Historical-Prior Generative Compression

• Historical-Prior Generative Compression is a generative coding paradigm that transmits minimal side information while using powerful learned priors for conditional signal reconstruction.
• It leverages various historical data sources—such as offline dataset priors, causal latent context, and past reconstructions—to reduce bitrate effectively.
• Empirical evaluations in image and video systems demonstrate significant bitrate reductions and improved perceptual quality compared to traditional codecs.

Searching arXiv for papers on historical-prior and generative compression to ground the article with current literature. Historical-Prior Generative Compression (HPGC) denotes a generative coding paradigm in which the sender transmits only minimal side information—compact latent codes, sparse keyframes, seeds, lightweight control tokens, or other low-entropy surrogates—while the receiver reconstructs the signal by invoking a powerful generative model conditioned on historically available information. In the broad synthesis used in generative visual compression, this “historical” information includes dataset priors learned offline, hyperpriors and autoregressive context in latent space, and temporal priors derived from previously decoded content; in later video work, the term is used more specifically for decoder-side synthesis conditioned on past reconstructions and learned video dynamics (Chen et al., 2024, Chen et al., 30 Dec 2025). The central premise is that transmitted bits need only specify what is novel relative to the prior, so reconstruction becomes a conditional generation problem rather than a residual-prediction problem.

1. Conceptual and historical foundations

The modern lineage begins with generative compression as a general idea: transmit short latent codes rather than pixels, and reconstruct through a learned generator. An early formulation encoded an image to a fixed-length latent vector, quantized it, and decoded through a pretrained GAN; for video, it transmitted fewer latent codes and reconstructed intermediate frames by latent interpolation (Santurkar et al., 2017). This baseline did not yet use an explicit learned temporal prior, but it established the core asymmetry of generative codecs: low-rate transmission at the encoder, powerful synthesis at the decoder.

Subsequent learned image compression made the prior explicit. In transform coding with quantized latents $\hat{y}$ and hyperlatents $\hat{z}$, bitrate is governed by the entropy model rather than solely by architectural choices. The joint autoregressive and hierarchical formulation

$$p(\hat{y}\mid \hat{z}, \text{context})=\prod_i \bigl(\mathcal{N}(\mu_i,\sigma_i^2)*\mathcal{U}(-1/2,1/2)\bigr)(\hat{y}_i)$$

showed that causal latent history and hyperprior side information are complementary sources of prior knowledge, yielding a 15.8% average reduction in file size over the previous learned state of the art, a 59.8% rate reduction relative to JPEG, more than 35% reduction relative to WebP and JPEG2000, and bitstreams 8.4% smaller than BPG (Minnen et al., 2018). In this sense, “historical prior” first appears as causal dependence on already decoded latent symbols.

The unifying objective across later systems is the rate–distortion Lagrangian

$L=R+\lambda D,$

with $R$ the expected code length and $D$ a distortion, perceptual, or task-aware term. In VAE-based formulations, the ELBO

$$\mathcal{L}_{ELBO}=\mathbb{E}_{q_\phi(z\mid x)}[\log p_\theta(x\mid z)]-\mathrm{KL}(q_\phi(z\mid x)\|p_\theta(z))$$

connects reconstruction to distortion and the KL term to rate under a learned prior (Chen et al., 2024). A later diffusion-based review generalized the notion further: historical priors are any decoder-side signals correlated with the source, including past reconstructions, preview images, side information, and common randomness, all of which modify the effective coding problem from unconditional reconstruction to conditional generation (Yang et al., 26 Jan 2026).

2. Priors in latent-space image compression

In image compression, historical priors most commonly appear as latent-space conditioning. The review literature groups scale hyperpriors, autoregressive context, Gaussian mixture likelihoods, channel-wise models, and space–channel contextual models under a common principle: latent entropy is reduced by conditioning on learned side information or decoded neighborhood history, so that

$H(y)\approx -\mathbb{E}[\log p(y\mid h)],$

where $h$ may denote hyper-latents or causal context (Chen et al., 2024). This is the image-domain analogue of historical conditioning: the decoder reconstructs not from $\hat{y}$ alone, but from $\hat{z}$0 interpreted through a learned, image-adaptive prior.

