Discrete Spatial Enhancement (DSE)

• DSE introduces a frequency-domain prior using DCT to preserve coarse spatial structure while isolating high-frequency semantic residuals.
• It fuses low-frequency components from previous logits with interpolated high-frequency details, avoiding issues like blockiness and oversmoothing.
• Empirical results show improved fidelity and diversity, with sharper object details and lower FID scores compared to spatial interpolation methods.

Discrete Spatial Enhancement (DSE) is the frequency-domain prior-construction mechanism introduced within Scaled Spatial Guidance (SSG) for multi-scale visual autoregressive (VAR) generation. In this setting, DSE takes the previous scale’s logits, which encode a coarser spatial structure, and constructs a prior at the current scale that is structurally faithful at low frequencies but also plausible at higher frequencies. This prior is then subtracted from the current-step logits to form the semantic residual that SSG amplifies. Within the SSG framework, DSE is therefore not the guidance rule itself, but the module that makes the residual meaningful in a coarse-to-fine generation hierarchy (Shin et al., 5 Feb 2026).

1. Definition and functional role

DSE is devised as the coarse-prior construction step for SSG in visual autoregressive models that generate images through next-scale prediction. Its immediate purpose is to estimate, at the current scale, what the previous step has already explained. The resulting prior, denoted $\ell_{\text{prior}}$, is used to define the semantic residual

$\Delta_k = \ell_k - \ell_{\text{prior}},$

where $\ell_k$ is the current-step logit tensor. SSG then amplifies this residual rather than directly modifying the model through retraining (Shin et al., 5 Feb 2026).

This design follows the paper’s central coarse-to-fine principle: each next-scale prediction step should focus on novel detail rather than redundantly re-predicting coarse content. DSE operationalizes that principle by constructing a prior from the previous logits $\ell_{\text{prev}}$ and the target size $S_k$. If the prior is poor, the residual becomes misaligned and may either suppress useful detail or amplify artifacts. DSE is introduced precisely to avoid that failure mode.

A plausible implication is that DSE serves as the structural bridge between the already-established coarse state and the current finer prediction. In the terminology of the paper, it helps isolate a residual that is more concentrated on genuinely new information.

2. Information-theoretic motivation and the need for a frequency-domain prior

SSG is motivated by an information-theoretic account of train-inference discrepancy in VAR generation. The paper states that the ideal residual at step $k$ should be informative about the final image’s high-frequency content while being uninformative about the already-established low-frequency structure. This is summarized by the frequency-domain approximation

$$\mathcal{L}_{\text{VAR-IB}} \approx \max_{z_k} \beta I(z_k; H(\hat{f}_K)) - I(z_k; L(\hat{f}_K)).$$

Under this framing, the current step should contribute high-frequency content not explained by earlier scales (Shin et al., 5 Feb 2026).

DSE is needed because SSG requires an estimate of “what the previous step already explained” at the current scale. The paper argues that naive spatial interpolation is not ideal for this purpose. Nearest neighbor introduces blocky discontinuities and spurious high frequencies, whereas linear interpolation overly smooths and attenuates structure. In both cases, the resulting prior can corrupt the residual that SSG is meant to emphasize.

DSE addresses this limitation by moving prior construction into the frequency domain, where coarse and fine components can be separated more cleanly. The paper emphasizes that the frequency domain provides a global, non-interfering representation: low-frequency coefficients represent coarse global structure, high-frequency coefficients represent fine detail and local variation, and the chosen transform is orthonormal, energy-preserving, and exact under forward and inverse mappings. This motivates spectral fusion rather than pixel-space resizing.

3. Algorithmic construction

Conceptually, DSE performs a form of spectral fusion. It uses the previous-step logits for reliable coarse structure, but supplements them with an upsampled interpolation that provides a plausible continuation of missing high-frequency content. The construction is defined as follows (Shin et al., 5 Feb 2026).

Input: previous logits $\ell_{\text{prev}}$, target size $S_k$ Output: prior logits $\ell_{\text{prior}}$

The procedure is:

1. If there is no previous logits tensor, DSE is skipped.
2. Compute a spatial interpolation:

$\Delta_k = \ell_k - \ell_{\text{prior}},$0

1. Apply the Discrete Cosine Transform (DCT) to both tensors:

$\Delta_k = \ell_k - \ell_{\text{prior}},$1

2. Initialize the fused spectrum with the interpolated spectrum:

$\Delta_k = \ell_k - \ell_{\text{prior}},$2

3. Overwrite the low-frequency block of $\Delta_k = \ell_k - \ell_{\text{prior}},$3 with the low-frequency block from $\Delta_k = \ell_k - \ell_{\text{prior}},$4:

$\Delta_k = \ell_k - \ell_{\text{prior}},$5

4. Reconstruct the prior with inverse DCT:

$\Delta_k = \ell_k - \ell_{\text{prior}},$6

The paper specifies the discrete transform as the DCT, chosen because it is orthonormal, energy-preserving, and cleanly separates frequency bands. The low-frequency coefficients from the original coarse logits are preserved as the trusted representation of global shape and layout, while the high-frequency coefficients from the interpolated logits provide a plausible continuation after upsampling. The fused signal is thus intended to be better than either naive interpolation or a raw resize.

