Hierarchical Quantized Autoencoder (HQA)

• HQA is a deep generative framework that employs multi-level quantization to capture both global and fine-grained structures in data.
• It uses vector or scalar quantization across hierarchical latent spaces to facilitate efficient compression and improved reconstruction quality.
• Empirical results show that HQA enhances rate–distortion trade-offs and enables progressive, scalable encoding for diverse application domains.

A Hierarchical Quantized Autoencoder (HQA) is a deep generative modeling framework that compresses data by learning multi-level, discrete latent representations organized in a hierarchy, with each level capturing progressively finer structure. HQAs employ vector quantization (VQ) or scalar quantization at multiple latent layers, yielding code sequences that can be entropy-encoded for compression or used as bottleneck representations for generative modeling. This architecture generalizes VQ-VAE and VQ-VAE-2 to deeper or more structured hierarchies, and has been employed for lossy image and video compression, generative modeling, graph representation learning, and structured text generation.

1. Architectural Principles and Hierarchical Latent Structure

HQAs comprise a sequence of encoder–quantizer–decoder blocks, where the encoder transforms the input into a stack of continuous latent feature maps at decreasing spatial (or spatiotemporal, or node) resolutions. Each latent level is quantized independently (or, in some variants, with explicit inter-level dependencies) by mapping features to discrete codebook entries. The hierarchy can take several forms:

• Coarse-to-fine ladder (VQ-VAE2 style): Each level encodes residual information not captured by coarser layers, with the top level (lowest resolution) storing global content and lower levels refining details (Duan et al., 2022, Adiban et al., 2023, Adiban et al., 2022).
• Residual-quantized hierarchy: At each level, the encoder quantizes the residual between its input and the decoded sum of higher-layer codes; the sum across layers approximates the full latent (Adiban et al., 2022, Adiban et al., 2023).
• Progressive quantization: Hierarchical bins with learned widths, each providing refined quantization within the interval defined by the previous coarser layer, enable scalable, quality-progressive coding (Lee et al., 2024).
• Domain-structured hierarchy: For graphs, discrete codebooks are hierarchically organized as clusters of clusters, supporting node-level and group-level representations (Zeng et al., 17 Apr 2025); for text, a tree or chain of discrete variables emulates script branching (Weber et al., 2018).

Table: Common HQA hierarchy layouts

Variant	Latent organization	Application domains
Ladder/stacked VQ	Multiresolution, parallel	Images, video, audio
Residual hierarchy	Residual at each layer	Images, video, audio
Tree/graph	Parent–child, cluster linkage	Graphs, structured text
Progressive bins	Nested, non-overlapping bins	Progressive compression

2. Quantization Mechanisms and Training Objectives

HQA architectures employ vector (or scalar) quantization at each hierarchy level. Quantization maps encoder outputs to codebook entries using the nearest-neighbor rule:

$k = \arg\min_j \|z_e - e_j\|_2$

where $z_e$ is the encoder output, and $\{e_j\}$ is the set of codebook embeddings. For video and image compression, spatial or spatiotemporal tensors are quantized either channelwise or positionwise.

Training objectives:

• VQ-VAE style: Minimize reconstruction loss plus codebook and commitment losses using the stop-gradient trick:

$$L = \|x - \hat{x}\|^2_2 + \|sg[z_e] - z_q\|^2_2 + \beta\|z_e - sg[z_q]\|^2_2$$

• Stochastic quantization and ELBO variants: Add noise (uniform or Gaussian) for differentiability, and employ a variational ELBO with entropy regularization to encourage codebook usage and mitigate collapse (Duan et al., 2022, Takida et al., 2023, Williams et al., 2020, Willetts et al., 2020).
• Residual or progressive HQA: Each quantization layer models either the residual or a subset/refinement of a previous coarse quantization (Adiban et al., 2022, Lee et al., 2024).
• Specialized domain losses: For graphs, node and edge reconstruction terms; for text, negative log-likelihood and sequence-level evaluation metrics (Zeng et al., 17 Apr 2025, Weber et al., 2018).