Several recent codecs deepen that conditioning. HVQ-CGIC introduces a hyperprior over VQ indices by predicting, at each spatial location, a Gaussian density in codebook embedding space and converting it to a categorical distribution through a Mahalanobis-distance Softmax. This exposes explicit and differentiable rate terms for VQ indices and hyper-latents, enabling controllable rate through both loss weights and routing masks; on Kodak, it reports the same LPIPS as Control-GIC, CDC, and HiFiC with an average of 61.3% fewer bits (yi et al., 8 Dec 2025). The same paper explicitly states that this is a hyperprior method and should not be confused with history-based priors used in predictive video coding.

Other image codecs operationalize “historical” in a broader, causal-within-image sense. DCIC-sgp first encodes a self-generated structural prior from the image itself, decodes that prior, and then uses it to condition the analysis transform, the entropy model, and the synthesis transform. The decoded prior thus functions as an encoded-first structural backbone rather than as a conventional hyperprior derived from the main latent; this design reports BD-rate reductions of 14.4%, 15.7%, and 15.1% against VTM-12.1 on Kodak, CLIC, and Tecnick (Zhao et al., 28 Oct 2025). AFP-GIC transfers an adaptive fused prior from a frozen pretrained AdaCode model, predicts a compatible fused prior at the decoder without transmitting it, and reduces decoder latency by 18.1% and total parameter count by 31.10 million relative to DC-VIC (Pei et al., 16 May 2026).

A separate branch uses large pretrained generative models as static decoder-side priors learned from historical corpora. FlowCodec plugs pretrained text-to-image priors such as Qwen-image-2512 and FLUX.1-dev into an ultra-low-bitrate codec through one-step latent transport, requiring neither auxiliary conditioning signals nor large retraining; it reports high visual quality below 0.05 bits per pixel and keeps trainable parameters below 0.54% of the generative backbone (Huang et al., 19 Jun 2026). This suggests that, in image compression, “historical prior” has at least three non-equivalent meanings: causal latent context, encoded-first structural side information, and frozen pretrained priors acquired offline.

3. Temporal historical priors in video codecs

Video is the setting in which HPGC becomes most literal. The review taxonomy includes rate–distortion autoencoders, sequential VAEs, GAN-based neural video compression, and domain-specific generative pipelines in which the current frame is synthesized from compact symbols conditioned on decoded history (Chen et al., 2024). In probabilistic form, the decoder models

$\hat{z}$1

where $\hat{z}$2 is built from past reconstructions, motion/structure codes, or other historical state. This differs from classical predictive codecs, which transmit motion-compensated residuals in pixel space; HPGC instead transmits compact semantic or dynamical surrogates and delegates texture synthesis to the decoder.

Recent systems realize this principle in distinct ways. The 2025 GVC framework defines HPGC explicitly as transmitting minimal side information while the receiver reconstructs video by invoking generative priors conditioned on temporal history. Its receiver uses a diffusion-based generative video model over GOPs, with compact tokens including compressed keyframes, high-level segment descriptors, low-level continuous features, and residual-coded streams. On MCL-JCV, it reports an average LPIPS of 0.214 at 0.008 bpp versus 0.271 for HEVC, and on DAVIS2017 VOS at 0.01 bpp it attains $\hat{z}$3, $\hat{z}$4, $\hat{z}$5, and $\hat{z}$6-Recall $\hat{z}$7, substantially above HEVC at the same bitrate (Chen et al., 30 Dec 2025).

ZeroGVC moves further toward training-free historical conditioning. It sends I-frames with OneDC, reconstructs subsequent P-frames through Codebook-Guided Autoregressive Latent Compression, and conditions a pretrained autoregressive diffusion prior on the KV-cache built from previously reconstructed latents. It also introduces an optional bidirectional reference mode that inserts the next I-frame latent into attention context without additional bitrate overhead, mitigating error propagation (Gao et al., 21 Jun 2026). A related but distinct zero-shot design, “Generation Is Compression,” converts a pretrained rectified-flow video model into the codec itself: the transmitted bitstream specifies the stochastic generative trajectory through reproducible codebook atoms, with I2V, T2V, and FLF2V variants spanning trade-offs between anchoring and bitrate. It reports high-quality reconstruction below 0.002 bpp and describes FLF2V boundary sharing as a historical-prior mechanism across GOPs (Zeng et al., 27 Mar 2026).

