Hierarchical Discrete Diffusion Model

• Hierarchical discrete diffusion models are generative frameworks that use multi-level abstraction to progressively denoise discrete variables, enabling structured data synthesis.
• They employ diverse hierarchy mechanisms—such as clean-to-group transitions, multi-level residual tokenization, and coarse-to-fine graph factorization—to capture semantic and structural granularity.
• Applications span image generation, speech enhancement, language modeling, and molecular graph synthesis, offering improved fidelity, efficiency, and control.

Hierarchical discrete diffusion models are denoising-based generative models in which the underlying variables are discrete and the generative process is explicitly organized across levels of abstraction, scales, or structured partitions. In the literature represented here, hierarchy appears in several forms: clean-to-group-to-mask state spaces, multi-level residual vector quantization token streams, coarse-to-fine graph constructions, semantic vocabularies, and discrete-time latent trajectories across scales. Representative instances include Hierarchical Koopman Diffusion for images, HiCoDiT for video-to-speech, ADDSE/RQDiT for speech enhancement, MolHIT for molecular graphs, HDLM for language modeling, and hierarchical discrete flow matching for graphs (Bai et al., 14 Oct 2025, Ye et al., 17 Apr 2026, Gonzalez, 25 Feb 2026, Jung et al., 19 Feb 2026, Zhou et al., 8 Oct 2025, Boget et al., 31 Mar 2026).

1. Definition and conceptual scope

In the narrowest sense, a hierarchical discrete diffusion model extends flat discrete diffusion by replacing a single categorical state space with multiple semantic or structural levels. MolHIT makes this explicit by enlarging the token space from clean states to mid-level states and then to a mask state, so that the forward process becomes a two-stage corruption chain $\mathcal{S}_0 \to \mathcal{S}_1 \to \mathcal{S}_2$ rather than a direct jump from clean token to mask (Jung et al., 19 Feb 2026). HDLM adopts the same principle for language: words map surjectively to cluster tokens and then to a mask token, so the forward process performs semantic abstraction and the reverse process performs “next semantic scale prediction” (Zhou et al., 8 Oct 2025).

A second usage places hierarchy in the discrete representation rather than only in the transition kernel. HiCoDiT models 12 RVQ codec levels, with low-level tokens encoding coarse speaker-aware semantics and high-level tokens capturing fine-grained prosody, while ADDSE models a $L \times D$ grid of RVQ codes with an architecture that separates temporal and depth structure (Ye et al., 17 Apr 2026, Gonzalez, 25 Feb 2026). In these systems, diffusion remains discrete, but the hierarchy is inherited from the codec itself.

A third usage couples hierarchy to multiscale structure in the generated object. Hierarchical discrete flow matching for graph generation defines a sequence of coarse graphs and fine graphs, with generation factorized across levels as $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$, where each $\mathcal{H}^\ell$ is a spanning supergraph induced by the next coarser level (Boget et al., 31 Mar 2026). Hierarchical Koopman Diffusion is presented as exactly an instance of what can be called a hierarchical discrete diffusion model because it models discrete-time diffusion evolution in a Koopman-lifted latent space and does so hierarchically across spatial scales (Bai et al., 14 Oct 2025).

A separate line of work locates hierarchy entirely in the forward process. “Improving Discrete Diffusion Models via Structured Preferential Generation” introduces a structured masking process in which token categories are destroyed in an information-based order, so the reverse model preferentially regenerates some categories before others (Rissanen et al., 2024). This suggests that hierarchy in discrete diffusion need not require a multilevel backbone; it can be induced by the corruption schedule alone.

2. Mathematical constructions

The most common starting point is absorbing or masking diffusion on discrete variables. HiCoDiT uses the SEDD framework, where each token dimension follows a continuous-time discrete Markov chain with a special absorbing token $[MASK]$, and the reverse model learns concrete probability ratios via Denoising Score Entropy (Ye et al., 17 Apr 2026). ADDSE uses absorbing discrete diffusion with token space $\{1,\dots,K+1\}$, where the $(K+1)$-th symbol is a mask token $M$, and the reverse process becomes repeated masked-token infilling trained by denoising cross-entropy (Gonzalez, 25 Feb 2026). Prism operates on the same masked discrete diffusion language-model setting, with per-token transition $$q_{t|0}(z_{t,i}\mid z_{0,i})=\mathrm{Cat}(z_{t,i}; \alpha_t \mathbf{e}(z_{0,i}) + (1-\alpha_t)\mathbf{e}_m)$$ (Bai et al., 2 Feb 2026).

