol itself is never specified: reference set composition and size, image resolution/alignment, Inception network and feature layer, and number of generated samples are all missing. The absolute values (>100 on both splits) are an order of magnitude above typical face-generation FIDs, which strongly suggests a non-standard setup that must be documented. Required: (a) full FID prot
  2. [§4.2] Identity/image disjointness of the 'unseen' LFW split is asserted as a best-effort dataset-level assumption with no audit. LAION-Face is a web crawl, and LFW images are heavily replicated on the public web, so the training pool may contain LFW images or near-duplicates outright. Since the unseen split is the only control for the in-domain confound above, this is load-bearing. A concrete, cheap fix: run an ArcFace nearest-neighbor / duplicate-detection pass between the 210K training pool and the 5,000 LFW evaluation images, report the overlap statistics, and remove any matches. If an audit is infeasible, the 'unseen' framing should be softened throughout (abstract, §4.2, §7).
  3. [§3.4, Eq. (15)–(17)] The pseudo-discriminator identity loss is under-specified to the point of being non-reproducible. Computing CosineSim(ArcFace(original), ArcFace(generated)) during training requires an image-space 'generated' sample at arbitrary timestep t, but the paper never states how this is obtained: is the predicted x0 decoded through the frozen VAE at every training step (expensive), is a one-step estimate used, is the ArcFace gradient path differentiable end-to-end, and does the exponential weighting apply to the sampling distribution of t or only to the loss magnitude? The description of Wt = exp(−kτ) is also internally ambiguous: with τ = t/T and t large = high noise, exp(−kτ) is smallest at high noise, which matches the text, but the value of k is never given. These details determine whether the loss behaves as claimed in Fig. 10 and must be added.
  4. [§4.1.2, Table 1] ArcFace serves simultaneously as (i) the training identity loss, (ii) the conditioning embedding source, and (iii) the sole evaluation metric for identity (FS). This is a mild but real circularity: Diff-ID is directly optimized for the metric it is ranked on, whereas baselines are not. The authors are transparent about InstantID winning raw FS, which mitigates the impact, but a second, independently trained face-recognition backbone (e.g., a non-ArcFace model such as a CurricularFace/MagFace or a different-architecture recognizer) used only at evaluation time would substantially strengthen the identity claims on both splits. This is a standard control in this literature and is feasible within the paper's scope.
minor comments (7)
  1. [§3.3.2, Eq. (5)] The claim that omitting softmax 'preserves raw magnitude relationships... helpful for fine-grained details' is unmotivated and never ablated; Table 3 varies embedding sources and the Fusion MLP but not the softmax. Either add a softmax-vs-no-softmax ablation or soften the claim to a design choice.
  2. [References] Several citations are placeholders or incorrect: [5], [6], [7] cite 'arXiv:2303.XXXX / 2304.XXXX'; [3] (GLIDE) cites arXiv:2112.10752, which is the Latent Diffusion paper ([2]); [8] attributes ControlNet to the PhotoMaker authors; [36] lists an author set for DreamBooth that does not match the actual paper (Ruiz et al.). Please audit the full bibliography.
  3. [§4.1.3, Eq. (23)] 'FIQ' collides with the established face-image-quality assessment literature (e.g., SER-FIQ and related ISO/IEC 29794-1 quality scores), where the term denotes a well-defined, standardized concept. A distinct name (e.g., 'identity–realism ratio') would avoid confusion, especially given the paper's stated intention not to validate the score against human or biometric judgments.
  4. [Table 3, §5.2] The adapter ablation reports FS only, on the in-domain validation split. Since the paper's thesis is the identity–realism trade-off, FID (and ideally unseen-split numbers) for the four variants would make the ablation commensurate with the main claim; the CLIP-only variant in particular might trade FS for FID.
  5. [§4.2] Inference configuration for evaluation is not reported: number of DDIM/sampling steps, guidance scale, seed count, and whether one or multiple images per identity were generated and how FS was aggregated. These affect both FS and FID comparability with baselines.
  6. [§3.1] 'resized to a uniform resolution of 512' — presumably 512×512 pixels; also state whether face crops/alignment were applied, since this affects both training and the FID reference statistics.
  7. [Fig. 6 / Fig. 8] Fig. 6 reports per-image FID values (e.g., 101.3 for a single image), but FID is a distributional statistic and is undefined for one image; clarify that these are per-subject estimates over a sample set, and state the sample count. In Fig. 8, please state the α values corresponding to each shown frame.

