Generative Face Video Coding (GFVC)

• GFVC is a video compression paradigm that leverages compact facial priors and deep generative decoders to efficiently reconstruct talking-head sequences.
• It integrates conventional key-frame coding with sparse inter-frame representations to achieve ultra-low bitrate transmission and perceptual optimization.
• GFVC methods employ diverse facial representations—such as 2D keypoints, 3D semantics, and latent spaces—and extend to multi-reference, layered, and joint audio-video coding.

Searching arXiv for recent GFVC papers to ground the article in current literature. Generative Face Video Coding (GFVC) is a model-based compression paradigm for talking-head sequences in which inter frames are represented by compact facial priors or latent codes and reconstructed by a deep generative decoder conditioned on one or more conventionally coded key-reference frames. In the unified formulation reported in recent reviews, a key frame is coded by a standard codec such as VVC, an analysis module extracts a low-dimensional representation for each non-key frame, and a generator synthesizes the reconstructed frame from the decoded key frame and decoded representation (Chen et al., 9 Jun 2025, Chen et al., 2023). Across the literature, GFVC is characterized by semantic-aware representation, perceptual optimization, and operation in ultra-low-bitrate regimes; several works also extend the paradigm toward layered scalability, multi-reference prediction, joint audio-video coding, standardization through Supplemental Enhancement Information (SEI), and low-complexity deployment (Chen et al., 2024, Chen et al., 24 Feb 2025).

1. Paradigm and formal definition

A generalized GFVC system partitions a face-only sequence into key-reference frames and generative inter frames. For key frames, a conventional codec produces reconstructed references. For each inter frame, the encoder extracts a compact facial prior $s_t$ or $\theta_1^t$, entropy-codes it, and the decoder synthesizes the frame with a generator conditioned on the decoded key frame and decoded prior (Chen et al., 2023, Chen et al., 9 Jun 2025). The formal decomposition reported in the surveys is

$\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$

$$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$

$$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$

The associated training criterion is a Lagrangian rate–distortion objective,

$$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$

or, in the alternate review formulation,

$\min_{E_\phi,G_\theta} R+\lambda\cdot D,$

with an optional perception term $P$ in $R+\lambda D+\beta P$ (Chen et al., 9 Jun 2025, Chen et al., 2023).

This abstraction generalizes early Model-Based Coding by replacing analytic synthesis with deep generative networks. The encoder no longer transmits dense block motion fields and residual transforms as in hybrid coding; instead it transmits sparse or compact face representations such as 2D landmarks, 3D landmarks, semantic maps, temporal-evolution features, or latent codes, and relies on a learned synthesizer for dense motion recovery and image generation (Chen et al., 2024, Chen et al., 2024). A plausible implication is that GFVC shifts the principal coding burden from pixel-domain residual transmission to representation design and decoder prior quality.

2. Representation families and decoder architectures

The surveys organize GFVC methods primarily by representation type. One major class uses 2D keypoints. In this setting, the encoder transmits a small set of facial keypoints, optionally with local affine motions or Jacobians, and the decoder predicts dense flow and occlusion before synthesizing the frame through a U-Net- or SPADE-style generator (Chen et al., 9 Jun 2025, Chen et al., 2023). In "Neural Face Video Compression using Multiple Views" (Volokitin et al., 2022), the single-view baseline follows the First Order Motion Model (FOMM): a keypoint detector $D$ outputs sparse 2D keypoints, a dense motion predictor $\theta_1^t$0 refines a coarse flow into a dense field $\theta_1^t$1 and predicts an occlusion map $\theta_1^t$2, an encoder $\theta_1^t$3 extracts source-view features, warp-and-occlusion $\theta_1^t$4 bilinearly samples the feature map, and a ResNet-based generator $\theta_1^t$5 produces the reconstruction. The coarse flow is

$\theta_1^t$6

with refinement

$\theta_1^t$7

and warped features

$\theta_1^t$8

In the multi-view extension, $\theta_1^t$9 source views are stored, each view is warped independently, and the warped features are fused by a permutation-invariant aggregation module $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$0 before decoding (Volokitin et al., 2022).