Sequence-level video priors also address temporal instability that frame-wise image priors cannot. GNVC-VD couples a spatio-temporal latent codec with a video-native diffusion transformer and sequence-level flow-matching refinement initialized from decoded latents rather than pure noise. It reports BD-Rate reductions relative to VVC of 89.4% in LPIPS and 94.5% in DISTS on HEVC-B, and improves temporal metrics over GLC-Video from warp error 86.5 to 66.6 and CLIP-F from 0.979 to 0.982 (Mao et al., 4 Dec 2025). By contrast, “Compressing Scene Dynamics” learns motion-pattern priors from common small scene dynamics rather than from content classes such as faces; it transmits 40 scalar motion-token parameters per inter frame before inter-prediction and CABAC, operates at 5–15 kbps, and reports BD-rate savings versus VVC of −35.55% for Rate-DISTS, −23.98% for Rate-LPIPS, and −29.91% for Rate-FVD (Yin et al., 2024).

4. Domain-specific and cross-domain realizations

One major application regime is ultra-low-bitrate semantic video. The review documents face- and body-centric pipelines based on FOMM, SPADE, landmarks, 3D semantics, compact temporal features, and PCA motion bases, all of which replace pixel-level residuals with structured condition signals such as keypoints, segmentation maps, or learned trajectories (Chen et al., 2024). These systems preserve identity and pose semantics by synthesizing frames from decoder-side priors rather than by reconstructing exact textures, and they support enhancements such as frame interpolation, residual-enhanced coding, multi-view aggregation, and multi-reference prediction.

A second regime is controllable or layered communication. Conceptual coding decomposes structure and texture so that reconstruction can proceed in stages—structure first, texture later—or can support user-specified filtering. The same review also treats machine-centric communication as a generative compression problem: layered or scalable designs deliver a base semantic representation for analysis, with optional enhancement for human viewing (Chen et al., 2024). This suggests that HPGC is not restricted to perceptual restoration; it also supports rate–accuracy optimization in which the decoder is a task-conditioned inference engine.

The most pronounced cross-domain extension is Earth observation. D2AR treats historical EO archives as an executable prior $\hat{z}$8 composed of geographic and temporal embeddings together with a large generative model trained on global archives. Only ultra-low-rate control tokens $\hat{z}$9 and minimal side information are stored or transmitted; reconstruction on the ground conditions on both $$p(\hat{y}\mid \hat{z}, \text{context})=\prod_i \bigl(\mathcal{N}(\mu_i,\sigma_i^2)*\mathcal{U}(-1/2,1/2)\bigr)(\hat{y}_i)$$0 and $$p(\hat{y}\mid \hat{z}, \text{context})=\prod_i \bigl(\mathcal{N}(\mu_i,\sigma_i^2)*\mathcal{U}(-1/2,1/2)\bigr)(\hat{y}_i)$$1 through a Diffusion Transformer operating in EQ-VAE latent space. The framework reports 100× to 10,000× data reduction, sustains 1.54 EFLOP/s and peaks at 2.16 EFLOP/s during training, and improves reconstruction quality at extreme compression ratios such as 6104×, 6295×, and 17,777× while largely preserving downstream DynamicEarthNet classification metrics (Zhang et al., 9 May 2026). Here the historical prior is neither local context nor temporal history within a clip, but a global archive prior learned from repeated measurements of the same evolving planet.

The review literature also places point clouds, light fields, 360° content, and stereo reconstruction in the same conceptual family, insofar as generative priors replace part of the transmitted signal with decoder-side synthesis (Chen et al., 2024). A plausible implication is that HPGC is best understood not as a modality-specific codec class, but as a family of conditional generative communication systems whose common operation is to externalize reusable information into a prior and to transmit only the unpredictable remainder.

5. Evaluation, robustness, and interpretive boundaries

Evaluation practice in HPGC departs from classical codec benchmarking because pixel fidelity and generative plausibility often diverge. The main review emphasizes that PSNR and SSIM do not align well with generative reconstructions trained in feature space, and identifies DISTS and LPIPS as more appropriate perceptual metrics, while also calling for temporal perceptual assessment tailored to generative video (Chen et al., 2024). Later systems accordingly combine perceptual metrics with downstream-task scores, as in GVC’s use of LPIPS and DAVIS2017 VOS rather than pixel distortion alone (Chen et al., 30 Dec 2025).