Circularity Check

0 steps flagged

No derivation-chain circularity: empirical ML comparisons on external metrics, with openly conceded FS loss to InstantID.

full rationale

Diff-ID is an empirical systems paper, not a first-principles derivation. Its load-bearing claims are experimental rankings on ArcFace Face Similarity, FID, and a secondary ratio FIQ = 100·FS/FID (Eq. 23, Table 1). FS and FID are standard external metrics computed against held-out/unseen image sets; the paper explicitly reports that InstantID wins raw FS and that Diff-ID’s advantage is FID/FIQ. Defining FIQ as a complementary summary and ranking by it is a metric-design choice, not a case where a predicted quantity equals a fitted input by construction—alternative composites in Table 2 are reported and InstantID could have won FIQ had its FS margin dominated. Using ArcFace both as a training identity loss and as the FS evaluator is common practice and creates possible evaluation bias, but it is not self-definitional circularity (the loss does not algebraically force the reported FS ranking against independently trained baselines). There is no self-citation uniqueness theorem, no ansatz smuggled in via author-overlapping prior work, and no renaming of a known closed-form result. Confounds such as fine-tuning on the same pool used for the validation FID reference (§4.2) are validity/generalization concerns, not circular reductions of a claimed derivation. Score 0 with empty steps is the warranted finding.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 4 invented entities

The claim is empirical systems work. It inherits face-recognition and diffusion tooling as domain assumptions, then adds hand-chosen training weights, an author-defined composite score, and architectural modules whose value is justified mainly by in-paper ablations rather than external theory.

free parameters (5)
  • lambda_identity = 0.10
    Balances denoising loss against ArcFace identity loss in L_total; chosen by qualitative sweep and fixed for final model.
  • identity_loss_decay_k = not numerically specified
    Controls exponential timestep weighting W_t = exp(-k τ); rate is a hand-set hyperparameter affecting when identity supervision dominates.
  • Adam learning rate / batch size / train steps = 1e-5; batch 16; 1,000,000 steps
    Optimization hyperparameters that determine the final adapter/UNet fit on the custom corpus.
  • FIQ scale factor 100 and composite form FS/FID = FIQ = 100 · FS / FID
    Author-chosen secondary metric form and scale used to declare the best identity–realism trade-off; not derived from theory or human studies.
  • projection/adapter weight matrices (W_arc, W_i, W_s, attention WQs/Ks/Vs, Fusion MLP) = trained end-to-end (values not reported)
    Learned parameters fit to the 210K dataset; the central performance claim depends on this fitted adapter rather than a closed-form predictor.
axioms (6)
  • domain assumption ArcFace cosine similarity is a valid primary proxy for identity preservation of generated faces.
    Used both as training pseudo-discriminator and as FS metric throughout §3.4 and §4.
  • domain assumption FID on Inception features is a valid primary proxy for photorealism of face images.
    Paired with FS to justify FIQ and realism claims in §4.
  • domain assumption Frozen CLIP text/image embeddings supply useful semantic conditioning without needing explicit attribute control modules.
    Core conditioning design in §3.3; paper disclaims fine-grained attribute editing.
  • ad hoc to paper Training and LFW evaluation identities can be treated as disjoint at dataset level without a formal identity-matching audit.
    Explicitly stated as assumption/best-effort in §4.2; load-bearing for 'unseen' generalization.
  • ad hoc to paper Omitting softmax in dual cross-attention preserves fine-grained magnitude relationships helpful for identity detail.
    Design justification in §3.3.2; not independently validated outside ablations.
  • standard math Standard latent diffusion / DDIM inversion mathematics from prior work holds for the fine-tuned UNet+adapter.
    Denoising loss and morph inversion/sampling follow Ho/Rombach/Song-style formulations cited in §3.
invented entities (4)
  • Diff-ID dual cross-attention adapter + Fusion MLP no independent evidence
    purpose: Fuse ArcFace identity and CLIP semantics into Stable Diffusion cross-attention for identity-consistent generation.
    Main architectural contribution; evidence is in-paper FS ablations (Table 3) and qualitative grids, not external theory.
  • Pseudo-discriminator ArcFace loss with exponential timestep weighting no independent evidence
    purpose: Increase identity supervision as noise decreases during denoising.
    Introduced in §3.4; supported by training-curve/weight sweeps, not an established standard loss.
  • Face Image Quality (FIQ) score no independent evidence
    purpose: Summarize identity–realism trade-off as a single secondary ratio.
    Defined as 100·FS/FID in §4.1.3; authors say it is diagnostic only and unvalidated against humans or biometric decisions.
  • 210K BLIP-captioned identity-centric face corpus no independent evidence
    purpose: Provide identity-aware training pairs from CelebA-HQ/FFHQ/LAION-Face after cleaning.
    Custom dataset central to training; not released or externally benchmarked in the paper.

