Scaled Spatial Guidance (SSG) in VAR

• Scaled Spatial Guidance is a training-free, inference-time method that refines per-scale residual logits to inject new high-frequency detail in multi-scale visual autoregressive generation.
• It leverages frequency-aware priors via Discrete Spatial Enhancement to preserve coarse structure while enabling controlled high-frequency extrapolation.
• Empirical analyses show that SSG improves FID, sFID, and IS on benchmarks, and its principle extends to varied spatially heterogeneous guidance methods in diffusion and video generation.

Searching arXiv for the cited SSG papers and closely related work to ground the article. arxiv_search query: (Shin et al., 5 Feb 2026) OR "Scaled Spatial Guidance" OR "SSG: Scaled Spatial Guidance for Multi-Scale Visual Autoregressive Generation" OR (Li et al., 29 Apr 2026) OR (Shen et al., 2024) OR (Mao et al., 8 Jan 2026) OR (Zhang et al., 9 Apr 2026) OR (Hu et al., 23 Aug 2025) Scaled Spatial Guidance (SSG) most specifically denotes a training-free, inference-time guidance method for multi-scale visual autoregressive (VAR) image generation that modifies per-scale residual logits so that each generation step contributes high-frequency content not already explained by coarser scales (Shin et al., 5 Feb 2026). In adjacent literatures, the phrase is also used more broadly for guidance schemes that replace a single global guidance strength with spatially varying or scale-aware control, including per-pixel, per-region, and cluster-level diffusion guidance (Li et al., 29 Apr 2026, Shen et al., 2024, Mao et al., 8 Jan 2026). The acronym is not stable across the literature: in some works, SSG instead expands to “Self-Swap Guidance” or “Spatial Signal Guided” (Zhang et al., 9 Apr 2026, Hu et al., 23 Aug 2025).

1. Terminological scope and acronym usage

The clearest formalization of SSG as a named method appears in “SSG: Scaled Spatial Guidance for Multi-Scale Visual Autoregressive Generation,” where SSG is defined for next-scale VAR models and paired with a frequency-domain prior construction called Discrete Spatial Enhancement (DSE) (Shin et al., 5 Feb 2026). Other papers use the same phrase descriptively rather than as a fixed algorithmic name, or use the acronym for different expansions entirely.

Paper	Expansion or usage of “SSG”	Role
(Shin et al., 5 Feb 2026)	Scaled Spatial Guidance	Specific VAR inference method
(Li et al., 29 Apr 2026)	SSG as spatially varying guidance; SAMG realization	Per-pixel diffusion guidance
(Shen et al., 2024)	Not named SSG, but presented as an operationalization of it	Region-wise CFG rescaling
(Mao et al., 8 Jan 2026)	SSG lens applied to FENCE	Cluster-level spatio-temporal guidance
(Zhang et al., 9 Apr 2026)	Self-Swap Guidance	Acronym collision; token-swap guidance
(Hu et al., 23 Aug 2025)	Spatial Signal Guided	Acronym collision in video DiT

This terminological dispersion matters because “Scaled Spatial Guidance” can refer either to the specific VAR method of (Shin et al., 5 Feb 2026) or, more loosely, to a broader design pattern in which guidance becomes spatially heterogeneous rather than globally uniform. A common misconception is that SSG names a single diffusion-time variant of classifier-free guidance; the literature instead shows at least three distinct usages: a logit-space VAR mechanism, a family resemblance among spatially adaptive guidance methods, and unrelated acronym expansions.

2. Information-theoretic formulation in multi-scale VAR

In the VAR setting of (Shin et al., 5 Feb 2026), image generation is factorized across coarse-to-fine scales. With token maps $r_k \in \{1,\dots,V\}^{h_k \times w_k}$, the joint distribution is

$$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$

A sampled token map is de-quantized to an embedding $z_k$, and the accumulated feature state evolves as

$$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$

The motivating problem is a train–inference discrepancy in next-scale VAR. Under teacher forcing, later scales learn to add progressively higher-frequency details. At inference, limited capacity and accumulated error can cause later scales to drift, re-predicting or distorting low-frequency structure instead of contributing scale-appropriate, novel detail. The paper reformulates this problem through an Information Bottleneck (IB) lens (Shin et al., 5 Feb 2026).