A second class uses 3D semantic or morphable-model representations. "Interactive Face Video Coding: A Generative Compression Framework" (Chen et al., 2023) represents each inter frame by a compact set of 3D-face semantic parameters after an Internal Dimension Increase (IDI) stage. The regressed latent $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$1 contains appearance and motion terms, while the transmitted compact semantics are $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$2. A 3D mesh reconstruction module produces a dense 2D facial mesh and eye-blink map, a mesh-based motion estimation network generates coarse and fine flow, and a CSSFT-GAN generator synthesizes the output frame (Chen et al., 2023). This line of work emphasizes editable semantics and direct control over mouth motion, eye blinking, head rotation, and head translation.

A third class uses learned compact features or latent spaces. The reviews list CFTE as a representative compact latent-matrix method and note that learned feature vectors or temporal-evolution features can replace explicit landmarks (Chen et al., 9 Jun 2025, Chen et al., 2023). "Video Coding Using Learned Latent GAN Compression" (Shukor et al., 2022) represents frames by latent codes in the StyleGAN2 $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$3 space and learns an invertible proxy space $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$4 through a RealNVP normalizing flow,

$\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$5

The encoder inverts each frame into $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$6, maps it to $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$7, quantizes and entropy-encodes $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$8, and reconstructs by $\hat X_0=\mathrm{Dec}(\mathrm{Enc}(X_0)),$9 using fixed StyleGAN2 encoder and generator components (Shukor et al., 2022). This establishes a distinct GFVC branch in which synthesis quality derives from a pretrained GAN prior rather than explicit motion warping alone.

Across these representation families, decoder design recurrently combines dense motion estimation, warping, occlusion handling, and a conditional generator. Depending on the method, realism is enforced by perceptual losses, adversarial losses, feature matching, equivariance constraints, or latent-space reconstruction objectives (Volokitin et al., 2022, Shukor et al., 2022, Chen et al., 2023).

3. Coding objectives, losses, and rate–distortion behavior

GFVC training is consistently expressed as rate–distortion optimization, but the distortion term is often perceptual rather than pixel-fidelity oriented. Reviews summarize this as

$$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$0

while standardization-oriented descriptions write

$$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$1

with $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$2 including both base-picture bits and generative side information (Chen et al., 2024, Chen et al., 2023). Typical distortion terms include $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$3 or $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$4 reconstruction, VGG-based perceptual loss, adversarial loss, and task-specific regularizers such as equivariance or identity preservation (Chen et al., 2024, Chen et al., 2023).

In the multi-view FOMM extension, the paper uses a multi-scale perceptual loss on VGG features and optionally a pixel-wise $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$5 or $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$6 term; no adversarial loss is used in that work (Volokitin et al., 2022). The perceptual term is

$$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$7

Its per-frame rate is explicitly derived from keypoint transmission: with $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$8 2D keypoints encoded as 16-bit floats, the inter-frame cost is $$\theta_1^t=\mathcal{E}(X_1^t), \quad R\approx \text{bits}(\theta_1^t),$$9 bits/frame (Volokitin et al., 2022).

By contrast, SGANC defines both rate and distortion in latent space. For intra coding,

$$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$0

with $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$1, and the distortion term is mean-squared error in $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$2,

$$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$3

For inter coding, successive latent differences are quantized, optional residuals are inserted every $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$4 frames, and the loss is

$$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$5

The paper justifies latent-space distortion by arguing that $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$6 encodes multi-scale, semantically disentangled face attributes and reports that latent-space MSE yields a better rate–distortion trade-off than image-space MSE or LPIPS within that framework (Shukor et al., 2022).

The empirical rate–distortion literature consistently notes that GFVC targets perceptual quality more than classical PSNR optimality. The 2025 survey reports average perceptual BD-rate savings at $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$7 of $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$8 for FOMM, $$\hat X_1^t=\zeta\!\Bigl(\hat X_0,\;\varpi\!\bigl(\hat\theta_1^t,\mathcal{E}(\hat X_0)\bigr)\Bigr)\equiv G(\hat\theta_1^t;\hat X_0).$$9 for CFTE, $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$0 for FV2V, $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$1 for DAC, and $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$2 for TPS relative to VVC in Rate-DISTS; corresponding Rate-LPIPS savings are $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$3, $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$4, $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$5, $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$6, and $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$7 (Chen et al., 9 Jun 2025). The same survey also states that GFVC lags in PSNR/SSIM and that this reflects the semantic-perceptual target rather than a failure of the paradigm (Chen et al., 9 Jun 2025). This addresses a recurrent misconception: lower pixel-domain fidelity does not preclude superior perceptual quality under GFVC’s design objective.