This construction implies a strict asymmetry between coarse and fine content. The previous logits are treated as authoritative for the coarse band, whereas the interpolated logits are used only as a scaffold for the missing higher-frequency band.

4. Mathematical placement inside SSG

Within SSG, DSE is the prior-construction module that enables residual-based guidance in logit space. Once the prior has been built, the guided logits are defined by

$\Delta_k = \ell_k - \ell_{\text{prior}},$7

where $\Delta_k = \ell_k - \ell_{\text{prior}},$8 is the stepwise guidance scale. The paper derives this rule from the MAP-style objective

$\Delta_k = \ell_k - \ell_{\text{prior}},$9

whose maximizer is

$\ell_k$0

DSE is therefore upstream of the guidance rule: it provides the prior term that makes $\ell_k$1 interpretable as a semantic residual rather than a generic logit difference (Shin et al., 5 Feb 2026).

During inference, the workflow is stepwise. At each autoregressive scale $\ell_k$2, the model produces raw logits $\ell_k$3. If $\ell_k$4, no prior is needed and SSG is not applied. For $\ell_k$5, DSE uses the previous logits $\ell_k$6 to construct $\ell_k$7 at the current scale. The residual is then computed, SSG is applied, and the next token map is sampled from the guided logits.

The paper emphasizes that this occurs online, at inference time, and does not require retraining or extra forward passes. Previous logits are cached and reused, so the overhead is very small. Because DSE works directly on logits and uses only cached values plus DCT/IDCT, it is described as model-agnostic, training-free, and fast.

5. Residual isolation, spectral behavior, and empirical ablation

The principal claim for DSE is that it improves the isolation of the semantic residual from a coarser prior. It does so in two stated ways: it preserves the coarse structure from the original previous logits, and it supplies a plausible high-frequency continuation through interpolation in frequency space. The paper’s interpretation is that the resulting difference

$\ell_k$8

becomes a better proxy for the step’s genuine high-frequency semantic content (Shin et al., 5 Feb 2026).

The paper’s frequency analysis supports that interpretation. It reports that SSG suppresses redundant low-frequency updates while boosting spectral energy above the previous step’s Nyquist threshold. This behavior is presented as consistent with DSE producing a prior that faithfully captures the coarse band and lets the residual concentrate on new detail.

The ablation on VAR-d16 directly compares prior constructions:

Prior construction	FID	IS
Baseline without SSG	3.42	275.6
Spatial nearest-neighbor prior	4.02	229.1
Spatial linear prior	3.79	234.8
Frequency-domain DSE with zero padding	3.34	277.6
Frequency-domain DSE without decay schedule	3.63	287.8
Full DSE + linear decay schedule	3.27	285.3

These results are used in the paper to support four conclusions. First, spatial interpolation harms performance, even relative to the no-SSG baseline. Second, frequency-domain prior construction is clearly better than spatial priors. Third, the full DSE formulation gives the best balance of fidelity and diversity when paired with the decay schedule. Fourth, DSE is important on its own, but its benefit is maximized when the guidance scale $\ell_k$9 is properly scheduled.

The paper also notes qualitative improvements: sharper object parts, fewer structural artifacts, and better preservation of coherent shapes. A plausible implication is that DSE’s contribution is most visible when the coarse-to-fine hierarchy would otherwise drift, because its prior construction directly regularizes that hierarchy at the level of the logits.

6. Scope, applicability, and terminological ambiguity

DSE is presented as broadly compatible with VAR models leveraging discrete visual tokens, regardless of tokenization design or conditioning modality. In the broader SSG pipeline, its role is fixed: input to DSE is previous-step logits $\ell_{\text{prev}}$0 and target resolution $\ell_{\text{prev}}$1; output of DSE is current-scale prior logits $\ell_{\text{prev}}$2; input to SSG is raw current logits $\ell_{\text{prev}}$3 and $\ell_{\text{prev}}$4; and output of SSG is guided logits $\ell_{\text{prev}}$5. This is why the paper states that SSG can be applied broadly to VAR, HART, and Infinity-style discrete-token generators (Shin et al., 5 Feb 2026).

The acronym “DSE” is, however, highly overloaded in the arXiv literature. In other works it denotes the Dyson-Schwinger equation framework in QCD (Tandy, 2014), design space exploration for accelerator tuning (Wang et al., 18 May 2025), directional speech enhancement in spatial speech perception (Shao et al., 2 Jul 2026), distribution system state estimation in power systems (Shafiei et al., 2017), and dynamic searchable encryption in encrypted databases (Liu et al., 2024). Within SSG, by contrast, DSE refers specifically to Discrete Spatial Enhancement, the DCT-based prior-construction mechanism for residual isolation in coarse-to-fine visual autoregressive generation.

This terminological ambiguity makes contextual definition essential. In the SSG paper, DSE is neither a general-purpose enhancement module nor a standalone generation method. It is the frequency-domain procedure that constructs a coarse prior so that residual guidance can emphasize genuinely new semantic detail at each scale.