Research map / Image generation / Autoregressive and latent generation

ExtraVAR: Stage-Aware RoPE Remapping for Resolution Extrapolation in Visual Autoregressive Models

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Next-scale and token autoregressive images · 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.

Training-free resolution extrapolation for VAR models via stage-aware RoPE frequency-band remapping plus an entropy-driven, closed-form per-head attention scaling calibration, evaluated on Infinity at up to 2560x2560.

Problem

VAR models are trained at a fixed maximum resolution and naive training-free extrapolation methods from LLMs or diffusion yield global repetition, local repetition, and detail degradation when generating beyond that resolution.

Approach

Each coarse-to-fine VAR stage is driven by a distinct dominant RoPE frequency band, so the RoPE remapping should be stage-specific (NoPE for very low bands, PI then interpolated to YaRN for others). Attention dispersion is calibrated with a resolution-invariant normalized entropy and a closed-form per-head scaling factor that matches the training-resolution entropy.

Key results

• On GenEval, ExtraVAR reaches 0.70 overall, improving over PI (0.57) by 0.13 when extrapolating Infinity to 2560x2560.
• On DPG-Bench, ExtraVAR reaches 83.58 overall, 0.89 above PI (82.69).
• On HPSv2.1, ExtraVAR improves over PI by 2.11 overall and achieves the best overall score among compared methods.

Surprising findings

• Zeroing or replacing very low-frequency RoPE dimensions with NoPE preserves visual quality when applied throughout generation, behaving near-NoPE.
• Naive PE (no extrapolation remapping) collapses to near-zero GenEval overall (0.02), and PE and RiFlex trail PI badly on HPSv2.1.

Limitations

• Evaluation limited to a single VAR model (Infinity) and maximum resolution of 2560x2560 due to hardware budget.
• Hyperparameters (k_l, k_h, m, tau_h=0.3, epsilon) are hand-set and rely on the three-stage partition specific to 13 scale steps.
• Attention calibration gains on GenEval are small on its own (+0.01 overall).

Experimental setup

Benchmarks: GenEval DPG-Bench HPSv2.1

Metrics: GenEval overall score DPG-Bench overall score HPSv2.1 score

Baselines: PE PI NTK-aware scaling YaRN RiFlex DyPE

Builds on: YaRN Position Interpolation NTK-aware scaling RiFlex DyPE UltraImage FastVAR

Closest papers by meaning

• SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers (NeurIPS 2026 · similarity 0.88)
• Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
• NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.87)
• HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.86)
• DREAM: Diffusion Rectification and Estimation-Adaptive Models (CVPR 2024 · similarity 0.86)
• ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.86)