pith-pipeline@v1.2.0-grok45-kimik3 · 21595 in / 4046 out tokens · 84576 ms · 2026-07-31T01:54:29.614446+00:00 · methodology

0 comments
read the original abstract

Generative diffusion models have revolutionized facial image synthesis, yet robust identity preservation in high resolution outputs remains a critical challenge. This issue is especially vital for security systems, biometric authentication, and privacy sensitive applications, where any drift in identity integrity can undermine trust and functionality. We introduce Diff-ID, a diffusion based framework that enforces identity consistency while delivering photorealistic quality. Central to our approach is a custom 210K image dataset synthesized from CelebA-HQ, FFHQ, and LAION-Face and captioned via a fine tuned BLIP model to bolster identity awareness during training. Diff-ID integrates ArcFace and CLIP embeddings through a dual cross attention adapter within a fine tuned Stable Diffusion UNet. To further reinforce identity fidelity, we propose a pseudo discriminator loss based on ArcFace cosine similarity with exponential timestep weighting. Experiments on held out and unseen faces show that Diff-ID does not exceed InstantID in raw ArcFace Face Similarity, but achieves substantially lower FID and the strongest FIQ based identity--realism trade off among the evaluated methods. We also present a unified DDIM based morphing pipeline that enables qualitative facial interpolation without per identity fine tuning. We further argue that identity preservation and photorealism should be evaluated jointly rather than in isolation, as high identity similarity alone does not guarantee realistic outputs. To make this trade off explicit, we report Face Image Quality (FIQ) as a complementary ratio based score that combines identity similarity and perceptual realism while keeping FS and FID as the primary metrics.

Figures

Figures reproduced from arXiv: 2607.25078 by Josef Kittler, Muhammad Awais, Sara Atito, Taimoor Rizwan, Zhenhua Feng.

Figure 1
Figure 1. Figure 1: Data pipeline used to curate and caption the identity centric dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diff-ID architecture. We extract semantic features from CLIP and identity features from ArcFace, project [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the exponential timestep weighting [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: DiffID Morph: identity preserving morphing pipeline built on DDIM inversion and dual embedding [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Face Similarity (FS) for one subject under four stylistic variations. FS remains high even for non-photorealistic [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Identity-specific comparison for one subject (ID1) across different models. Metrics (FS, FID, and FIQ) are [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Identity preservation: comparison across IP Adapter, PhotoMaker, InstantID, Arc2Face, and Diff-ID. All [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of morphing methods. Top: embedding space interpolation. Middle: linear interpolation. Bottom: [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Adapter ablation: identity preservation across variants. DiffID Joint combines ArcFace and CLIP image [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Training dynamics with and without the identity loss term. Exponential timestep weighting concentrates [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Effect of identity regularization strength ( [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 10 linked inside Pith

  1. [1]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems, 2020

    2020

  2. [3]

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

    Alexander Quinn Nichol and Prafulla Dhariwal. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. InarXiv preprint arXiv:2112.10752, 2021

    Pith/arXiv arXiv 2021

  3. [4]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Ghasemipour, Aditya Ramesh, Ilya Sutskever, Ruslan Salakhutdinov, et al. Photorealistic text-to-image diffusion models with deep language understanding. InProceedings of the 39th International Conference on Machine Learning (ICML), 2022

