Research map / Image generation / Super-resolution

NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution

NeurIPS 2026 · Image generation › Super-resolution · approach: Perceptual and real-world SR · 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 NSARM, a pure autoregressive Real-ISR model that builds on the pre-trained 8B Infinity bitwise next-scale model by replacing its first seven scales with a transformation network's output from the LR image, trained in two stages, to improve robustness and speed.

Problem

Real-world image super-resolution with diffusion-based methods suffers from over-enhancement, hallucinations, and failure cases across varying inputs, while existing AR Real-ISR methods are slow or limited by discrete codebooks.

Approach

Use the LR image to directly generate the preliminary coarse scales of a bitwise next-scale AR model (via a lightweight transformation network), then fully fine-tune the whole Infinity model with its original bitwise cross-entropy objective, so the generation pathway starts from the LR content.

Key results

• On RP60 at 512x512, NSARM achieves the best CLIPIQA (0.8564), MUSIQ (73.95) and TOPIQ (0.7615) averages among compared methods, with lowest TOPIQ variance (0.0021).
• NSARM reports 1.2 s inference at 1024x1024 versus 11.7-13 s for SUPIR and SeeSR, and 380 s for PURE at 512x512; total parameters are 8,760M, the largest in the comparison.
• In a 12-participant user study, NSARM received 55.6% of first-choice votes (267 of 480), with VARSR second at 14%.

Surprising findings

• On full-reference metrics NSARM trails SD-based methods substantially on DRealSR and RealSR (e.g., PSNR 26.66 vs 28.88 for PiSA-SR on DRealSR), attributed to bitwise quantization reducing pixel fidelity.
• Prompt choice has only marginal effect on NSARM; a fixed generic prompt performs nearly as well as LLaVA-generated captions.

Limitations

• Bitwise generation reduces pixel-level fidelity compared with continuous-space methods, with weaker PSNR/SSIM/LPIPS.
• Fixed output size of 1024x1024 from the base Infinity model; patch-based inference or resizing needed for other sizes.
• Very large parameter count (8.76B) with no speed or size advantage over lightweight methods.

Experimental setup

Datasets: Unsplash (1M images) DIV2K Validation (100 crops, test) RealSR DRealSR RP60

Benchmarks: DIV2K DRealSR RealSR RP60

Metrics: PSNR SSIM LPIPS NIQE MANIQA MUSIQ CLIPIQA TOPIQ variance Deficient/Poor/Collapse failure counts

Baselines: SeeSR SUPIR OSEDiff PiSA-SR PURE VARSR

Builds on: Infinity VAR VARSR PURE binary spherical quantization

Closest papers by meaning

• Visual Autoregressive Modeling for Image Super-Resolution (ICML 2025 · similarity 0.91)
• Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.90)
• Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
• TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.89)
• HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.89)
• SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.89)