4. Multi-reference, layered, and bandwidth-intelligent extensions

Several works extend baseline GFVC to address instability, narrow bitrate range, or limited bandwidth adaptation. One line of development introduces multiple references. The multi-view extension of FOMM stores $$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$8 source views, computes a dense flow and warped feature map for each, and fuses them with either pooling or cross-view self-attention (Volokitin et al., 2022). The self-attention formulation computes

$$\mathcal{L}_{\rm tot}=\sum_t D(X_1^t,\hat X_1^t)+\lambda R(\theta_1^t),$$9

followed by

$\min_{E_\phi,G_\theta} R+\lambda\cdot D,$0

and weighted merging

$\min_{E_\phi,G_\theta} R+\lambda\cdot D,$1

On VoxCeleb2, the single-view baseline reported $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$2, $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$3 dB, $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$4, and $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$5, whereas the 3-view model MAX-RS-3 reported $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$6, $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$7 dB, $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$8, and $\min_{E_\phi,G_\theta} R+\lambda\cdot D,$9 (Volokitin et al., 2022).

A more explicit multi-reference animation framework is MRDAC, which augments a conventional HEVC or VVC pipeline by carrying model side information in SEI, maintaining a small buffer of decoded references, predicting dense flow and occlusion from each reference in parallel, and aggregating warped deep features through weighted max-pooling (Konuko et al., 2024). Its per-reference warping is

$P$0

and aggregation is

$P$1

MRDAC additionally introduces a contrastive loss on warped representations from different references for the same target frame, with cosine similarity and temperature $P$2, to align multi-reference features and reduce reconstruction drift (Konuko et al., 2024). The paper states that, in standard long-GOP open-loop tests, MRDAC improves LPIPS by $P$3 across $P$4 kbps, improves MS-SSIM by $P$5 over DAC/MVAC, improves VMAF by $P$6 points at $P$7 kbps, and yields approximate BD-rate gains of $P$8 versus DAC and $P$9 versus MVAC; bi-directional mode provides an additional $R+\lambda D+\beta P$0 MS-SSIM gain at the cost of $R+\lambda D+\beta P$1 s added latency (Konuko et al., 2024).

A second extension introduces layered or progressive coding to widen bitrate coverage. "A Hybrid Deep Animation Codec for Low-bitrate Video Conferencing" (Konuko et al., 2022) combines a deep animation branch with a low-bitrate HEVC base-layer video and a learned fusion module. Keypoints and Jacobians are transmitted for animation, while HEVC P-frames at $R+\lambda D+\beta P$2 provide a $R+\lambda D+\beta P$3 kbps auxiliary stream for $R+\lambda D+\beta P$4 fps video. Over VoxCeleb2 and Xiph.org sequences, the method reports BD-Rate gains over HEVC of $R+\lambda D+\beta P$5 and $R+\lambda D+\beta P$6 in PSNR, $R+\lambda D+\beta P$7 and $R+\lambda D+\beta P$8 in SSIM, and $R+\lambda D+\beta P$9 and $D$0 in msVGG, and the authors state that H-DAC extends DAC’s operational range up to $D$1 kbps (Konuko et al., 2022).

The notion of bandwidth intelligence is developed further by PGen and PFVC. PGen adds optional enhancement layers of auxiliary facial features at progressively coarser spatial resolutions such as $D$2, $D$3, and $D$4, yielding a layered bitstream of base GFVC latents plus scalable feature maps (Chen et al., 24 Feb 2025). The enhancement path includes a U-Net + GDN feature descriptor, a learned hyperprior and autoregressive entropy model, attention-guided signal enhancement via Hadamard product, and coarse-to-fine GAN-based generation with SPADE and multi-scale warping (Chen et al., 24 Feb 2025). The paper reports that PGen extends GFVC’s useful perceptual rate range from $D$5 kbps up to $D$6 kbps and that PGen+CFTE achieves average BD-rate improvement against VVC of DISTS $D$7, FVD $D$8, and MANIQA $D$9 (Chen et al., 24 Feb 2025).

PFVC replaces fixed-dimensional inter-frame codes with adaptive visual tokens at four granularities: $\theta_1^t$00 The encoder selects the token prefix up to granularity level $\theta_1^t$01, and the decoder reconstructs motion features, dense flow, occlusion mask, and final frame through a GAN-based generator (Chen et al., 2024). The paper states that PFVC strictly outperforms VVC, CFTE, and FOMM in DISTS, LPIPS, FVD, and MANIQA across the entire bitrate range, matches or slightly outperforms VVC in PSNR/SSIM below $\theta_1^t$02 kbps, and covers from ultra-low $\theta_1^t$03 kbps up to medium $\theta_1^t$04 kbps seamlessly (Chen et al., 2024). This suggests that hierarchical inter-frame representations address one of the central constraints identified for early GFVC systems: coarse rate control tied mainly to key-frame QP.

