W4A4 Quantization for Inference on Wan2.2-I2V-A14B

Chengyu Shi; Jiahao Liu; Yidong Chen

arxiv: 2606.29337 · v1 · pith:YRZKUDBU · submitted 2026-06-28 · cs.CV · cs.DC

W4A4 Quantization for Inference on Wan2.2-I2V-A14B

Yidong Chen , Chengyu Shi , Jiahao Liu This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-30 07:25 UTCgrok-4.3pith:YRZKUDBU record.json open to challenge →

classification cs.CV cs.DC

keywords W4A4 quantizationvideo generation modelactivation smoothingmixed precisioninference efficiencyVBench metricsWan2.2-I2V-A14B

0 comments

The pith

A W4A4 quantization pipeline for Wan2.2-I2V-A14B stays within 2-3.5 percent of FP16 quality on VBench metrics.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows how to quantize a large image-to-video model to 4-bit weights and activations while keeping output quality close to the original 16-bit version. The method uses calibration to find channels with extreme activation values, then smooths their range and routes outliers through a higher-precision path in matrix multiplications. On standard video quality tests, most metrics drop only a small amount and motion smoothness even improves. A basic 4-bit approach without these steps loses about 5 percent across the board. The work demonstrates that existing ideas from language model quantization can transfer to video generation models for efficient inference.

Core claim

Adapting mixed-precision handling of activation outliers via dual-branch GEMM and per-channel smoothing, combined with block-wise HiF4 packing, allows the 4-bit pipeline to match FP16 performance within 2-3.5 percent on most VBench I2V axes and to improve motion smoothness, while a standard HiFloat4 implementation falls about 5 percent behind FP16 on all scores.

What carries the argument

Dual-branch GEMM that keeps outlier columns in higher precision while applying W4A4 to the rest, after per-channel smoothing of heavy-tailed activations identified by calibration.

If this is right

The quantized model supports inference with lower memory and compute demands.
Motion smoothness can improve rather than degrade under this quantization scheme.
The approach requires only calibration data and no additional model training.
Outperforms a direct application of HiFloat4 without the mixed-precision and smoothing steps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This indicates that activation outlier management is critical for maintaining quality in quantized video diffusion models.
Similar calibration-based smoothing might reduce the quality gap in other generative tasks like text-to-video.
Deployment on edge devices could become feasible if the memory savings scale with model size.

Load-bearing premise

That calibration batches from OpenS2V-5M capture the activation statistics well enough for the target model without needing retraining or more data.

What would settle it

Evaluating the quantized model on a new set of video prompts outside the calibration distribution and checking whether the quality degradation exceeds 3.5 percent on VBench metrics.

Figures

Figures reproduced from arXiv: 2606.29337 by Chengyu Shi, Jiahao Liu, Yidong Chen.

**Figure 4.** Figure 4: Schematic per-channel maxima before/after SmoothQuant fold [PITH_FULL_IMAGE:figures/full_fig_p002_4.png] view at source ↗

**Figure 8.** Figure 8: Block layouts explored for HiF4/MXFP4 weight packing after [PITH_FULL_IMAGE:figures/full_fig_p003_8.png] view at source ↗

**Figure 7.** Figure 7: Sorted per-channel scorej on an FFN projection: a few dominant columns justify a MixQ-style split. branches see reduced activation tails. Deeper blocks exhibit similar but attenuated tails ( [PITH_FULL_IMAGE:figures/full_fig_p003_7.png] view at source ↗

read the original abstract

We summarize our submission to Sub-Challenge 1: W4A4 Quantization for Inference (HiF4 / MXFP4) of the ICME 2026 Low-Bit-width Large-Model Quantization Challenge. The sub-challenge targets 4-bit weight and 4-bit activation inference on Wan-AI/Wan2.2-I2V-A14B under HiF4 or MXFP4 numerical formats. We adapt two complementary ideas from LLM quantization, MixQ-style mixed precision for sparse activation outliers and SmoothQuant-style per-channel smoothing, together with block-wise HiF4 packing for Wan2.2 feed-forward linear layers. Calibration on representative OpenS2V-5M batches identifies heavy-tailed activation channels; smoothing rebalances dynamic range before W4A4 rounding; and a dual-branch GEMM preserves outlier columns in higher precision while the bulk of channels use strict W4A4. On official VBench I2V metrics, our pipeline stays within 2-3.5 percent of FP16 on most quality axes and improves motion smoothness, outperforming a native HiFloat4 baseline that degrades roughly 5 percent relative to FP16 across all reported scores.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper applies MixQ and SmoothQuant to Wan2.2-I2V-A14B with block packing but supplies almost no implementation details or validation.

