VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution (ICLR 2026)

📄 Paper    📝 arXiv    🏠 Project Page    💻 Code

Yixuan Zhu*, Shilin Ma*, Haolin Wang, Ao Li, Yanzhe Jing, Yansong Tang†, Lei Chen, Jiwen Lu, Jie Zhou

(* Equal contribution   † Corresponding author)

Tsinghua University

VARestorer is the official Hugging Face model repository for the ICLR 2026 paper "VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution." It distills a pre-trained text-to-image visual autoregressive (VAR) model into a single-step real-world image super-resolution system.

This Hugging Face repository includes the released checkpoint assets together with a runnable snapshot of the official codebase for convenient download and inference. The primary development home remains the official GitHub repository.

Real-World Restoration at a Glance

Street Scene	Landscape	Corgi Portrait

_{Left half: real degraded input  |  Right half: VARestorer one-step output.
Want to drag the divider yourself? → Try the interactive slider on the project page.}

1 step	0.23 s	~10× faster	27.3 M params
one-pass inference	per 512×512 image	than VAR baseline	trainable (1.2% of total)

Pipeline

Download and Use

You can either clone this Hugging Face repository directly or use the primary GitHub repository. The commands below assume the usual GitHub workflow, but the same directory layout is now mirrored here as well.

1. Clone the repository and install the dependencies:

# Option A: clone the primary GitHub repository
git clone https://github.com/EternalEvan/VARestorer.git
cd VARestorer
pip install -r requirements.txt
pip install --no-build-isolation git+https://github.com/cloneofsimo/lora.git
pip install --no-build-isolation flash_attn==2.8.3

If you prefer to clone the Hugging Face mirror instead, use:

git clone https://huggingface.co/EternalEvan/VARestorer
cd VARestorer
pip install -r requirements.txt
pip install --no-build-isolation git+https://github.com/cloneofsimo/lora.git
pip install --no-build-isolation flash_attn==2.8.3

2. Download the main checkpoint from this repository (weights) or from Google Drive:

huggingface-cli download EvanEternal/VARestorer varestorer.pth --local-dir ./weights

3. Download the additional dependencies required by the official release:

• google/flan-t5-xl into ./weights/flan-t5-xl
• lxq007/DiffBIR as ./weights/general_swinir_v1.ckpt
• FoundationVision/Infinity as ./weights/infinity_vae_d32reg.pth

4. Run inference:

bash scripts/infer.sh

For the latest updates, issue tracking, and future development, please refer to the official GitHub repository.

Links

• Paper (OpenReview)
• arXiv
• Project Page
• Code Repository

Citation

@inproceedings{zhu2026varestorer,
  title     = {VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution},
  author    = {Zhu, Yixuan and Ma, Shilin and Wang, Haolin and Li, Ao and
               Jing, Yanzhe and Tang, Yansong and Chen, Lei and Lu, Jiwen and Zhou, Jie},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://openreview.net/forum?id=T2Oihh7zN8}
}