Introduction

Test-time scaling, which has demonstrated great success in improving reasoning for LLMs (e.g., OpenAI o1/o3, DeepSeek-R1), can hold significant promise for computer vision models as well. By allocating more computational resources during inference, vision models could achieve greater accuracy, robustness, and interpretability in complex tasks ranging from perception and understanding to reasoning and decision-making. This approach could enhance performance in high-stakes domains such as medical imaging, autonomous driving, and security surveillance, where precision and interpretability are crucial. Additionally, extending test-time scaling to multimodal models and generative architectures could foster more sophisticated cross-modal reasoning and higher-quality content generation. However, applying test-time scaling to vision models presents unique challenges. Vision tasks typically involve high-dimensional inputs, making computational scaling at test time more resource-intensive. Efficient algorithms will be necessary to ensure that the increased computation does not lead to impractical processing times or energy consumption. Moreover, ensuring robustness and safety when models are subjected to increased inference computation—particularly in dynamic or adversarial environments—will be crucial.

The Workshop on Test-time Scaling for Computer Vision (ViSCALE) at CVPR2025 aims to explore the frontiers of scaling test-time computation in vision models, addressing both theoretical advancements and practical implementations. We will discuss the suitability of test-time scaling for traditional vision tasks like perception and the extensions to multimodal and generative models, towards enhancing performance in critical domains. It will also cover solutions for efficient algorithms, considerations of robustness and safety, and novel problems in computer vision posed by test-time scaling. By bringing together experts, the workshop seeks to foster collaboration and innovation in applying this paradigm to push the limits of computer vision.

Important Dates

Call for Papers

We welcome submissions related to different aspects of test-time scaling for computer vision, including but not limited to the following topics:

• Theoretical analysis of test-time scaling in computer vision.
• Suitability of test-time scaling in traditional computer vision tasks.
• Extensions of test-time scaling to multimodal foundation models and generative models.
• Efficient algorithms for vision models with test-time scaling.
• Trustworthiness issues like robustness, safety, and interpretability.
• Novel challenges posed by test-time scaling in computer vision.
• Evaluations and benchmarks on test-time scaling techniques.

Information for Submission

Format: We consider 3 types of submissions:

• Extended Abstracts (1 page excluding references)
• Short Papers (4 pages excluding references)
• Long Papers (8 pages excluding references)

Submission Site: OpenReview

Submission Deadline: March 15, 2025 (AoE)

Best Paper Award: We will select 1~3 best paper awards as well as oral presetation according to the quality of the submissions. There will be a prize upto $1000 for each awarded paper.

Speakers

Ludwig Schmidt Stanford University

Saining Xie New York University

Trevor Darrell U.C., Berkeley

Organizers

Hang Su Tsinghua University

Yinpeng Dong Tsinghua University

Yichi Zhang Tsinghua University

Jindong Gu University of Oxford

Cihang Xie U.C., Santa Cruz

Lingjuan Lyu Sony

Bolei Zhou U.C., Los Angelos

Jun Wang University College London

Jun Zhu Tsinghua University

Philip Torr University of Oxford

Shiguang Shan Chinese Academy of Sciences

Wanli Ouyang Shanghai AI Laboratory

Shuicheng Yan National University of Singapore

Sponsors

We are sincerely grateful for the supports from all our sponsors.

Contact

For any inquiries, please contact the official email: viscalecvpr@gmail.com or our organizers, Yinpeng Dong: dongyinpeng@mail.tsinghua.edu.cn and Yichi Zhang: zyc22@mails.tsinghua.edu.cn