Deepfake Generation and Detection: A Benchmark and Survey

Generation landscape

• Four mainstream tasks are face swapping [11] [13], face reenactment [15], talking face generation [16], and facial attribute editing [18] [19] [20]; methods evolve from GANs [5] to diffusion [7] [22] [23], images to video [8], and use NeRF [24] for multi-view consistency [26].Earlier VAE [2] [4] and GAN [5] [6] methods enabled realistic generation but had unsatisfactory performance; diffusion later greatly enhanced image and video generation.Source: §1 · pp. 1–2

Detection motivation and evolution

• Non-consensual and misleading uses create privacy, impersonation, and dissemination risks, driving demand for detection that keeps evolving alongside increasingly realistic generation.Detection has progressed from handcrafted features through deep learning to hybrid techniques, while expanding from spatial and frequency domains to temporal analysis.Source: §1 · p. 2

Contribution and scope

• The survey unifies task definitions and dataset and metric comparisons, reviews four major generation fields and forgery detection, and benchmarks representative works, especially diffusion-based approaches.It focuses on influential published works and also discusses related fields such as head swapping, face reconstruction, body animation, and adversarial sample detection.Source: §1 · p. 2

Relation to prior surveys

• The survey broadens earlier reviews by covering more tasks than character generation, a wider range of models—particularly diffusion—than related surveys, and related generation and detection sub-fields.It focuses on representative, attention-grabbing works because of the large number of published papers and preprints.Source: §1 · pp. 2–3

Survey methods and pipeline

• The authors searched major academic databases for studies published from 2020 to 2025, grouped pre-2020 work separately, and selected representative studies by venue quality, citation count, and technical novelty.The survey proceeds from background, definitions, datasets, metrics, applications, and ethics to task reviews, benchmark comparisons, challenges, and future directions.Source: §1 · p. 3

2.1 Problem Definition

• Deepfake generation is framed as conditional content generation, while detection is image- or pixel-level classification; the survey distinguishes five tasks by their targets and manipulated facial content.The tasks are face swapping, face reenactment, talking face generation, facial attribute editing, and forgery detection.Source: §2.1 · p. 4

2.2 Technological History and Roadmap

• VAEs model latent distributions, GANs generate through adversarial competition, and diffusion models generate by progressively denoising data.Conditional inputs and vector quantization extend VAE approaches; CGAN, Pix2Pix, StyleGAN, and StyleGAN2 extend GANs, while DDPM, LDM, SVD, and AnimateDiff develop diffusion generation [4] [5] [6] [22] [23] [7] [8].Source: §2.2 · pp. 4–5
• The technological roadmap also reviews CNNs, Transformers, and NeRF, including their roles in image modeling and 3D scene representation [24].Source: §2.2 · p. 5

2.3 Datasets, Metrics, and Losses

• The survey catalogs datasets for generation and forgery detection, task-specific evaluation metrics, and losses associated with VAE-, GAN-, and diffusion-based methods.Examples include PSNR, SSIM, LPIPS, and FID for generation; AUC, ACC, EER, AP, and F1-Score for forgery detection; and reconstruction, KL-divergence, adversarial, and denoising losses.Source: §2.3 · p. 6

2.4 Related Research Domains

• Related domains include head swapping, face super-resolution, 3D face reconstruction, and face inpainting, which support identity transfer, image quality, controllable manipulation, or restoration.Head swapping replaces the head while preserving non-identity attributes; super-resolution improves synthesized output quality [16]; reconstruction supports manipulation and mitigates pose artifacts; inpainting restores missing regions while preserving texture.Source: §2.4 · pp. 6–7
• Body animation transfers a driving signal to alter body pose, while portrait style and makeup transfer adapt source-image styles to a target.Makeup methods allow quantitative intensity control but perform poorly on extreme styles. Deepfake detectors often rely on a single cue, leaving adversarial-sample generalization and evaluation relatively understudied.Source: §2.4 · p. 7

2.5 Applications of Deepfake Detection

• YouTube uses automated deepfake classifiers for content moderation, and the Bank of China employs liveness detection to counter video-based impersonation fraud.Source: §2.5 · p. 7

2.6 Ethical and Societal Considerations

• The survey describes regulatory responses to synthetic-media risks, including EU transparency and platform obligations and China’s rules on watermarking, traceability, identity verification, and accountability.Source: §2.6 · p. 8

3.1 Deepfake Generation

• The generation survey organizes face-swapping methods by basic architecture and summarizes the categories in Table 2.The section also reviews methods for face reenactment, talking-face generation, and facial-attribute editing.Source: §3.1.1 · p. 8