5. Standardization, interoperability, and systems aspects

GFVC has moved from isolated research codecs toward standardization through SEI carriage in standard hybrid bitstreams. The SEI-based framework proposed in 2024 defines a GFV SEI message for compact facial parameter streams such as 2D/3D keypoints, facial semantics, or compact features, inserted into a VVC bitstream and adopted as a technology under consideration in the JVET work item for Versatile Supplemental Enhancement Information, to be standardized as a new version of "ITU-T H.274 | ISO/IEC 23002-7" (Chen et al., 2024). The message syntax includes flags for coordinate and matrix parameters, optional prediction, a precision factor, and Uniform Resource Identifiers for TranslatorNN and GenerativeNN: $\theta_1^t$46 The design is backward-compatible: decoders unaware of the SEI ignore it and display the hybrid-coded base pictures, while SEI-aware decoders invoke GFV post-processing (Chen et al., 2024).

The surveys describe this as a high-level syntax for transporting facial parameters $\theta_1^t$05 and mapping codes, enabling interoperable signaling of representation type, latent dimension or point count, temporal indexing, quantization parameters, and translator flags (Chen et al., 9 Jun 2025, Chen et al., 2023). In MRDAC, GFVC side information rides in the SEI payload while the conventional HEVC/VVC bitstream remains unchanged, and the SEI payload typically remains $\theta_1^t$06 bytes/frame (Konuko et al., 2024). This standardization path is closely related to application-layer flexibility: user-specified animation or filtering, metaverse avatars, progressive enhancement, translator-network signaling, and model upgrades by URI changes are all explicitly identified as enabled functionalities (Chen et al., 2024).

GFVC research has also increasingly addressed decoder complexity and deployment constraints. The 2025 survey summarizes a low-complexity CFTE implementation based on depthwise separable convolutions and network slimming, reducing parameters from $\theta_1^t$07 M to $\theta_1^t$08 M and MACs/pixel from $\theta_1^t$09 to $\theta_1^t$10 kMAC while keeping Rate-DISTS and Rate-LPIPS virtually unchanged and still superior to VVC at low bitrate (Chen et al., 9 Jun 2025). "A Lightweight Dual-Mode Optimization for Generative Face Video Coding" (Zhang et al., 19 Aug 2025) gives a more detailed complexity-reduction study for a CFTE baseline. It replaces $\theta_1^t$11 convolutions with depthwise separable convolutions, halves bottleneck channels in the Generation Module, and applies two-stage adaptive channel pruning using a MaskedBatchNorm2d with trainable scales $\theta_1^t$12 and a learnable threshold $\theta_1^t$13. The sparsity term is

$\theta_1^t$14

with total objective

$\theta_1^t$15

The resulting model reduces parameters from $\theta_1^t$16 M to $\theta_1^t$17 M and KMACs/pixel from $\theta_1^t$18 to $\theta_1^t$19, corresponding to approximately $\theta_1^t$20 parameter reduction and $\theta_1^t$21 computation saving, while the pruned CFTE-Lite model achieves BD-Rate DISTS $\theta_1^t$22 and BD-Rate LPIPS $\theta_1^t$23 against VVC (Zhang et al., 19 Aug 2025).

A hardware-centric realization is provided by GRACE, which targets FPGA deployment for an animation-based generative codec decoder (Wan et al., 12 Nov 2025). The system applies layer fusion and post-training static quantization, partitions lightweight tasks to the processor side and heavy operators such as Conv2D, grid-sample, upsample, and Hadamard product to programmable logic, and uses double buffering and loop unrolling for acceleration (Wan et al., 12 Nov 2025). On a PYNQ-Z1 platform, the prototype reports $\theta_1^t$24 Frame/J, $\theta_1^t$25 per reconstructed pixel, and energy-efficiency gains of $\theta_1^t$26 over CPU and $\theta_1^t$27 over GPU, with $\theta_1^t$28 FPS at $\theta_1^t$29 and a $\theta_1^t$30 MHz PL clock (Wan et al., 12 Nov 2025). These studies make clear that GFVC’s practical feasibility depends not only on coding theory but also on model compression, quantization, and hardware/software co-design.

