Research map / Architectures and efficiency / RNNs and normalization

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

NeurIPS 2025 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TTS-VAR, a test-time scaling framework for Infinity visual auto-regressive text-to-image models combining adaptive descending batch sizes, DINOv2-feature clustering-based diversity search at early scales, and reward-weighted resampling at late scales.

Problem

Test-time scaling for visual auto-regressive (VAR) image generation is unexplored; early-scale tokens are hard to evaluate with reward models and cannot be revised once generated, so diffusion-style best-of-N or importance sampling underperforms.

Approach

Align test-time search with VAR's coarse-to-fine causal generation: cluster intermediate samples by DINOv2 structural features at coarse scales to keep diversity, and resample by a path-aware potential score (VALUE, exp of reward) at late scales, under a descending batch-size schedule.

Key results

• On Infinity2B, TTS-VAR at N=8 raises GenEval from 0.6946 to 0.7530 (about 8.7% relative gain), above Best-of-N's 0.7364 at N=8.
• TTS-VAR with N=2 reaches GenEval 0.7403, exceeding Best-of-N at N=8 (0.7364) with a quarter of the samples.
• On Infinity8B, GenEval improves from 0.7646 to 0.7931 at N=1 and reaches 0.8189 at N=4.

Measured effects reported in the paper

Surprising findings

• Resampling at early scales (e.g., scale 3) noticeably degrades final results, so resampling must be restricted to later scales.
• Clustering-based diversity search alone gives only moderate gains, and on DPG-Bench at N=4 TTS-VAR's Global score (83.37) falls below Best-of-N's (88.02).

Limitations

• Remaining failure cases in text-image position alignment, as acknowledged by the authors.
• Applicability to other coarse-to-fine models such as 1-D tokenizer autoregressive models is unexplored.
• Gains are reported on a single VAR family (Infinity) with ImageReward as the main guidance reward; user study is small (21 respondents).

Experimental setup

Datasets: GenEval prompts T2I-CompBench DPG-Bench

Benchmarks: GenEval T2I-CompBench DPG-Bench

Metrics: GenEval overall score ImageReward HPSv2.1 Aesthetic V2.5 CLIP-Score CLIP-Div user preference votes

Baselines: Best-of-N Importance Sampling Infinity2B Stable Diffusion 3 FLUX.1 DALLE 3

Builds on: Infinity Visual Auto-Regressive Modeling (VAR) DINOv2 K-Means++ ImageReward

Closest papers by meaning

• Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
• Markovian Scale Prediction: A New Era of Visual Autoregressive Generation (CVPR 2026 · similarity 0.90)
• ScalingAR: Scaling Confidence for Autoregressive Image Generation (ICML 2026 · similarity 0.90)
• Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction (CVPR 2025 · similarity 0.89)
• Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.87)
• When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.87)