The classical IB objective is

$$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$

For VAR step $k$, the paper sets $X=\hat f_{k-1}$, $\tilde X=z_k$, and $Y=\hat f_K$, then reverses the usual compression perspective to maximize novel utility while penalizing redundancy:

$$\mathcal{L}_{\text{VAR-IB}} = \max_{z_k} \beta I(z_k;\hat f_K \mid \hat f_{k-1}) - I(\hat f_{k-1}; z_k).$$

Using the chain rule and the coarse-state approximation $$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$0, this becomes

$$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$1

and, in the frequency-domain interpretation with ideal low/high-pass filters $$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$2 and $$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$3 and $$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$4,

$$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$5

The interpretation given in the paper is precise: the step-$$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$6 residual should be informative about new high-frequency content while remaining uninformative about low-frequency content already established. This is the conceptual core of SSG in its original form.

At the operational level, SSG acts in logit space. If $$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$7 are the pre-softmax residual logits at scale $$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$8, then the scale-$$p(r_1,\dots,r_K) = p(r_1)\prod_{k=2}^K p(r_k \mid r_{<k}).$$9 semantic residual is defined as

$z_k$0

where $z_k$1 is a frequency-aware prior derived from the previous scale.

3. Discrete Spatial Enhancement and the SSG logit update

The prior construction in (Shin et al., 5 Feb 2026) is Discrete Spatial Enhancement (DSE), whose stated purpose is to transport only the coarse structure from the previous scale to the current resolution while avoiding over-smoothing from linear interpolation or blockiness from nearest-neighbor upsampling. DSE applies a 2D DCT-II with orthonormal normalization over the spatial dimensions of the logits for each vocabulary channel, combines spectra by hard low-frequency replacement, and returns to spatial logits with the inverse DCT-III.

Let $z_k$2 be previous-step logits and $z_k$3 the current-resolution interpolant. After taking DCTs of both tensors, DSE replaces the top-left low-frequency block of the interpolated spectrum by the exact low-frequency coefficients of the coarse spectrum:

$z_k$4

where $z_k$5 for $z_k$6 and $z_k$7, and $z_k$8 otherwise. The prior is then

$z_k$9

The paper emphasizes two consequences of this construction: DSE preserves coarse structure exactly through hard low-frequency replacement, and it carries a plausible high-frequency extrapolation from the interpolant rather than injecting artificial blockiness (Shin et al., 5 Feb 2026). A zero-padding variant that removes extrapolated detail is denoted DSE†.

SSG then converts the information-theoretic objective into a closed-form logit-space update. Given the semantic residual $$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$0, the paper defines a concave MAP-style surrogate

$$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$1

whose unique maximizer is

$$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$2

A per-scale schedule is recommended:

$$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$3

The paper states that this linear decay is effective and stable in practice, whereas a fixed $$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$4 can overemphasize late high-frequency updates and degrade FID even if IS rises. After the update, sampling proceeds as usual:

$$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$5

optionally with top-$$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$6 or nucleus sampling if those are already present in the base system. No additional normalization or clipping is required beyond the base temperature $$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$7 (Shin et al., 5 Feb 2026).

Algorithmically, SSG is minimal. At each scale $$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$8, the model computes base logits $$\hat f_k = \hat f_{k-1} + U(z_k), \qquad \hat f_0 = 0.$$9, constructs $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$0 from cached raw logits $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$1 by DSE, forms $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$2, applies the affine update, samples tokens, de-quantizes them to $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$3, accumulates $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$4, and caches raw logits for the next step. Its extra cost per scale is an interpolation plus DCT/IDCT on an $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$5 tensor, with GPU-batched transform complexity $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$6 per channel (Shin et al., 5 Feb 2026).

4. Empirical profile, ablations, and operating characteristics

The empirical profile reported in (Shin et al., 5 Feb 2026) is consistent across class-conditional ImageNet, higher-resolution ImageNet, and text-to-image benchmarks. On ImageNet $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$7 with 10-step class-conditional generation, SSG improves FID, sFID, and IS across all reported VAR scales; the strongest reported configuration is VAR-d30, where FID improves from 2.02 to 1.68 and IS from 302.9 to 313.2.

