TTS-VAR: Adaptive Visual & Time Series Models

• TTS-VAR is a framework that unifies test-time scaling for visual autoregressive decoding with time-varying vector autoregression for dynamic time series analysis.
• In visual generative modeling, it employs hierarchical decoding with adaptive batch scheduling and clustering-based diversity to balance exploration and reward maximization.
• In time series analysis, it uses local kernel smoothing and tensor decompositions to capture evolving dynamics and improve parameter estimation in nonstationary data.

TTS-VAR refers to a family of methodologies across diverse research fields, most prominently: (i) test-time scaling frameworks for visual auto-regressive generation, and (ii) time-varying vector autoregression for nonstationary time series analysis. The acronym "TTS-VAR" has distinct, domain-specific interpretations, yet centers on addressing challenges of scalability, adaptation, and efficient exploration in high-dimensional generative or temporal problems. In visual generative modeling, TTS-VAR (Test-Time Scaling for Visual Auto-Regressive generation) provides an algorithmic framework to optimize exploration and sample selection during hierarchical, coarse-to-fine autoregressive decoding. In time series econometrics and statistics, TTS-VAR often denotes time-varying vector autoregressive models, with a major focus on locally-stationary multivariate processes where mean and coefficient matrices evolve smoothly over time.

1. Problem Formulation and Objectives

TTS-VAR in the visual generative context conceptualizes the sampling process of a hierarchical VAR (visual auto-regressive) model as a $K$-step path search. Each step corresponds to decoding a set of discrete residual tokens $r_k$, yielding sequences $x = (r_1, ..., r_K)$ with progression across spatial scales. Given an explicit computational budget $T$, the challenge is to allocate sampling and selection costs across the causal chain such that the expected sample quality, as measured by a reward function $R(\cdot, c)$ (e.g., ImageReward or alignment to prompt $c$), is maximized upon complete generation. The effective search distribution is reweighted according to a potential:

$$p_{\theta'}(x) \propto p_\theta(x) \cdot \exp(\lambda R(D(x), c)), \qquad \text{subject to}~\mathrm{cost}(p_{\theta'}) \leq T$$

where $D(x)$ decodes partial generations, and $\lambda$ controls the temperature of reward biasing. The core goal is to design efficient, sample- and scale-adaptive test-time procedures that balance path diversity (exploration) and reward maximization (exploitation), improving alignment and quality under strict cost constraints (Chen et al., 24 Jul 2025).

In time series, TTS-VAR (time-varying VAR) models describe processes $X_t \in \mathbb{R}^r$ with smoothly evolving mean vector $r_k$0 and VAR matrix $r_k$1, for rescaled time $r_k$2:

$r_k$3

Parameters are estimated locally in time, addressing nonstationarity and temporal dynamics that static VARs cannot capture (Motta, 2021).

2. Hierarchical Generation and Adaptive Batch Scheduling

Visual TTS-VAR leverages the multi-scale, coarse-to-fine structure of models such as Infinity—an autoregressive generator that progressively refines image detail over $r_k$4 quantization scales. Early steps correspond to low-resolution layouts, while later steps add fine-grained content. Due to the quadratic scaling of memory and FLOPs with spatial resolution, batch sizes are scheduled to be largest at the coarsest scales and to decline adaptively at finer scales. For example, a typical 13-scale schedule is:

$r_k$5

where $r_k$6 is the candidate batch size at scale $r_k$7. This aggressive early exploration maintains diversity until costs become prohibitive (Chen et al., 24 Jul 2025).

3. Diversity Search and Resampling: Clustering and Potential Selection

A dual-phase candidate selection mechanism is implemented:

3.1. Clustering at Coarse Scales

At coarse scales ($r_k$8, such as $r_k$9), sample selection using reward functions is unreliable because partial images lack detailed semantics. TTS-VAR preserves structural variety through semantic clustering in feature space. Each candidate partial image is encoded (e.g., DINOv2 patch features, PCA reduction), then grouped via K-means++ into $x = (r_1, ..., r_K)$0 clusters. The candidate closest to each centroid is advanced. This ensures that varied spatial layouts ("structural seeds") are not pruned prematurely (Chen et al., 24 Jul 2025).

3.2. Potential-Based Resampling at Fine Scales

Once partial images become semantically rich enough ($x = (r_1, ..., r_K)$1, e.g. $x = (r_1, ..., r_K)$2), reward-based selection becomes informative. Each candidate is scored by $x = (r_1, ..., r_K)$3, and $x = (r_1, ..., r_K)$4 candidates are resampled using the induced distribution $x = (r_1, ..., r_K)$5. This procedure exploits potential-based importance sampling, efficiently concentrating resources on likely high-reward continuations (Chen et al., 24 Jul 2025).

