TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation

A Test-Time Scaling Framework for Visual Auto-Regressive Generation[Paper]

Zhekai Chen, Ruihang Chu, Yukang Chen, Shiwei Zhang, Yujie Wei, Yingya Zhang, Xihui Liu

Scaling visual generation models is essential for real-world content creation, yet requires substantial training and computational expenses. Alternatively, test-time scaling has garnered growing attention due to resource efficiency and promising performance. In this work, we present TTS-VAR, the first general test-time scaling framework for visual auto-regressive (VAR) models, modeling the generation process as a path searching problem.

To dynamically balance computational efficiency with exploration capacity, we first introduce an adaptive descending batch size schedule throughout the causal generation process.Besides, inspired by VAR's hierarchical coarse-to-fine multi-scale generation, our framework integrates two key components: (i) At coarse scales, we observe that generated tokens are hard for evaluation, possibly leading to erroneous acceptance of inferior samples or rejection of superior samples. Noticing that the coarse scales contain sufficient structural information, we propose clustering-based diversity search. It preserves structural variety through semantic feature clustering, enabling later selection on samples with higher potential. (ii) In fine scales, resampling-based potential selection prioritizes promising candidates using potential scores, which are defined as reward functions incorporating multi-scale generation history.

Experiments on the powerful VAR model Infinity show a notable 8.7% GenEval score improvement (0.69→0.75). Key insights reveal that early-stage structural features effectively influence final quality, and resampling efficacy varies across generation scales.

News

[2025.9.19] Accepted by NeurIPS 2025!

• [2025.7.24] We release Paper and Github repo TTS-VAR.

Quick Start

1. Prepare the environment

   cd TTS-VAR
   conda create -n tts-var python==3.10.12
   conda activate tts-var
   

2. Clone submodule Infinity

   git submodule init
   git submodule update
   

3. Install dependencies

   cd Infinity
   pip install torch==2.5.1
   pip install -r requirements.txt
   
   cd ..
   pip install -r requirements.txt
   

4. Download the pretrained models and put it in ./pretrained_models/

   git-lfs install
   bash download_models.bash
   

5. Inference with TTS-VAR

   bash run.bash
   

Inference Args

The inference arguments are list in run.bash, with specific explanation in tts-var/main.py. Here we show arguments for TTS-VAR. For Infinity's arguments, please refer to Infinity.

Process Arguements

Argument	Description	Default	Explanation
--reward_type	Type of reward model to use for reward function.	ir	ImageReward
--cal_type	Calculation type for reward-based resampling.	value
--resample_sis	List of step indices for resampling during inference.	[6,9]
--extract_type	Feature extraction type for reward guidance.	pca	PCA from DINOv2 features
--extract_sis	List of step indices for feature clustering during inference.	[2,5]
--bs_sis	List of batch sizes for different scales of generation.	[8, 8, 6, 6, 6, 4, 2, 2, 2, 1, 1, 1, 1]	For 13 scales
--lam	Lambda value for potential weights in resampling.	10.0

Path Arguements

Argument	Description	Default
--aesthetic_predictor_path	Path to aesthetic predictor model when using aes reward.	./pretrained_models/aesthetic_predictor_v2_5.pth
--siglip_encoder_path	Path to SigLIP encoder model when using aes reward.	./pretrained_models/siglip-so400m-patch14-384
--image_reward_model_path	Path to ImageReward model when using ir reward.	./pretrained_models/ImageReward.pt
--dinov2_hub_path	Github or locol repo of DINOv2	facebookresearch/dinov2

Acknowledgement

This code is built on top of Infinity. We thank the authors for their great work.

Citation

If you find this code useful for your research, please cite our paper:

@article{chen2025tts,
  title={TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation},
  author={Chen, Zhekai and Chu, Ruihang and Chen, Yukang and Zhang, Shiwei and Wei, Yujie and Zhang, Yingya and Liu, Xihui},
  journal={arXiv preprint arXiv:2507.18537},
  year={2025}
}