Setting	Baseline $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$8 SSG	Latency
ImageNet $$\mathcal{L}_{\text{IB}} = \min_{\tilde X} I(X;\tilde X) - \beta I(\tilde X;Y).$$9, VAR-d16	FID 3.42 $k$0 3.27; sFID 8.70 $k$1 8.39; IS 275.6 $k$2 285.3	Time unchanged
ImageNet $k$3, VAR-d30	FID 2.02 $k$4 1.68; sFID 8.52 $k$5 8.50; IS 302.9 $k$6 313.2	10 steps
ImageNet $k$7, VAR-d36	FID 2.70 $k$8 2.39; IS 290.6 $k$9 320.6	Time unchanged
MJHQ-30K, HART-0.7B	FID 8.46 $X=\hat f_{k-1}$0 7.28; CLIPScore 0.2819 $X=\hat f_{k-1}$1 0.2834	Time $X=\hat f_{k-1}$2 constant
MJHQ-30K, Infinity-2B	FID 10.01 $X=\hat f_{k-1}$3 9.68; CLIPScore 0.2754 $X=\hat f_{k-1}$4 0.2767	Time $X=\hat f_{k-1}$5 constant

The same paper reports that SSG+VAR-d30, with FID 1.68 at 10 steps, is competitive with or surpasses diffusion and masked autoregressive baselines such as DiffiT at FID 1.73 with 250 steps and MAR-H at FID 1.78 with 64 steps (Shin et al., 5 Feb 2026). The result is important less as a cross-family ranking than as evidence that guidance on the inherited coarse-to-fine hierarchy can unlock substantial gains without retraining or additional denoising iterations.

The ablations isolate the two core design choices. First, purely spatial priors degrade performance: nearest-neighbor prior gives FID 4.02 and linear interpolation prior gives FID 3.79 against a baseline FID of 3.42. Frequency-domain priors instead improve it: DSE† yields FID 3.34 and full DSE yields FID 3.27, with IS gains and unchanged latency. Second, the $X=\hat f_{k-1}$6 schedule matters: fixed $X=\hat f_{k-1}$7 overguides, producing FID 3.63 and IS 287.8, whereas linear decay gives FID 3.27 and IS 285.3 (Shin et al., 5 Feb 2026).

The paper also reports a spectral analysis: SSG suppresses spectral energy below the previous Nyquist frequency and boosts it above that frequency, aligning the generation step with the intended high-frequency novelty. Pixel-level spectra align better to the reference dataset while avoiding excessive high-frequency noise. Across temperature sweeps, the FID–IS Pareto frontier improves: for similar IS, SSG gives lower FID, and for similar FID, it gives higher IS. Mean wall-clock time changes by at most 1–2%, which the paper describes as effectively negligible (Shin et al., 5 Feb 2026).

5. Diffusion-era generalizations of scaled spatial guidance

Outside VAR, several papers instantiate the same broad principle—spatially heterogeneous guidance—through different mechanisms. This suggests a broader, concept-level reading of “scaled spatial guidance”: not a single update rule, but a family of methods that localize guidance strength according to structure, semantics, or uncertainty.

“Delta Score Matters! Spatial Adaptive Multi Guidance in Diffusion Models” derives a per-pixel guidance energy from the conditional–unconditional difference,

$X=\hat f_{k-1}$8

normalizes it, and maps it affinely to a local guidance scale between $X=\hat f_{k-1}$9 and $\tilde X=z_k$0. The final guided prediction is

$\tilde X=z_k$1

The paper interprets standard CFG as tangential linear extrapolation on a curved data manifold and uses the resulting deviation analysis to motivate conservative scaling at high-energy boundaries and aggressive scaling in low-energy regions. Reported gains include SDXL COCO FID 27.98 to 25.41 with CLIPScore 19.51 to 20.14, and SD3.5-M GenEval All score 0.63 to 0.66; in video, ModelScope-1.7B improves CHScore Flow 70.47 to 72.10, Frame LPIPS 8.12 to 7.35, Frame SSIM 79.13 to 80.50, and MTScore CLIP 10.56 to 10.78 (Li et al., 29 Apr 2026).

“Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance” formulates a region-wise version of CFG. It constructs semantic masks from cross-attention and self-attention, then composes guidance as

$\tilde X=z_k$2

The per-region scales are chosen to uniformize aggregate guidance magnitude across regions relative to a benchmark mask. On SD-v1.5 with DPMSolver++ and $\tilde X=z_k$3, the paper reports FID 12.466 to 12.059 and CLIP 0.3223 to 0.3226; on IF with DPMSolver++ and $\tilde X=z_k$4, it reports FID 15.31 to 13.99 at unchanged CLIP 0.3280 (Shen et al., 2024).

“Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation” extends the same logic to spatio-temporal data. FENCE uses a dynamic posterior-based guidance scale,

$\tilde X=z_k$5

and computes it at cluster level after grouping nodes by attention-derived spatial correlation features. The resulting guided prediction is

$\tilde X=z_k$6

Across PEMS04/07/08 with SR-TC and SC-TC missingness at $\tilde X=z_k$7, FENCE is reported to beat the second-best method by an average 6.26% in MAPE, with cluster-level scaling outperforming a uniform global scale in ablations (Mao et al., 8 Jan 2026).

These methods are mechanistically different. SAMG is energy-driven and per-pixel, S-CFG is mask-driven and per-region, and FENCE is posterior-driven and cluster-level. The shared principle is spatial heterogeneity in guidance strength.

6. Related acronym collisions, limitations, and interpretive boundaries

The acronym SSG is not unique to Scaled Spatial Guidance. In “Guiding a Diffusion Model by Swapping Its Tokens,” SSG means “Self-Swap Guidance.” The method creates a perturbed branch by swapping semantically dissimilar spatial or channel tokens, defines

$\tilde X=z_k$8

and guides with

$\tilde X=z_k$9

The paper explicitly states that this is not Scaled Spatial Guidance by name, although it is a scaled spatial guidance mechanism “in the practical sense” because it uses spatially targeted perturbations and a tunable guidance strength. Reported conditional SDXL results on COCO 2014 include FID 21.73, CLIP 0.313, IS 34.63, AES 5.902, PickScore 22.17, and ImageReward 0.276; unconditional COCO 2014 results include FID 70.91 and IS 16.44 (Zhang et al., 9 Apr 2026).

In “SSG-DiT: A Spatial Signal Guided Framework for Controllable Video Generation,” SSG expands to “Spatial Signal Guided,” not Scaled Spatial Guidance. The method uses CLIP-derived $Y=\hat f_K$0 attention and MLP masks to build a spatially prompted image and injects the resulting visual condition into a frozen video DiT through a dual-branch attention adapter:

$Y=\hat f_K$1

The paper explicitly states that it does not introduce an explicit scalar guidance coefficient in its reported experiments; any multiplier on $Y=\hat f_K$2 would be a practical extension rather than part of the reported setup. On VBench, the full model reports Subject Consistency 97.40, Spatial Relationship 78.17, and Overall Consistency 26.31, while the ablation without SSG drops Overall Consistency to 18.91 (Hu et al., 23 Aug 2025).

For the original VAR SSG, the limitations are specific and concrete. The method is sensitive to the quality of the previous step’s logits: if they are distorted, the prior can misalign the semantic residual and suppress useful detail. It cannot fully correct severe initial mistakes or tokenizer bottlenecks, does not force separation for ambiguous prompts or classes that blend with the background, and can overguide when $Y=\hat f_K$3 is large and non-decayed, raising IS while harming FID. Methods that do not expose logits at sampling, such as continuous-time diffusion without categorical stages, require adaptation for direct application (Shin et al., 5 Feb 2026).

Across the broader literature, limitations recur at the level of spatial decomposition. SAMG notes dense semantic overlaps as a failure mode; S-CFG notes mis-segmentation, token ambiguity, and instability for extremely small regions; FENCE notes noisy attention and mis-clustering under extreme sparsity (Li et al., 29 Apr 2026, Shen et al., 2024, Mao et al., 8 Jan 2026). A plausible implication is that “scaled spatial guidance” is most reliable when the mechanism used to localize guidance—frequency separation, attention-derived masks, energy maps, or clusters—faithfully captures the structure that the model should preserve.

In that sense, Scaled Spatial Guidance is best understood not merely as one named method, but as a general principle of replacing uniform guidance with structure-aware, location-sensitive guidance. Its most formal and compact realization remains the VAR logit update of (Shin et al., 5 Feb 2026), where the principle is expressed as a frequency-aware prior, a semantic residual, and a single closed-form affine update.