Codebook optimization: Either EMA updates (Adiban et al., 2023, Adiban et al., 2022) or direct gradient descent; codebook collapse is mitigated by entropy bonuses or periodic dead-code resets (Reyhanian et al., 29 Jan 2026, Zeng et al., 17 Apr 2025).

3. Entropy Coding, Compression, and Rate–Distortion Tradeoff

HQAs are natively suited for source coding. Quantized latents define a discrete symbol sequence to be entropy-coded:

• Arithmetic/range coding: Each hierarchy level's code is encoded using prior distributions parameterized by coarser level codes (typically logistic or Gaussian CDF models) (Duan et al., 2022).
• GPU parallelization: As each position in a given coarse level can be encoded independently when conditioned on higher-level codes, batched GPU arithmetic coding enables high throughput (Duan et al., 2022).
• Progressive coding: In progressive codecs, only the most significant quantization bins or selected components are transmitted at each layer, supporting scalable rate–distortion tradeoffs and adaptive quality (Lee et al., 2024).

Training is guided by a Lagrangian objective

$\mathcal{L} = R + \lambda D$

where $R$ is the bitrate (as estimated via entropy models), $D$ is distortion (MSE or MS-SSIM), and $\lambda$ controls the tradeoff.

4. Codebook Design, Hierarchy Efficiency, and Collapse Prevention

Hierarchical codebooks enable large representational capacity with tractable lookup cost and address the codebook underutilization ("collapse") endemic to high-dimensional, non-hierarchical VQ-VAEs (Adiban et al., 2022, Zeng et al., 17 Apr 2025, Reyhanian et al., 29 Jan 2026, Takida et al., 2023, Willetts et al., 2020).

• Hierarchical search: Instead of $O(M^n)$ complexity for a flat codebook of size $M^n$, an $z_e$0-level hierarchy with $z_e$1 codes per level reduces lookup to $z_e$2 per position.
• Residual decomposition: Each layer models only information not captured in coarser layers, reducing redundancy and improving code utilization (Adiban et al., 2022, Adiban et al., 2023).
• Soft/annealed selection: Early in training, stochastic or annealed posterior sampling encourages exploration and codebook spread; as annealing proceeds, hard selections dominate, yielding efficient compression (Zeng et al., 17 Apr 2025, Takida et al., 2023).
• Entropy/commitment bonuses: ELBOs with entropy terms (or explicit perplexity tracking) ensure dispersive use of available codes (Takida et al., 2023, Williams et al., 2020).
• Lightweight interventions: Codebook initialization from data, dead-code reset strategies, and careful hyperparameter tuning (favoring $z_e$3 regimes) are crucial to maximizing effective capacity (Reyhanian et al., 29 Jan 2026, Adiban et al., 2022).

5. Applications and Empirical Results

HQAs have been instantiated and empirically validated in diverse domains:

Image and Video Compression

Experiments on standard datasets (Kodak, Tecnick, UCF101, CelebA, ImageNet, FFHQ) demonstrate that HQA and associated hierarchical quantized VAEs:

• Achieve state-of-the-art or near-state-of-the-art PSNR and MS-SSIM at low bitrates, outperforming classic and learned codecs such as BPG and Ballé's hyperprior (Duan et al., 2022, Lee et al., 2024, Kotthapalli et al., 31 Dec 2025).
• Remove common artifacts (blocking, ringing) seen in classic codecs and preserve better perceptual detail (Duan et al., 2022).
• Allow rapid, massively parallel encoding and decoding on GPU (runtime ~30–50 ms for 1024×1024 images) (Duan et al., 2022).
• Support fine-grained rate–distortion control and quality scalability with a single progressively decodable code stream (Lee et al., 2024).
• Two-level hierarchical VQ-VAEs improve PSNR by 0.2–0.5 dB over strong non-hierarchical baselines at similar rates (Kotthapalli et al., 31 Dec 2025).
• In ablations with codebook collapse prevented and matched capacity, single-level VQ-VAEs can recover reconstruction performance close to hierarchical models (within 0.5 dB) (Reyhanian et al., 29 Jan 2026).