    2022

  4. [5]

    Ip-adapter: Learning to adapt pretrained text-to-image models for personalized image editing

    Jian Liu et al. Ip-adapter: Learning to adapt pretrained text-to-image models for personalized image editing. In arXiv preprint arXiv:2303.XXXX, 2023. 18 Diff-ID: Identity Consistent Facial Generation and Morphing

    2023

  5. [6]

    Photomaker: Stacked-id embeddings for enhanced identity preservation in diffusion models

    Xi Zhang et al. Photomaker: Stacked-id embeddings for enhanced identity preservation in diffusion models. In arXiv preprint arXiv:2303.XXXX, 2023

    2023

  6. [7]

    Instantid: Fast and accurate identity preservation in text-to-image models

    Wei Chen et al. Instantid: Fast and accurate identity preservation in text-to-image models. InarXiv preprint arXiv:2304.XXXX, 2023

    2023

  7. [8]

    Controlnet: Adding conditional control to diffusion models

    Xi Zhang et al. Controlnet: Adding conditional control to diffusion models. InarXiv preprint arXiv:2301.11368, 2023

    Pith/arXiv arXiv 2023

  8. [9]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, and Mark Chen. Zero-shot text-to-image generation. InarXiv preprint arXiv:2102.12092, 2021

    Pith/arXiv arXiv 2021

  9. [10]

    Cogview: Mastering text-to-image generation via transformers

    Zizhao Ding, Yunzhi Wang, Yuxiao Dong, Xiaodong Dong, Shengjie Tang, Feng Wu, Xunyu Lu, Gang Liu, Xiong Zhu, and Jian Gao. Cogview: Mastering text-to-image generation via transformers. InarXiv preprint arXiv:2105.13290, 2021

    Pith/arXiv arXiv 2021

  10. [12]

    Score-based generative modeling through stochastic differential equations

    Yang Song and Stefano Ermon. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations (ICLR), 2021

    2021

  11. [13]

    Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 2022

    Pith/arXiv arXiv 2022

  12. [14]

    Adding conditional control to text-to-image diffusion models.arXiv preprint arXiv:2302.05543, 2023

    Ze Zhang, Ding Zhu, Jorge Martinez, Jian Yang, Yixin Wang, Yi Zhang, and Li Fei-Fei Sun. Adding conditional control to text-to-image diffusion models.arXiv preprint arXiv:2302.05543, 2023

    Pith/arXiv arXiv 2023

  13. [15]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning (ICML), 2021

    2021

  14. [16]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Yuxiao Ni, Stefanos Zafeiriou, and Sameer Kumar. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2019

    2019

  15. [17]

    Uv-gan: Adversarial facial uv map completion for pose-invariant face recognition.arXiv preprint arXiv:1905.04150, 2019

    Jiankang Deng, Grigorios Chrysos, Evangelos Ververas, Jie Shen, Irene Kotsia, and Stefanos Zafeiriou. Uv-gan: Adversarial facial uv map completion for pose-invariant face recognition.arXiv preprint arXiv:1905.04150, 2019. Available athttps://arxiv.org/abs/1905.04150

    Pith/arXiv arXiv 1905

  16. [18]

    Ms-celeb-1m: A dataset and benchmark for large-scale face recognition

    Yandong Guo, Lei Zhang, Yuxiao Yuan, Jun Liu, Tao Wang, Yu Cai, Jessey Chen, Zhangyang Wang, Zhen Zhang, and Huabin Su. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. InEuropean Conference on Computer Vision (ECCV), pages 87–102, 2016

    2016

  17. [19]

    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. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4401–4410, 2019

    2019

  18. [20]

    Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows

    Rinon Abdal, Yipeng Qin, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13153–13162, 2021

    2021

  19. [21]

    Ganspace: Discovering interpretable gan controls

    Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, pages 9841–9850, 2020

    2020

  20. [22]

    Diffmorpher: Identity morphing via diffusion models

    Jinwoo Lee, Hyun Kim, and Junseok Choi. Diffmorpher: Identity morphing via diffusion models. InNeurIPS Workshop on Generative Models, 2024

