TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation

TL;DR

TTS-VAR introduces a general test-time scaling framework for Visual Auto-Regressive (VAR) models by framing generation as path searching. It couples an adaptive descending batch-size schedule with two scale-specific strategies: clustering-based diversity at coarse scales to preserve structural variety, and resampling-based potential selection at finer scales to steer toward high-quality outputs. Empirical results on Infinity VAR show notable GenEval gains and robustness across model sizes, with analyses revealing the importance of early structural features and scale-dependent resampling effectiveness. The approach offers a resource-efficient route to improved VAR-based image synthesis and provides open-source code for reproduction.

Abstract

Scaling visual generation models is essential for real-world content creation, yet requires substantial training and computational expenses. Alternatively, test-time scaling has garnered growing attention due to resource efficiency and promising performance. In this work, we present TTS-VAR, the first general test-time scaling framework for visual auto-regressive (VAR) models, modeling the generation process as a path searching problem. To dynamically balance computational efficiency with exploration capacity, we first introduce an adaptive descending batch size schedule throughout the causal generation process. Besides, inspired by VAR's hierarchical coarse-to-fine multi-scale generation, our framework integrates two key components: (i) At coarse scales, we observe that generated tokens are hard for evaluation, possibly leading to erroneous acceptance of inferior samples or rejection of superior samples. Noticing that the coarse scales contain sufficient structural information, we propose clustering-based diversity search. It preserves structural variety through semantic feature clustering, enabling later selection on samples with higher potential. (ii) In fine scales, resampling-based potential selection prioritizes promising candidates using potential scores, which are defined as reward functions incorporating multi-scale generation history. Experiments on the powerful VAR model Infinity show a notable 8.7% GenEval score improvement (from 0.69 to 0.75). Key insights reveal that early-stage structural features effectively influence final quality, and resampling efficacy varies across generation scales. Code is available at https://github.com/ali-vilab/TTS-VAR.

Figures (10)

• Figure 1: TTS-VAR generates several samples concurrently like Best-of-N (BoN). In TTS-VAR, we adopt an adaptive descending batch size schedule to make the most of AR efficiency, with feature clustering at early scales to ensure diversity, and resampling according to potentials at late scales for more valuable samples. (1-3) are overviews showing the difference between raw inference, BoN, and TTS-VAR. (a) is a detailed example of the generation process of our method.
• Figure 1: Quantitative evaluation on GenEval.
• Figure 2: Different Batch Size Schedules. We visualize the memory usage in (a) and computation complexity in (b) for 13 scales during the generation of Infinity, with fixed batch size 1 and adaptive batch size. Specifically, the adaptive batch size here is [8,8,6,6,6,4,2,2,2,1,1,1,1]. This batch size schedule enables more possibilities with little additional consumption.
• Figure 3: Score Curves over Sample Number $N$.
• Figure 4: Qualitative Comparison. Each line shows results generated by Stable Diffusion 3 (SD3) stable-diffusion3, Infinity, and Infinity with test-time scaling strategies, with objects marked blue and relationships marked green.