3.1.1 Face Swapping

• Traditional face swapping matches facial keypoints or fits a 3D prior, but remains constrained by pose and lighting similarity, low resolution, attribute editing, occlusion, and extreme conditions.3DMM-based methods parameterize facial information, while other approaches render frontal-face models at arbitrary poses or map 2D graphics densely.Source: §3.1.1 · p. 9
• GAN-based face swapping improves generalization, resolution, and geometric detail with region-specific processing, layered encodings, semantic constraints, and 3D priors, yet occlusion and identity–attribute trade-offs persist.FSGAN combines reenactment and swapping; MegaFS targets million-pixel output, while HifiFace and FlowFace incorporate shape-aware information.Source: §3.1.1 · pp. 9–10

3.1.2 Face Reenactment

• Face reenactment methods use 3DMMs, landmark mapping, latent-feature decoupling, and self-supervision to transfer facial motion while addressing artifacts, large poses, and occlusion.Examples include PECHead’s motion-aware feature alignment, MetaPortrait’s dense keypoint matching, and StyleMask’s separation of pose and expression from identity.Source: §3.1.2 · pp. 10–12

3.1.3 Talking Face Generation

• Talking-face methods extend audio-driven lip synthesis toward control of identity, pose, style, and emotion, while multimodal conditioning can improve vividness but still faces quality and integration challenges.MakeItTalk predicts landmarks from audio, SadTalker uses audio-derived 3DMM motion, and EMMN links emotion embeddings with lip movements; emotion intensity remains difficult to control.Source: §3.1.3 · pp. 12–13
• Diffusion and 3D representations support fine-grained or emotion-editable talking faces, direct audio-to-video generation, and explicit facial reconstruction and motion modeling.DAE-Talker uses learned latent representations, EmoTalker adds emotion-intensity blocks, and 3D Gaussian Splatting is applied to refine facial reconstruction and motion details.Source: §3.1.3 · p. 13

3.1.4 Facial Attribute Editing

• Facial-attribute editing moves from single-attribute operations toward comprehensive editing, with disentanglement and retention of unrelated target information as central challenges.HifaFace uses wavelet-based editing, TransEditor uses a dual-space GAN, and IA-FaceS separates component-invariant detail embeddings from component-specific operations.Source: §3.1.4 · p. 14
• Text-driven and diffusion-based methods expand facial editing through text-to-style mapping, cross-modal alignment, multimodal generation, and identity-preserving personalized editing.TG-3DFace uses global contrastive learning and fine-grained alignment, while DiffusionRig conditions an initial 3D face model to preserve identity.Source: §3.1.4 · p. 15

3.2.1 Space Domain .

• Spatial methods detect local image-detail discrepancies and altered noise, using cues such as color, artifacts, gradients, textures, and noise traces.RECCE uses learned representations of real samples to identify reconstruction differences, while LGrad classifies gradient visualizations of artifacts.Source: §3.2.1 · pp. 15–16

3.2.2 Time Domain .

• Temporal detectors assess physiological plausibility, frame-to-frame inconsistencies, and multimodal mismatches.Methods use cues including blinking, head pose, gaze, inter-frame feature or style changes, and audio-visual relationships.Source: §3.2.2 · pp. 16–17

3.2.3 Frequency Domain .

• Frequency-domain methods use periodic features and frequency distributions to detect forgeries; F3-Net, HFI-Net, and FreqNet learn frequency-specific clues or source-independent features.Other approaches combine spatial, temporal, and frequency cues, including Gu et al.’s spatial-frequency interaction convolution.Source: §3.2.3 · p. 17

3.2.4 Data Driven .

• Data-driven methods learn distinguishing patterns from datasets, exploit model fingerprints, monitor neuron activation with FakeSpotter, or classify component sources and identity differences.FakeSpotter applies a binary classifier to neuron behavior in a deep face-recognition system.Source: §3.2.4 · p. 17

3.3.1 Face Super-resolution .

• For face super-resolution, CNNs learn low-to-high-resolution mappings using global, local, or mixed methods, while GANs dominate for flexible, efficient architectures.Source: §3.3.1 · p. 17

3.3.2 Portrait Style Transfer .

• GANs are the most mature approach to portrait style transfer, but instability can produce artifacts and implausible components.3DAvatarGAN bridges pre-trained source-domain 3D-GANs with 2D-GANs trained on artistic data; Scenimefy guides StyleGAN with CLIP semantic constraints and patch-based contrastive style loss. Diffusion-based cross-domain transfer includes DiffusionGAN3D, which combines 3DGAN with text-to-graphics diffusion, relative-distance loss, and learnable tri-planes.Source: §3.3.2 · pp. 17–18
• Body-animation GANs transfer information from reference images to target images, but remain sample-dependent and perform worse on rare poses.Diffusion-based body animation is an active research topic; LEO uses a Latent Motion Diffusion Model to represent motion as flow graphs and focus on spatiotemporal continuity.Source: §3.3.3 · p. 18

