Research map / Architectures and efficiency / Quantization

QuantDemoire: Quantization with Outlier Aware for Image Demoiréing

NeurIPS 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 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 QuantDemoire, a post-training quantization framework for image demoiréing that combines a sampling-based outlier-aware quantizer with FP16 weight outliers and a frequency-aware calibration that emphasizes low/mid-frequency components.

Problem

Full-precision image demoiréing networks are too costly for edge devices, and existing PTQ methods degrade severely on demoiréing models at low bit-widths due to activation/weight outliers and banding artifacts in smooth regions.

Approach

Estimate activation quantization ranges and SmoothQuant factors from randomly sampled subsets to ignore outliers, keep the top/bottom 0.5% of weights in FP16, and calibrate quantizer bounds on frequency-decomposed (low/mid-frequency, via iterated dilated binomial kernel) outputs to reduce banding.

Key results

• At W4A4 on UHDM, QuantDemoire reaches 21.08 dB PSNR versus 17.07 dB for the best prior baseline (2DQuant), a gain of over 4 dB.
• At W4A4, QuantDemoire keeps 0.79M params and 1.80G ops (86.6% and 86.7% reductions) relative to ESDNet FP32.
• At W3A3 on UHDM, QuantDemoire achieves 19.12 dB PSNR versus 14.83 dB for SVDQuant.

Surprising findings

• Discarding weight outliers instead of storing them in FP16 drops UHDM W4A4 PSNR from 20.52 to 16.29 dB.
• Using only high-frequency components in calibration reduces PSNR (20.62 dB) below the no-calibration baseline (20.92 dB).

Limitations

• Only one backbone (ESDNet) is evaluated.
• Quantizer parameters are calibrated on a small set (200 pairs per dataset) and only activation bounds are optimized during calibration.
• Ablations are reported only on UHDM at W4A4.

Experimental setup

Datasets: UHDM FHDMi LCDMoiré

Benchmarks: UHDM FHDMi LCDMoiré

Metrics: PSNR SSIM LPIPS Params Ops

Baselines: MinMax Percentile 2DQuant SVDQuant

Builds on: SmoothQuant ESDNet LSQ SVDQuant percentile quantization

Closest papers by meaning

• SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.89)
• Outlier-Aware Post-Training Quantization for Discrete Graph Diffusion Models (ICML 2025 · similarity 0.87)
• QuantSR: Accurate Low-bit Quantization for Efficient Image Super-Resolution (NeurIPS 2023 · similarity 0.87)
• DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.86)
• Outlier-aware Slicing for Post-Training Quantization in Vision Transformer (ICML 2024 · similarity 0.86)
• Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.86)