Research map / Image generation / Fast diffusion sampling

Guiding Visual Autoregressive Models through Spectrum Weakening

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 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 Spectrum Weakening Guidance (SWG), a training-free, condition-free guidance method for visual autoregressive models that builds a weak model by DFT-based spectral selection along the channel dimension, with renormalization for stability.

Problem

Classifier-free guidance in visual autoregressive models requires a separately trained unconditional branch and is ill-defined for unconditional generation; existing weak-model guidance methods rely on diffusion-specific assumptions (timesteps, full-image layouts) that do not transfer to AR decoding

Approach

Construct a weak branch by applying a binary spectral selection mask (after DFT along the channel dimension) to intermediate features of the same model, reducing information in a controlled way, then extrapolate away from the weak prediction; two renormalization schemes (spectral and spatial) keep magnitudes stable.

Key results

• On NOVA (COCO), SWG improves unconditional FID from 85.34 to 50.37 and conditional FID from 46.86 to 15.80.
• On Lumina-mGPT, SWG improves conditional FID from 90.85 to 82.19 and CLIP score from 28.08 to 29.28.
• On RandAR, best FID of 6.55 at SWG scale 3.0; combining SWG with CFG reaches FID 5.26 at omega_c=3.8, omega_s=0.4.

Surprising findings

• Retaining only 10% of spectral components still yields a valid guidance signal, while retaining 90% degrades performance.
• Overly weak guidance appears to degrade fine details in generated images.

Limitations

• Information-loss theorem's strictness holds only under conditions that rarely hold in practice, so the theoretical justification is heuristic.
• Lumina-mGPT unconditional IS slightly decreases, attributed to Inception domain mismatch.
• Renormalization does not recover removed information; the method relies on empirically tuned selection ratios and guidance scales.

Experimental setup

Datasets: COCO ImageNet

Benchmarks: COCO ImageNet

Metrics: FID IS CLIP score aesthetic score precision recall

Baselines: classifier-free guidance Lumina-mGPT without guidance NOVA without guidance RandAR without guidance average-feature weak branch layer-pruning weak branch

Builds on: classifier-free guidance contrastive decoding discrete Fourier transform

Closest papers by meaning

• NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · similarity 0.87)
• Toward Guidance-Free AR Visual Generation via Condition Contrastive Alignment (ICLR 2025 · similarity 0.87)
• Continuous Visual Autoregressive Generation via Score Maximization (ICML 2025 · similarity 0.86)
• Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.86)
• FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.86)
• Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities (ICML 2025 · similarity 0.86)