Video Prediction and Generative Modeling

HQAs with autoregressive or spatiotemporal priors (e.g., ST-PixelCNN) produce high-fidelity, temporally coherent video or sequence predictions, with hierarchical codes improving both sharpness and diversity over single-level VQ-VAEs at equivalent bitrates (Adiban et al., 2023).

Graph and Text Representation Learning

• HQA-GAE achieves leading performance on self-supervised node embedding, link prediction, and node classification benchmarks, outperforming 16 baselines across eight graphs (Zeng et al., 17 Apr 2025).
• In script generation, hierarchical VQ provides superior perplexity and semantic branching quality over language-modeling and standard VAEs (Weber et al., 2018).

6. Limitations, Open Questions, and Future Directions

Notable findings and open issues in the HQA literature include:

• Hierarchy is not inherently reconstructive: When representational budget is controlled and codebook collapse is prevented, hierarchy does not automatically guarantee superior reconstruction; rather, it stabilizes codebook usage and makes optimization tractable (Reyhanian et al., 29 Jan 2026).
• Perceptual utility: While reconstruction fidelity can be matched by single-level models, hierarchical structure may still benefit perceptual downstream tasks (e.g., autoregressive priors) due to its inductive bias for multi-scale feature separation.
• Generalization to new modalities: HQA frameworks have been successfully ported to video, audio, graphs, and text with suitable domain adaptations (Takida et al., 2023, Kotthapalli et al., 31 Dec 2025, Zeng et al., 17 Apr 2025, Weber et al., 2018).
• Entropy coding and adaptive quantization: Many HQA variants do not model entropy explicitly for their discrete codes; integrating learned entropy models or adaptive quantization could further optimize the rate–distortion tradeoff (Kotthapalli et al., 31 Dec 2025, Lee et al., 2024).
• Scalability and parameter efficiency: Depthwise quantization and hierarchical clustering can exponentially increase representational capacity with only linear codebook growth, offering efficient scaling for high-dimensional or high-resolution data (Fostiropoulos et al., 2022, Adiban et al., 2022).
• Progressive coding and real-time throughput: Hierarchies allow efficient progressive transmission and adaptive rate control, with learned quantization intervals and masking enabling near-instantaneous decoding and high throughput (Lee et al., 2024).

7. Comparative Summary of Key Results

Paper / Variant	Application	Notable Achievements	Reference
Quantized Hierarchical VAE (3-layer)	Image compression	+0.8 dB PSNR vs. BPG, GPU parallel entropy encoding	(Duan et al., 2022)
S-HR-VQVAE	Video prediction	PSNR↑, SSIM↑ at lower bpp, robustness to code collapse	(Adiban et al., 2023)
HQA-GAE (2-layer, graph)	Graph learning	AP, classification ↑20%/2pts vs. baselines	(Zeng et al., 17 Apr 2025)
HQ-VAE (variational)	Image/audio reconstruction	Codebook perplexity↑, SOTA FID (ImageNet)	(Takida et al., 2023)
DeepHQ (progressive HQA)	Progressive compression	SOTA rate, real-time decode, 8-level scalability	(Lee et al., 2024)
Capacity-matched VQ-VAE	Comparative study	Single-level matches hierarchical for MSE/PSNR	(Reyhanian et al., 29 Jan 2026)

Hierarchical Quantized Autoencoders thus unify multi-level vector/scalar quantization, residual learning, and variational inference into a powerful framework for high-fidelity, scalable, and interpretable discrete representation learning across vision, graph, and sequential domains. Their practical impact depends on optimized codebook design, entropy modeling, and architecture–domain alignment.