MolHIT generalizes this flat setup by defining a forward kernel

$$Q_t = \alpha_t I + (\beta_t - \alpha_t) Q^{(1)} + (1-\beta_t) Q^{(2)},$$

where $L \times D$0 projects clean states to group states and $L \times D$1 projects to the mask state (Jung et al., 19 Feb 2026). The semigroup-consistent family $L \times D$2 preserves closed-form marginals and posterior expressions, so the model remains a valid discrete diffusion process while traversing intermediate chemically meaningful abstractions.

HDLM uses a continuous-time Markov chain over a hierarchical vocabulary. For a token $L \times D$3, the forward marginal is

$L \times D$4

with $L \times D$5 (Zhou et al., 8 Oct 2025). The reverse model predicts a word-level distribution $L \times D$6, while cluster-level posteriors are obtained by Bayesian marginalization through the surjective map $L \times D$7. The resulting ELBO decomposes into a cluster-level cross-entropy term for mask inputs and a within-cluster word-level cross-entropy term for cluster inputs.

Hierarchical discrete flow matching replaces explicit forward Markov chains with a continuous-time probability path

$L \times D$8

and a probability velocity

$L \times D$9

so generation becomes a discrete denoising process implemented through flow matching rather than through a long discrete-time chain (Boget et al., 31 Mar 2026).

Hierarchical Koopman Diffusion departs from categorical-state corruption but still instantiates a discrete-time hierarchical diffusion viewpoint. It encodes an image into multiscale latent observables $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$0 and assumes linear latent dynamics

$$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$1

so the entire denoising trajectory has a closed-form evolution operator at each scale (Bai et al., 14 Oct 2025).

3. Hierarchy mechanisms and representative architectures

The literature suggests that no single hierarchy mechanism dominates. Some models place hierarchy in the state space, some in the tokenization, some in graph coarsening, and some in inference over denoising trajectories.

Model	Discrete object	Hierarchy mechanism
HKD (Bai et al., 14 Oct 2025)	Diffusion trajectory states in Koopman latent space	Scale-specific Koopman subspaces and operators across spatial resolutions
HiCoDiT (Ye et al., 17 Apr 2026)	12 RVQ codec token streams	Low-level $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$2 and high-level $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$3 blocks with dual-scale conditioning
ADDSE / RQDiT (Gonzalez, 25 Feb 2026)	$$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$4 RVQ code grid	Frame-DiT over time and Depth-DiT over RVQ depth
MolHIT (Jung et al., 19 Feb 2026)	Atom and bond categories	Clean $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$5 chemical-group $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$6 mask hierarchy plus decoupled atom encoding
HDLM (Zhou et al., 8 Oct 2025)	Word tokens	Word $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$7 cluster $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$8 mask semantic vocabulary
HDFM (Boget et al., 31 Mar 2026)	Node and edge labels	Coarse graph $$p(\mathcal{G})=\prod_{\ell=0}^{L} p_{\theta_\ell}(\mathcal{G}^\ell \mid \mathcal{H}^\ell)$$9 spanning supergraph $\mathcal{H}^\ell$0 fine graph factorization

A distinct architecture-free variant is structured preferential generation, where the hierarchy is a token-type ordering in the forward process. Different categories are masked at different times according to an information-based schedule, so the reverse process regenerates some token types earlier than others (Rissanen et al., 2024).

Prism occupies a different position. It is built on masked discrete diffusion LLMs, but it is not a new generative architecture; it is a test-time scaling framework layered on top. Its hierarchy lies in inference: exploration, thinning, and refinement stages over denoising trajectories, with pruning, partial remasking, and self-verification (Bai et al., 2 Feb 2026). This is still relevant because it shows that hierarchy can enter discrete diffusion at decoding time rather than only in training-time model design.

4. Objectives, samplers, and controllability

Despite the diversity of forward processes, the training objectives often reduce to denoising losses with structure-aware conditioning. HKD is trained from trajectories produced by a well-trained diffusion teacher and uses a reconstruction loss for the one-step map together with a trajectory consistency loss

$\mathcal{H}^\ell$1

which forces any teacher state $\mathcal{H}^\ell$2 to evolve to the same target clean image $\mathcal{H}^\ell$3 after Koopman propagation (Bai et al., 14 Oct 2025).

