VARQ: Training-Free KV-Cache Quantization for Visual Autoregressive Generation

VARQ is a training-free, plug-in KV-cache quantizer for visual autoregressive (VAR) generators. It compresses the key/value cache to 2–8 bits while preserving fidelity far better than prior post-hoc quantizers (KIVI, FlexGen/GPTQ), across class-conditional image, text-to-image, and autoregressive video models.

model families
VAR · Infinity · InfinityStar · Self-Forcing · LongLive

2–8 bit

KV-cache precision
aggressive low-bit compression

up to +16 dB

PSNR over the best prior quantizer
on VAR image models at 3–4 bit

Qualitative Comparison

Same prompt / class, same bit-width — only the KV-cache quantizer differs. Pick a model and precision to compare Baseline (BF16), prior methods, and VARQ (Ours) side by side. The VAR image models at 3–4 bit show the largest gap: VARQ stays faithful to the baseline while KIVI and FlexGen collapse.

Model

KV-cache bits

Why VARQ

Visual autoregressive models spend most of their memory and bandwidth on the KV cache. VARQ makes that cache cheap without retraining and without a quality collapse at low bit-widths.

Training-Free

A drop-in quantizer — no fine-tuning, no calibration on the target model required.

Low-Bit KV Cache

Compresses keys and values down to 2–8 bits, shrinking cache memory and bandwidth.

Fidelity Preserving

Substantially higher PSNR than KIVI and FlexGen at the same precision, avoiding artifacts.

Broadly Applicable

One method across class-conditional, text-to-image, and streaming video AR generators.

Code

Coming Soon

Source code and quantization kernels will be released.

Citation

BibTeX

@article{xu2026varq,
  title={VARQ: Training-Free KV-Cache Quantization for Visual Autoregressive Generation},
  author={Xu, Boxun and Lu, Jiaji and Yin, Yuxuan and Wang, Zihu and Liu, Ziyue and Wang, Yu and Liu, Zirui and Li, Peng},
  journal={Under review},
  year={2026}
}