6. Evaluation practice, misconceptions, and current research directions

Evaluation in GFVC has shifted decisively toward perceptual and temporal metrics. The 2025 survey reports a large-scale GFVC-compressed face video database with subjective Mean Opinion Scores based on $\theta_1^t$31 test sequences, $\theta_1^t$32 algorithms, and $\theta_1^t$33 QPs, evaluated under a DSCQE protocol with $\theta_1^t$34 participants, yielding final MOS on $\theta_1^t$35 sequences and $\theta_1^t$36 subjects after outlier processing (Chen et al., 9 Jun 2025). The study finds that DISTS, LPIPS, TOPIQ, FVD, and no-reference MANIQA/FAVOR correlate best with MOS, with PLCC/SRCC approximately $\theta_1^t$37, while PSNR and SSIM correlate poorly at approximately $\theta_1^t$38 (Chen et al., 9 Jun 2025). This empirical result reinforces the earlier survey observation that classic pixel-domain fidelity measures are often inadequate for generative face reconstruction (Chen et al., 2023).

A common misconception is that GFVC should be judged primarily by PSNR, in the same way as block-based hybrid codecs. The literature repeatedly qualifies this point. PFVC notes that GAN-based perceptual coding may remain below VVC in PSNR at very high rates even while outperforming in perceptual metrics (Chen et al., 2024). The 2025 survey states that GFVC lags in PSNR/SSIM because it targets semantic-perceptual reconstruction rather than strict pixel reproduction (Chen et al., 9 Jun 2025). Likewise, SGANC reports that at low BPP its PSNR is comparable to H.265/VTM while LPIPS is drastically lower and outputs are artifact-free and photo-realistic; at higher BPP GFVC leads on PSNR and MS-SSIM as well (Shukor et al., 2022).

Another persistent issue is reconstruction instability and drift. PFVC identifies unstable reconstruction, narrow bitrate envelope, and lack of bandwidth adaptation as key limitations of prior GFVC algorithms (Chen et al., 2024). MRDAC frames drift as the growth of reconstruction error with increasing temporal distance from reference frames and argues that multi-reference aggregation and contrastive alignment reduce drift growth rate (Konuko et al., 2024). H-DAC observes that sparse-keypoint animation can drift over long intervals and fail on non-facial motion, while the auxiliary base layer regularizes background changes and disocclusions (Konuko et al., 2022). These critiques are not contradictions of the paradigm; rather, they define the main axes along which later GFVC systems have evolved.

The application space is broader than ultra-low-bitrate face teleconferencing. IFVC emphasizes editable semantic bitstreams and interactive coding without additional manipulation processes (Chen et al., 2023). The SEI standardization paper enumerates user-specified facial animation or filtering, metaverse avatars, progressive enhancement, chroma-key fusion, and model upgrades via URIs (Chen et al., 2024). AVCC extends GFVC to joint audio-video compression. It encodes an identity code $\theta_1^t$39, pose code $\theta_1^t$40, and audio tokens $\theta_1^t$41, and decodes them through an Audio-Visual Diffusion module with cross-attention between audio and visual latents (Xu et al., 17 Dec 2025). The cross-modal diffusion objective denoises both modalities jointly, and the training loss adds perceptual, adversarial, feature-matching, and lip-sync terms,

$\theta_1^t$42

The paper reports that AVCC achieves uniformly better video reconstruction quality than VVC and leading GFVC methods in DISTS, LPIPS, and FID, saves up to $\theta_1^t$43 bitrate versus VVC on selected datasets, and improves audio quality at $\theta_1^t$44 kbps relative to Encodec at $\theta_1^t$45 kbps in MEL, STFT, VIS, and WER (Xu et al., 17 Dec 2025). A plausible implication is that GFVC is beginning to evolve from face-only semantic video coding toward multimodal communication systems.

Current challenges remain explicit in the survey literature: computational demands and edge deployment, interpretability and robustness, quality assessment, generalization beyond celebrity-style training data, and privacy and ethics related to realistic face generation (Chen et al., 9 Jun 2025, Chen et al., 2023). The same sources identify likely future directions: diffusion and transformer generators, temporal models, full-bandwidth coverage, improved GFVC-tailored metrics, hardware/software co-design, and extension beyond faces to full scenes or human bodies (Chen et al., 9 Jun 2025, Chen et al., 2023).