HiCoDiT uses multi-level Denoising Score Entropy,

$\mathcal{H}^\ell$4

plus an identity alignment term

$\mathcal{H}^\ell$5

and total loss

$\mathcal{H}^\ell$6

It also applies conditioning dropout so that lip, identity, and emotion inputs can be dropped during training to enable predictor-free guidance at inference (Ye et al., 17 Apr 2026).

ADDSE uses noisy-speech-conditioned denoising cross-entropy

$\mathcal{H}^\ell$7

with a time-independent network that predicts $\mathcal{H}^\ell$8. Because the network is time-independent, predictions can be reused when no new code becomes unmasked, reducing average network function evaluations; the paper reports average NFE $\mathcal{H}^\ell$9 for $[MASK]$0 (Gonzalez, 25 Feb 2026).

MolHIT derives a continuous-time NELBO for HDDM, but in practice it trains with the standard DiGress-style cross-entropy on clean nodes and edges. Its distinctive feature appears at sampling time through the Project-and-Noise sampler: sample a clean graph $[MASK]$1 from the model’s logits, then re-apply the forward kernel to obtain the next noisy graph. Temperature scaling and top-$[MASK]$2 truncation are applied to atom predictions, and classifier-free guidance is used for property-conditioned generation (Jung et al., 19 Feb 2026).

HDLM derives closed-form ELBOs in which the mask-level term is a cluster classification loss and the cluster-level term is a within-cluster word classification loss. Practical training then adds force-transition constraints during decoding, loss-weight clipping for numerical stability, and stochastic cluster perturbations with probability $[MASK]$3 to improve self-correction under mismatched contexts (Zhou et al., 8 Oct 2025).

HDFM uses simple cross-entropy over clean node and edge labels at each hierarchy level,

$[MASK]$4

which is notable because the hierarchical discrete diffusion effect comes primarily from the coarse-to-fine graph factorization and sparse supergraph conditioning rather than from a complicated loss (Boget et al., 31 Mar 2026).

5. Empirical regimes and applications

In image generation, HKD is positioned as a one-step model that retains trajectory access. On CIFAR-10, it reports FID-50k $[MASK]$5 in one step, compared with $[MASK]$6 for DMD and $[MASK]$7 for CD (LPIPS), while on FFHQ $[MASK]$8 it reports $[MASK]$9 in one step versus $\{1,\dots,K+1\}$0 for ECM. Its qualitative contribution is equally central: because the Koopman dynamics are explicit, intermediate states can be reconstructed analytically, spectral modes can be masked or injected, and “frequency-aware” editing can modify fine details while leaving coarse structure intact (Bai et al., 14 Oct 2025).

In speech and audio, hierarchical discrete diffusion is closely tied to residual vector quantization. HiCoDiT exploits the hierarchy of 12 codec levels, aligning lip motion and identity with low-level blocks and emotion with high-level blocks through dual-scale adaptive instance layer normalization, and the paper reports improved fidelity and expressiveness over baselines in video-to-speech generation together with stronger out-of-domain performance on film data (Ye et al., 17 Apr 2026). ADDSE uses a hierarchical discrete latent space for speech enhancement, where non-intrusive metrics plateau around $\{1,\dots,K+1\}$1–$\{1,\dots,K+1\}$2 diffusion steps and performance is particularly strong at low signal-to-noise ratios (Gonzalez, 25 Feb 2026).

In language modeling, HDLM reports consistently lower validation and generative perplexity than masked or flat discrete diffusion baselines; for example, HDLM-small-64 achieves validation perplexity $\{1,\dots,K+1\}$3 and generative perplexity $\{1,\dots,K+1\}$4, compared with MDLM-small at $\{1,\dots,K+1\}$5 and $\{1,\dots,K+1\}$6, and HDLM-base-128 reaches validation perplexity $\{1,\dots,K+1\}$7 and generative perplexity $\{1,\dots,K+1\}$8 (Zhou et al., 8 Oct 2025). Prism addresses a different problem—test-time scaling for masked diffusion LLMs—and shows that hierarchical trajectory search can match or exceed best-of-16 performance with much lower NFE; for LLaDA 8B Instruct on GSM8K, Prism with $\{1,\dots,K+1\}$9 reaches $(K+1)$0 at $(K+1)$1 NFE, compared with best-of-16 at $(K+1)$2 and $(K+1)$3 NFE (Bai et al., 2 Feb 2026).

