Research map / Architectures and efficiency / Quantization

PermuQuant: Lowering Per-Group Quantization Error by Reordering Channels for Diffusion Models

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 PermuQuant, a PTQ method for low-bit diffusion models that reorders channels by a joint activation-weight second-moment criterion before per-group quantization, accepting reorderings only when they reduce calibration error.

Problem

Extremely low-bit (W3A3, W4A4) post-training quantization of diffusion transformers suffers severe quality loss, and per-group quantization error depends on how channels are grouped into contiguous blocks sharing a scale.

Approach

Channel order matters under per-group quantization; sorting channels by a joint activation/weight second moment before grouping places similarly scaled channels together, with a calibration-based acceptance rule and offline or normalization-fused permutations so no runtime permutation is needed.

Key results

• On Z-Image-Turbo under W3A3, PermuQuant-H reduces FID from 86.5 (ConvRot) to 80.4 and improves GenEval overall from 0.81 to 0.83.
• On SANA-1.5-1.6B, PermuQuant-H reduces FID from 98.1 to 68.0 and raises ImageReward from 0.440 to 0.919 versus ConvRot.
• On FLUX.1-dev with an RTX 5090, W4A4 NVFP4 PermuQuant gives 3.5x DiT memory reduction and up to 6.3x E2E latency speedup over BF16, largely from avoiding CPU offloading.

Surprising findings

• Much of the reported E2E speedup over BF16 comes from quantization eliminating CPU offloading rather than from low-bit arithmetic alone.
• Plain reordering without Hadamard transform sometimes gives worse FID than ConvRot on FLUX.1-dev and Z-Image-Turbo (e.g., PermuQuant FID 134.9 vs ConvRot 86.5 on Z-Image-Turbo).

Limitations

• The reordering strategy is heuristic and not guaranteed globally optimal.
• Depends on calibration data representativeness for channel statistics and acceptance decisions.
• Attention quantization is not addressed; only linear layers are handled.

Experimental setup

Datasets: COCO Captions

Benchmarks: MJHQ-30K GenEval

Metrics: FID KID ImageReward GenEval overall

Baselines: RTN SmoothQuant OmniQuant SVDQuant ConvRot

Builds on: SmoothQuant OmniQuant QuaRot ConvRot SVDQuant

Closest papers by meaning

• DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
• Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.88)
• EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.88)
• ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.88)
• PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.88)
• TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.88)