4. Theoretical Considerations and Cost-Quality Tradeoffs

The design of TTS-VAR is informed by empirical consistency analysis: reward functions at early scales exhibit low predictive consistency with final reward ($x = (r_1, ..., r_K)$6 for $x = (r_1, ..., r_K)$7), justifying delayed resampling. The strict causal structure of VAR sampling precludes retroactive correction—erroneous early decisions are unrecoverable. Consequently, clustering-based diversity preservation at early scales is essential. Overall, the dual-stage strategy delivers most of the alignment and quality improvement of computationally expensive "Best-of-$x = (r_1, ..., r_K)$8" reranking, but at less than half the cost. No closed-form bounds are provided, but scale-wise analysis indicates near-optimality under resource constraints (Chen et al., 24 Jul 2025).

5. Experimental Validation and Comparative Results

On Generation Evaluation (GenEval) tasks with powerful VAR backbones (e.g., Infinity-2B), TTS-VAR achieves substantial empirical gains. For $x = (r_1, ..., r_K)$9 samples/prompt, the alignment score improves from $T$0 (raw) to $T$1 (+8.7%), outperforming both Best-of-8 ($T$2), Importance Sampling ($T$3), and large-scale diffusion models such as SD3 ($T$4), despite a smaller model footprint. Benchmarks confirm that TTS-VAR's combination of adaptive sampling, clustering-based diversity, and potential selection at appropriate scales is highly effective (Chen et al., 24 Jul 2025).

6. Implementation Details

Key implementation elements include:

• Batch size schedules hard-coded per scale.
• Semantic features from pre-trained frozen DINOv2 models, PCA-reduced to $T$5, and K-means++ clustering.
• Fixed reward model (e.g., ImageReward), with scaling parameter $T$6.
• No use of beam search or KV-caching; algorithms operate within the vanilla VAR decoding pipeline.
• Complete codebase is available at https://github.com/ali-vilab/TTS-VAR (Chen et al., 24 Jul 2025).

Pseudocode for the entire algorithm fits within approximately 20 lines, attesting to its modularity and ease of integration.

7. TTS-VAR in Time Series: Locally Stationary VAR Models

In time series, TTS-VAR denotes time-varying VAR models where both the mean vector $T$7 and VAR matrix $T$8 vary smoothly with rescaled time $T$9. The estimation procedure uses kernel-weighted least squares, with both local-constant and local-linear approaches. The estimator's mean squared error converges at rate $R(\cdot, c)$0 (local-constant) or $R(\cdot, c)$1 (local-linear, for bandwidth $R(\cdot, c)$2 optimally tuned), and offers closed-form solution matrices for practical implementation (Motta, 2021). Simulation studies confirm that joint estimation of $R(\cdot, c)$3 and $R(\cdot, c)$4 is crucial: local-linear smoothing substantially reduces bias, especially at boundaries, and provides sharper confidence bands than local-constant alternatives.

Applications to macroeconomic and high-dimensional functional data exploit the flexibility of TTS-VAR to account for gradual regime shifts, drift, or structural evolution, which are intractable for stationary VARs.

8. Extensions: High-Dimensional and Tensor Time-Varying VAR

To address over-parameterization in high-dimensional VARs, recent work employs tensor decompositions (e.g., CANDECOMP/PARAFAC, CP) for the time-varying coefficient tensor. The time-varying parameter tensor VAR (TVP-TVAR) allows only one factor matrix of the tensor decomposition to evolve over time, dramatically reducing the number of free parameters. Bayesian selection among competing configurations, together with "kneedle" rank selection based on DIC, yields models that recover true system structure and dynamics with over 90% parameter reduction. Empirical application to fMRI data demonstrates time-resolved Granger causality networks consistent with cognitive state transitions (Luo et al., 12 May 2025).

Summary Table: TTS-VAR Across Domains

Domain	Core Problem	Key Ideas
Visual Gen	Test-time scaling for VAR	Coarse-to-fine clustering, batch scheduling, reward-based resampling (Chen et al., 24 Jul 2025)
Time Series	Locally stationary VAR	Joint mean/matrix estimation via kernel smoothing (Motta, 2021)
High-dim VAR	TVP Tensor VAR (TVP-TVAR)	Low-rank CP decomposition, Bayesian rank/model selection (Luo et al., 12 May 2025)

TTS-VAR frameworks, whether in generative modeling or statistical time series analysis, address the challenge of dynamic adaptation to nonstationarity or hierarchical structure, optimizing performance under complex computational and inferential constraints.