read the letter

The main takeaway is that the authors combine two established LLM quantization ideas—MixQ mixed precision for outliers and SmoothQuant per-channel smoothing—plus block-wise HiF4 packing, then calibrate on OpenS2V-5M batches to run Wan2.2-I2V-A14B at W4A4. They report VBench I2V scores within 2-3.5% of FP16 on most axes and better motion smoothness than a native HiFloat4 baseline.

The work does a clean job of targeting a real 14B video model and stating the high-level pipeline without claiming new primitives. The dual-branch GEMM for keeping outlier columns in higher precision is a practical choice for the feed-forward layers.

The soft spots are the missing substance. The text gives no description of the exact outlier threshold, how the smoothing scales are computed, what the block packing looks like in code, or any ablation showing that the heavy-tailed channels found on OpenS2V-5M match those arising during actual I2V generation. The stress-test point about possible distribution shift is not addressed, and there are no error bars, multiple seeds, or cross-model checks. Without those, the closeness to FP16 cannot be evaluated.

This paper is for engineers who need to deploy exactly this model under tight memory limits. A reader already working on similar video diffusion quantization might extract the high-level recipe, but the lack of detail limits reuse.

I would send it to peer review so the authors can add the implementation steps and any validation that would let others judge whether the calibration transfers.

Referee Report

2 major / 1 minor

Summary. The manuscript describes a W4A4 quantization approach for the Wan2.2-I2V-A14B image-to-video model submitted to the ICME 2026 challenge. It adapts MixQ-style mixed precision for outliers and SmoothQuant-style per-channel smoothing, with calibration on OpenS2V-5M batches to identify heavy-tailed channels, followed by smoothing and dual-branch GEMM for inference. The key empirical claim is that on VBench I2V metrics, the quantized model stays within 2-3.5% of FP16 performance on most axes, improves motion smoothness, and outperforms a native HiFloat4 baseline which degrades by ~5%.

Significance. If the reported performance holds under scrutiny, the work would provide a practical demonstration of effective 4-bit weight and activation quantization for a large-scale video generation model without retraining, which could have significant implications for deploying such models efficiently. The adaptation of LLM quantization techniques to this domain is a reasonable extension, but the absence of detailed methodology, ablations, or statistical validation limits the immediate impact.

major comments (2)

[Abstract] Abstract: The central performance claims (within 2-3.5% of FP16 on VBench metrics) are presented without any implementation details, error bars, statistical tests, ablation results, or description of the exact calibration procedure, making it impossible to evaluate or reproduce the results.
[Abstract] Abstract: The pipeline relies on the assumption that heavy-tailed activation channels identified via calibration on OpenS2V-5M batches are representative of those in Wan2.2-I2V-A14B during I2V generation; no cross-model validation or ablation is provided to support transfer of these statistics, which directly bears on whether the reported closeness to FP16 can be achieved.

minor comments (1)

[Abstract] Abstract: The term 'native HiFloat4 baseline' is not defined; clarify what this baseline entails and how it differs from the proposed pipeline.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our ICME 2026 challenge submission. This work is presented as a concise summary due to strict page limits, but we address the concerns about missing details and assumptions below. We clarify aspects of the methodology while noting constraints of the format.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims (within 2-3.5% of FP16 on VBench metrics) are presented without any implementation details, error bars, statistical tests, ablation results, or description of the exact calibration procedure, making it impossible to evaluate or reproduce the results.

Authors: The submission format for the challenge requires a compact abstract-style summary, which limits inclusion of error bars, statistical tests, and full ablations. The manuscript does outline the core pipeline: MixQ-style mixed precision for outliers, SmoothQuant-style per-channel smoothing, block-wise HiF4 packing, and dual-branch GEMM. The calibration procedure is described as processing OpenS2V-5M batches to detect heavy-tailed channels before applying smoothing and mixed-precision inference. We agree additional quantitative validation would improve reproducibility and can expand the description of the calibration steps in a revised version if space allows. revision: partial
Referee: [Abstract] Abstract: The pipeline relies on the assumption that heavy-tailed activation channels identified via calibration on OpenS2V-5M batches are representative of those in Wan2.2-I2V-A14B during I2V generation; no cross-model validation or ablation is provided to support transfer of these statistics, which directly bears on whether the reported closeness to FP16 can be achieved.