    2024

  21. [23]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 5966–5978, 2021

    2021

  22. [24]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li et al. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InarXiv preprint arXiv:2201.12086, 2022

    Pith/arXiv arXiv 2022

  23. [25]

    Facial geometric detail recovery via implicit representation

    Xingyu Ren, Alexandros Lattas, Baris Gecer, Jiankang Deng, Chao Ma, and Xiaokang Yang. Facial geometric detail recovery via implicit representation. In2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), 2023

    2023

  24. [26]

    Sample and computation redistribution for efficient face detection.arXiv preprint arXiv:2105.04714, 2021

    Jia Guo, Jiankang Deng, Alexandros Lattas, and Stefanos Zafeiriou. Sample and computation redistribution for efficient face detection.arXiv preprint arXiv:2105.04714, 2021. 19 Diff-ID: Identity Consistent Facial Generation and Morphing

    Pith/arXiv arXiv 2021

  25. [27]

    Ostec: One-shot texture completion

    Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Ostec: One-shot texture completion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  26. [28]

    Killing two birds with one stone: Efficient and robust training of face recognition cnns by partial fc

    Xiang An, Jiangkang Deng, Jia Guo, Ziyong Feng, Xuhan Zhu, Yang Jing, and Liu Tongliang. Killing two birds with one stone: Efficient and robust training of face recognition cnns by partial fc. InCVPR, 2022

    2022

  27. [29]

    Partial fc: Training 10 million identities on a single machine

    Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial fc: Training 10 million identities on a single machine. InICCVW, 2021

    2021

  28. [30]

    Sub-center arcface: Boosting face recognition by large-scale noisy web faces

    Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. InProceedings of the IEEE Conference on European Conference on Computer Vision, 2020

    2020

  29. [31]

    Retinaface: Single-shot multi-level face localisation in the wild

    Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. InCVPR, 2020

    2020

  30. [32]

    Stacked dense u-nets with dual transformers for robust face alignment

    Jia Guo, Jiankang Deng, Niannan Xue, and Stefanos Zafeiriou. Stacked dense u-nets with dual transformers for robust face alignment. InBMVC, 2018

    2018

  31. [33]

    The menpo benchmark for multi-pose 2d and 3d facial landmark localisation and tracking.IJCV, 2018

    Jiankang Deng, Anastasios Roussos, Grigorios Chrysos, Evangelos Ververas, Irene Kotsia, Jie Shen, and Stefanos Zafeiriou. The menpo benchmark for multi-pose 2d and 3d facial landmark localisation and tracking.IJCV, 2018

    2018

  32. [34]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InCVPR, 2019

    2019

  33. [35]

    Denoising diffusion implicit models

    Yang Song and Stefano Ermon. Denoising diffusion implicit models. InAdvances in Neural Information Processing Systems, volume 33, 2020

    2020

  34. [36]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation.arXiv preprint arXiv:2208.12242, 2022

    Yatao Ru, Samuel Lombardi, Brendan McMahan, Alice Shih, Xiang Gao, Antoine Miech, and Christoph Feichtenhofer. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation.arXiv preprint arXiv:2208.12242, 2022

    Pith/arXiv arXiv 2022

  35. [37]

    A survey on face morphing attacks: Vulnerability exposure in biometric systems.IEEE Transactions on Information F orensics and Security, 14(8):2156–2179, 2019

    Celine Rathgeb and Christoph Busch. A survey on face morphing attacks: Vulnerability exposure in biometric systems.IEEE Transactions on Information F orensics and Security, 14(8):2156–2179, 2019

    2019

  36. [38]

    Franco, and Davide Maltoni

    Matteo Ferrara, Giovanni L. Franco, and Davide Maltoni. The magic passport. InProc. International Workshop on Biometrics and F orensics (IWBF), pages 1–6. IEEE, 2014

    2014

  37. [39]

    A morphable model for the synthesis of 3d faces

    V olker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. InACM SIGGRAPH, pages 187–194. ACM, 1999

    1999

  38. [40]

    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. InAdvances in Neural Information Processing Systems, volume 30, 2017. 20

    2017