3.3.4 Makeup Transfer .

• Traditional graphics-based makeup transfer warps reference makeup using facial landmarks, but manual operators can cause artifacts and unintended background changes.Early deep-learning methods struggled with large pose or expression differences and extreme makeup; PSGAN++ [161] adds makeup distillation, attentive morphing, style transfer, and identity extraction to PSGAN, while ELeGANt, CUMTGAN, and HT-ASE explore detail preservation.Source: §3.3.4 · p. 18

4.1 Metrics

• Metrics are task-specific: swapping assesses identity, expression, pose, and image quality; reenactment adds similarity, reconstruction, and lip-shape measures, while talking-face and editing metrics target synchronization and text consistency.Talking-face metrics also assess identity, pose, expression, and gaze; facial attribute editing commonly uses FID, LPIPS, KID, PSNR, and SSIM.Source: §4.1 · pp. 18–19

4.2 Benchmark Protocol

• Face swapping is benchmarked on FF++, while reenactment uses VoxCeleb for self- and cross-identity tests and VoxCeleb2 for quality assessment.The swapping test samples 10 frames from each of 1,000 videos; cross-identity reenactment selects 35 different-identity video pairs. The survey reports training-data differences and relevant preprocessing because they affect comparability.Source: §4.2 · pp. 19–20, Table 7 · p. 20, Table 9 · p. 20
• Talking-face generation uses MEAD, attribute editing reports original results without fixed training or validation splits, and forgery detection evaluates self-dataset and cross-dataset performance.Talking-face methods differ in MEAD use and validation splits; detection reports AUC and ACC on FF++ HQ/LQ and AUC on DFDC, Celeb-DF, Celeb-DFv2, and DeeperForensics-1.0.Source: §4.2 · p. 20, Table 11 · p. 21, Table 13 · p. 21

4.3 Main Results on Deepfake Generation

• WSC-Swap preserves identity and pose but has expression-error weaknesses; AVFR-GAN benefits from multimodal modeling, while HiDe-NeRF decouples pose and expression for accurate attribute modeling.The survey reports face-swapping results on FF++ and reenactment results for self-reenactment, cross-identity reenactment, and quality assessment.Source: Table 7 · p. 20, Table 9 · p. 20, Table 8 · p. 20
• AMIGO uses emotion-tagged landmark generation and landmark-to-image translation for textured talking faces; attribute-editing results are demonstrations, not fair comparisons across datasets.The MEAD benchmark results vary across methods, while attribute-editing tables report FID and reconstruction metrics.Source: §4.3 · p. 21, Table 11 · p. 21, Table 13 · p. 21

4.4 Main Results on Forgery Detection

• On FF++, LipForensics performs robustly on LQ, Guo et al. perform best on HQ; AVoiD-DF and Zhao et al. generalize well cross-dataset, though DFDC remains challenging.The cross-dataset results cover DFDC, Celeb-DF, Celeb-DFv2, and DeeperForensics-1.0 using AUC.Source: §4.4 · pp. 21–22, Table 14 · p. 21, Table 15 · p. 22

Face Swapping

• Face swapping still suffers cross-dataset performance degradation and lacks a unified evaluation protocol, motivating comprehensive frameworks for fair comparison.Models that perform well on training data may degrade on different test datasets.Source: §5 · p. 22

Face Reenactment

• Face reenactment must balance lightweight deployment with quality, maintain authenticity under rapid pose changes and extreme lighting, and reduce computational costs.These challenges impede industrial adoption and practical use.Source: §5 · p. 22

Talking Face Generation

• Talking-face methods need finer emotional control and better alignment of intonation, audio semantics, head pose, and expression to reduce artifacts in strongly emotional speech.Source: §5 · p. 22

Facial Attribute Editing

• Facial attribute editing must better disentangle attributes to avoid unintended changes to other features, and needs an accepted benchmark dataset and evaluation framework.GAN-based decoupling is mainstream, while diffusion models are gradually being introduced.Source: §5 · p. 22

Forgery Detection

• Forgery detection needs single models that identify diverse manipulations robustly under disturbances such as compression, while overcoming limited generalization and evasion as techniques evolve.Current generic approaches target common steps such as face-swapping integration or temporal inconsistencies, which limits generalization.Source: §5 · pp. 22–23

Discussion

• Future systems should combine multimodal generation and detection, broader datasets, feedback-driven optimization, and finer temporal modeling, while using transparency, misuse detection, and governance to address ethical risks.Generation may integrate visual, auditory, and textual cues; detection may use audio–video synchronization, speech semantics, and physiological signals, with self- or weakly supervised learning reducing annotation needs. Potential applications include digital humans, telepresence, assistive media creation, and privacy-preserving data synthesis.Source: §5 · p. 23

• The survey summarizes task concepts, generative-model and neural-network development, and other background from closely related fields.Source: §6 · p. 23
• It classifies and discusses technical approaches across four mainstream generation fields and one detection field.Source: §6 · p. 23
• It organizes and benchmarks representative methods in each field.Source: §6 · p. 23
• It summarizes current challenges and future research directions for each field.Source: §6 · p. 23