Authors: OpenS2V-5M was chosen as calibration data because it provides diverse, large-scale video content aligned with I2V tasks. The empirical results on VBench I2V metrics for the target Wan2.2-I2V-A14B model (within 2-3.5% of FP16 and outperforming the HiFloat4 baseline) provide indirect support for the effectiveness of the transferred statistics. We acknowledge that no explicit ablation on calibration dataset choice or cross-validation is included, which is a limitation of the concise format. The assumption is standard in post-training quantization but could be discussed more explicitly if a revision is requested. revision: no

Circularity Check

0 steps flagged

No circularity: purely empirical pipeline with external benchmarks

full rationale

The manuscript contains no equations, derivations, or predictions. It describes an empirical adaptation of prior LLM quantization techniques (MixQ, SmoothQuant) applied via calibration on OpenS2V-5M followed by direct measurement on VBench I2V metrics for Wan2.2-I2V-A14B. All reported results are comparisons against FP16 and a native HiFloat4 baseline; none reduce to fitted parameters or self-citations by construction. The central claims rest on observable performance deltas rather than any definitional or fitted-input loop.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on calibration-derived parameters and the domain assumption that activation statistics from OpenS2V-5M generalize to the target model; these are not independently verified in the abstract.

free parameters (2)

per-channel smoothing scales
Fitted during calibration to rebalance dynamic range before W4A4 rounding.
outlier identification threshold
Used to decide which activation columns receive higher-precision treatment in the dual-branch GEMM.

axioms (1)

domain assumption Heavy-tailed activation channels exist and can be reliably identified from OpenS2V-5M calibration batches
Invoked to justify the mixed-precision branch and smoothing step.

pith-pipeline@v0.9.1-grok · 5751 in / 1359 out tokens · 36482 ms · 2026-06-30T07:25:45.756546+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Wan2.2: Open large-scale video generation models,

Wan-AI, “Wan2.2: Open large-scale video generation models,” https: //github.com/Wan-Video/Wan2.2, 2025

2025
[2]

Smoothquant: Accurate and efficient post-training quantization for large language models,

Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han, “Smoothquant: Accurate and efficient post-training quantization for large language models,” inInternational Conference on Machine Learning (PMLR), 2023, pp. 38087–38099

2023
[3]

MixQ: Taming dynamic outliers in mixed-precision quantization by online prediction,

Yidong Chen, Chen Zhang, Rongchao Dong, Haoyuan Zhang, Yonghua Zhang, Zhonghua Lu, and Jidong Zhai, “MixQ: Taming dynamic outliers in mixed-precision quantization by online prediction,” inProc. Int. Conf. High Performance Computing, Networking, Storage and Analysis (SC), 2024

2024
[4]

Quantizing deep convolutional networks for efficient inference: A whitepaper

Raghuraman Krishnamoorthi, “Quantizing deep convolutional networks for efficient inference: A whitepaper,”arXiv preprint arXiv:1806.08342, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018

[1] [1]

Wan2.2: Open large-scale video generation models,

Wan-AI, “Wan2.2: Open large-scale video generation models,” https: //github.com/Wan-Video/Wan2.2, 2025

2025

[2] [2]

Smoothquant: Accurate and efficient post-training quantization for large language models,

Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han, “Smoothquant: Accurate and efficient post-training quantization for large language models,” inInternational Conference on Machine Learning (PMLR), 2023, pp. 38087–38099

2023

[3] [3]

MixQ: Taming dynamic outliers in mixed-precision quantization by online prediction,

Yidong Chen, Chen Zhang, Rongchao Dong, Haoyuan Zhang, Yonghua Zhang, Zhonghua Lu, and Jidong Zhai, “MixQ: Taming dynamic outliers in mixed-precision quantization by online prediction,” inProc. Int. Conf. High Performance Computing, Networking, Storage and Analysis (SC), 2024

2024

[4] [4]

Quantizing deep convolutional networks for efficient inference: A whitepaper

Raghuraman Krishnamoorthi, “Quantizing deep convolutional networks for efficient inference: A whitepaper,”arXiv preprint arXiv:1806.08342, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018