Recent studies on Visual Autoregressive (VAR) models have highlighted that high-frequency components, or later steps, in the generation process contribute disproportionately to inference latency. However, the underlying computational redundancy involved in these steps has yet to be thoroughly investigated. In this paper, we conduct an in-depth analysis of the VAR inference process and identify two primary sources of inefficiency: step redundancy and unconditional branch redundancy. To address step redundancy, we propose an automatic step-skipping strategy that selectively omits unnecessary generation steps to improve efficiency. For unconditional branch redundancy, we observe that the information gap between the conditional and unconditional branches is minimal. Leveraging this insight, we introduce unconditional branch replacement, a technique that bypasses the unconditional branch to reduce computational cost. Notably, we observe that the effectiveness of acceleration strategies varies significantly across different samples. Motivated by this, we propose SkipVAR, a sample-adaptive framework that leverages frequency information to dynamically select the most suitable acceleration strategy for each instance. To systematically evaluate the importance of high-frequency information, we further introduce a suite of high-variation benchmark datasets that expose models' sensitivity to fine details. Extensive experiments demonstrate that SkipVAR achieves an average SSIM above 0.88 while providing a (1.81\times) overall acceleration, and achieves up to (2.62\times) speedup while maintaining model performance on the GenEval benchmark, by applying our acceleration strategies alone. These results demonstrate the effectiveness of frequency-aware, training-free adaptive acceleration for scalable autoregressive image generation.
Our project SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping has been officially released. You can explore the code and preprint through the following links:
- 📂 GitHub Repository – Source code and implementation details
- 📄 arXiv Preprint – Paper published on arXiv
- Welcome to our website! For more display effects, please visit our website: Click here to visit our website
-
Pretrained Decision Models
We provide pretrained decision models under the./SkipVARdirectory. These models are trained on a 2B model with 3k samples, using an input downsampled to a size of (8,8), and adopting the 9th step as the decision step. Please refer to./SkipVAR/readme.mdfor instructions on how to use the appropriate model based on your target SSIM threshold. -
One-Line Image Generation
We have added a new generation functiongen_one_img_SkipVAR, which is now used across all generation scripts. Simply runinfer.shoreval.shto see the accelerated generation in action — no further modification is needed.
- Infinity uses FlexAttention to speedup training, which requires
torch>=2.5.1. - Install other pip packages via
pip3 install -r requirements.txt. - Download weights from huggingface. Besides vae & transformers weights on https://huggingface.co/FoundationVision/infinity, you should also download flan-t5-xl.
from transformers import T5Tokenizer, T5ForConditionalGeneration
tokenizer = T5Tokenizer.from_pretrained("google/flan-t5-xl")
model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-xl")These three lines will download flan-t5-xl to your ~/.cache/huggingface directory.
Note: To successfully run our work following the original Infinity method, simply download the additional packages already listed in requirements.txt.
We provide eval.sh for evaluation on various benchmarks with only one command. In particular, eval.sh supports evaluation on commonly used metrics such as GenEval, ImageReward, HPSv2.1, FID and Validation Loss. Please refer to evaluation/README.md for more details.
bash scripts/eval.shThis project is a minimal modification of Infinity, which provides the original VAR-based text-to-image generation framework. All original code remains under the MIT License.
Please cite us if our work is useful for your research.
@misc{li2025skipvaracceleratingvisualautoregressive,
title={SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping},
author={Jiajun Li and Yue Ma and Xinyu Zhang and Qingyan Wei and Songhua Liu and Linfeng Zhang},
year={2025},
eprint={2506.08908},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2506.08908},
}If you have any questions, feel free to approach me at fakerone.li@foxmail.com