Robustness has multiple meanings in this literature. The 2017 generative compression work demonstrated unusually strong bit-error resilience because image latents were fixed-length, uniformly quantized vectors rather than variable-length entropy codes: even at bit error rates $$p(\hat{y}\mid \hat{z}, \text{context})=\prod_i \bigl(\mathcal{N}(\mu_i,\sigma_i^2)*\mathcal{U}(-1/2,1/2)\bigr)(\hat{y}_i)$$2, PSNR degraded by only approximately 1 dB, whereas JPEG degraded by more than 7 dB at $$p(\hat{y}\mid \hat{z}, \text{context})=\prod_i \bigl(\mathcal{N}(\mu_i,\sigma_i^2)*\mathcal{U}(-1/2,1/2)\bigr)(\hat{y}_i)$$3 due to catastrophic desynchronization (Santurkar et al., 2017). More recent HPGC systems shift the dominant failure modes away from channel noise and toward semantic drift, conditioning sensitivity, temporal instability, and domain shift.

The term “historical prior” itself is interpretively unstable. In video systems such as GVC, ZeroGVC, GNVC-VD, and motion-prior codecs, it usually denotes temporal conditioning on previously decoded content or on structured dynamics (Chen et al., 30 Dec 2025). In image systems such as FlowCodec and AFP-GIC, by contrast, it refers to pretrained prior spaces learned offline from large corpora rather than to stream-internal history (Huang et al., 19 Jun 2026, Pei et al., 16 May 2026). DCIC-sgp uses the term in yet another sense: the prior is “historical” only because it is encoded and decoded first, then reused causally within the same image (Zhao et al., 28 Oct 2025). HVQ-CGIC sharpens the distinction by explicitly warning that its hyperprior over VQ indices is not a history-based predictive-video prior (yi et al., 8 Dec 2025). This plurality is not merely terminological; it marks distinct causal structures in the codec.

6. Open problems and research directions

A persistent technical tension is the fidelity–realism trade-off. Adversarial and perceptual objectives improve visual plausibility at very low rates, but they can hurt pixel fidelity; the review identifies the choice of $$p(\hat{y}\mid \hat{z}, \text{context})=\prod_i \bigl(\mathcal{N}(\mu_i,\sigma_i^2)*\mathcal{U}(-1/2,1/2)\bigr)(\hat{y}_i)$$4 and loss composition as a central control variable, and also notes rate–distortion–complexity optimization as an emerging requirement for flexible decoding (Chen et al., 2024). Diffusion-based reviews add a complementary theoretical perspective: when realism is treated explicitly, side information and common randomness can shift the attainable rate–distortion–perception frontier, but efficient channel simulation and reliable perceptual objectives remain open problems (Yang et al., 26 Jan 2026).

Generalization is equally unresolved. Many of the strongest systems are scenario-specific—faces, bodies, small oscillatory scenes, or geographically well-covered EO regions—and degrade when prior coverage is sparse or when content falls outside the training regime. Earth-observation experiments quantify this dependence directly: increasing prior coverage from 900 to 1000 cities improves PSNR from 12.1516 to 12.4708 and LPIPS from 0.3273 to 0.2990 on held-out cities, confirming sensitivity to prior completeness (Zhang et al., 9 May 2026). In video, ZeroGVC identifies tiling, runtime, exposure bias, and abrupt scene changes as remaining constraints even with pretrained autoregressive diffusion priors (Gao et al., 21 Jun 2026).

A recurrent direction is hybridization. The review highlights stronger temporal priors and hybrid predictive–generative codecs, including VVC-assisted key frames with generative inter-frame synthesis, as a natural path forward (Chen et al., 2024). Several current systems already instantiate parts of this program: GVC trades bitrate against practicality by sending richer tokens when latency is constrained (Chen et al., 30 Dec 2025); ZeroGVC adds future I-frame context at zero bitrate cost (Gao et al., 21 Jun 2026); GNVC-VD unifies spatio-temporal latent compression with sequence-level generative refinement (Mao et al., 4 Dec 2025). Standardization and deployment are also active themes, with JVET AHG efforts on generative face video coding, interoperability studies, and hardware–software co-design identified as necessary for broader adoption (Chen et al., 2024).

Taken together, these developments indicate that HPGC is converging toward a general principle: compress by externalizing reusable information into learned priors, whether those priors are causal latent context, decoded temporal history, frozen generative backbones, or archive-scale spatiotemporal memory. What remains unsettled is not the utility of the principle, but the optimal form of the prior, the guarantees that can be attached to prior-guided reconstructions, and the balance between controllability, realism, semantic faithfulness, and computational cost.