In graph and molecular generation, the hierarchy is both structural and computational. HDFM reports large speed gains by combining sparse hierarchical supergraphs with discrete flow matching: on ZINC250k, HDFM requires $(K+1)$4 seconds at $(K+1)$5 NFE and $(K+1)$6 seconds at $(K+1)$7 NFE, versus $(K+1)$8 seconds for DiGress at $(K+1)$9 NFE and $M$0 seconds for SparseDiff at $M$1 NFE (Boget et al., 31 Mar 2026). MolHIT, which explicitly names Hierarchical Discrete Diffusion Models as its foundation, reports MOSES validity $M$2, quality $M$3, and FCD $M$4, and further shows strong downstream performance in multi-property guided generation and scaffold extension (Jung et al., 19 Feb 2026).

6. Related research, misconceptions, and outlook

A recurring misconception is to equate hierarchy exclusively with multiresolution image pyramids. The surveyed literature places hierarchy over token categories, RVQ levels, graph coarsenings, semantic vocabularies, chemical groupings, and even inference trajectories. Another is to treat every hierarchical diffusion model as discrete. Several influential neighbors are hierarchical but continuous: Hierarchical Diffuser uses Gaussian diffusion for temporally abstract planning with jumpy sub-goals (Chen et al., 2024); SIHD builds multi-scale continuous trajectory diffusions from structural entropy and state communities in offline RL (Zeng et al., 26 Sep 2025); Groupwise Diffusion Model uses continuous flow matching with non-overlapping group schedules and a frequency-domain hierarchy (Lee et al., 2023); and Diffuse-TreeVAE conditions a continuous DDPM on a learned latent tree with a discrete leaf index (Goncalves et al., 2024).

These adjacent works matter because they supply design templates even when the noise process is not discrete. Groupwise Diffusion Model shows that grouping and order of generation are first-class design choices, and its frequency-domain version demonstrates a coarse-to-fine hierarchy in which low-frequency groups encode high-level semantics while higher bands refine local details (Lee et al., 2023). The phase-transition analysis of diffusion on a Random Hierarchy Model goes further: it studies a fully discrete hierarchical latent tree with Gaussian leaf diffusion and finds a critical time beyond which high-level features such as class labels abruptly become unrecoverable while low-level features decay smoothly. This provides a theoretical explanation for why diffusion time can behave like a scale parameter in hierarchical data (Sclocchi et al., 2024).

The main limitations are equally diverse. Some models depend on hand-designed or pretrained hierarchies, as in MolHIT’s chemically curated grouping and HDLM’s embedding-based semantic clustering (Jung et al., 19 Feb 2026, Zhou et al., 8 Oct 2025). Others rely on teacher trajectories or two-stage pipelines, as in HKD and Diffuse-TreeVAE (Bai et al., 14 Oct 2025, Goncalves et al., 2024). Inference wrappers such as Prism introduce hyperparameter sensitivity in pruning windows, survivor counts, and remasking policies (Bai et al., 2 Feb 2026). Hierarchical graph generators can become harder to train in low-data regimes, and continuous analogues such as SIHD and Hierarchical Diffuser still face the usual offline-RL dependence on dataset support (Zeng et al., 26 Sep 2025, Chen et al., 2024).

The outlook suggested by these papers is toward richer learned hierarchies, not merely deeper ones. HDLM explicitly points to learned semantic mappings and more than one intermediate level; MolHIT highlights applicability beyond molecules to text and images; HKD proposes higher-resolution images, videos, and 3D data; and HDFM suggests that coarse-to-fine discrete denoising on sparse structures can scale to broader discrete domains (Zhou et al., 8 Oct 2025, Jung et al., 19 Feb 2026, Bai et al., 14 Oct 2025, Boget et al., 31 Mar 2026). A plausible implication is that “hierarchical discrete diffusion model” is best understood not as a single architecture, but as a general design principle: denoising should proceed through semantically or structurally meaningful intermediate discrete states rather than through a flat categorical space alone.