CVPR 2026 Compute Reporting Form - Author Guidelines

The Compute Reporting Form (CRF) is a pilot initiative designed to increase transparency around computational resources in computer vision research. By collecting data on GPU/CPU usage, training time, and efficiency metrics, we aim to understand the computational landscape of our research - from resource requirements to efficiency innovations. This reporting will not affect your paper's acceptance decision - the data is reviewed only by a separate Compute Reporting Committee and will not be shared with reviewers, Area Chairs, or Senior Area Chairs.

Submitting the CRF is mandatory. The CRF is due with your supplementary materials. Section 1 (Hardware Specifications) and Section 5 (Verification and Submission) are required and take approximately 5 minutes to complete, covering basic information about your hardware infrastructure. Sections 2-4 are optional but encouraged, allowing you to report compute costs, performance metrics, development history, and anonymized training logs. Papers that complete Sections 1-3 & 5 are eligible for recognition awards, while only papers that complete all sections are eligible for all Compute Reporting recognition awards.

 

Recognition Program

Papers that complete sections of the CRF may receive recognition for computational excellence. The "Efficient CVPR" Badge and "CVPR Compute Gold Star" awards recognize strong efficiency within task categories and outstanding efficiency achievements, and are available to papers that complete Sections 1-3. The "CVPR Compute Transparency Champion" award recognizes high-quality anonymized W&B log submissions and requires completing all sections including Section 4. Awards are determined by the Compute Reporting Committee using objective metrics (such as compute per performance improvement) along with other criteria as determined by the committee, with comparisons made within similar task categories to ensure fair evaluation. Recognized papers will receive special presentation highlights during the conference and featured placement on the conference website. Award results will be announced during the conference.

 

To Complete the Form

Review the prefilled example (CVPR26 CRF prefilled example.pdf) to see how to report each section with realistic values. Fill out the CRF (CVPR26 Compute Reporting Form.pdf) with your paper's computational details. Section 1 (Hardware Specifications) is mandatory. Section 2 (Task and Compute Reporting) and subsequent sections are optional but strongly encouraged for authors seeking recognition awards. If you cannot or prefer not to report computational data, you may provide an explanation in Section 5.2 (Additional Comments). Optionally, you may submit anonymized Weights & Biases logs using the provided anonymization tool. Submit the completed form and any anonymized logs through the supplementary materials submission system by the deadline. For detailed instructions, including W&B log anonymization guidance, consult the full CRF form and accompanying FAQ.

Note on heterogeneous environments: If you used multiple GPU types or varied hardware configurations, report the primary or most-used configuration in Section 1. You can describe the full range of hardware in the Additional Comments section (5.2).

Important: Do not flatten or manipulate the PDF - save the original PDF file after filling it out. Flattening (such as using "Print to PDF" in macOS Preview app) removes the form fields needed for automated data processing.

 

Data Protection and Privacy

The Compute Reporting initiative follows strict data protection standards. CRF data is processed independently from the paper review process and is never shared with reviewers. All analyses use anonymized data, and published results are appropriately aggregated. For complete details on how your data is collected, used, stored, and protected, please review the Data Protection and Retention Policy.

 

CVPR 2026 Compute Reporting Form - FAQ

Contents

1. Why are we collecting compute information?
2. Will my compute reporting affect my paper's acceptance decision?
3. How is my CRF data accessed and protected?
4. How will the CRF data be used?
5. What if my research used a lot of computational resources?
6. Is there an example of a completed form I can reference?
7. How should I save and submit the completed CRF?
8. What sections are required vs. optional?
9. My compute environment is heterogeneous (multiple GPU types, different machines). How should I fill out Section 1?
10. I don't have exact numbers for everything. What should I do?
11. What's the difference between reporting just my final model vs. total development compute?
12. Why are you asking for Weights & Biases logs, and do I have to submit them?
13. What are the recognition awards, and how are they determined?
14. How long will this form take to complete?
15. What if my paper doesn't fit neatly into your categories?
16. Who can I contact if I have questions or concerns?

 

Q. Why are we collecting compute information?

A. The Compute Reporting Form (CRF) is a pilot initiative aimed at increasing transparency around computational resources in computer vision research. By collecting this data, we hope to:

• Help the community understand the true computational costs of different approaches
• Provide benchmarking data that can guide future research directions
• Recognize papers that achieve strong results efficiently
• Promote more sustainable and accessible AI research practices

This is an experiment in transparency, not a judgment mechanism.

Q. Will my compute reporting affect my paper's acceptance decision?

A. No, absolutely not. Your compute reporting data will NOT be shared with reviewers, Area Chairs, or Senior Area Chairs, and will NOT influence the peer review or acceptance process in any way. Papers are evaluated solely on their scientific merit. The Compute Reporting Committee is completely separate from the review process. For complete information on how your data is protected and used, see our Data Protection and Retention Policy.

Q. How is my CRF data accessed and protected?

A. CRF data is processed by the Compute Reporting Committee, which operates independently from the paper review process. Your data is never shared with reviewers, Area Chairs, or Senior Area Chairs. All analyses use anonymized data, and published results are appropriately aggregated to prevent identification of individual submissions. Links between your CRF submission and your paper identity are permanently deleted 6 months after the conference. For complete details on data collection, use, retention, and security measures, please review our Data Protection and Retention Policy.

Q. How will the CRF data be used?

A. The CRF data will be analyzed to understand computational trends or disparities in computer vision research and may inform policy decisions. Key uses include: identifying relationships between computational resources and research outcomes, benchmarking requirements across task categories to reveal barriers to entry, recognizing efficiency innovations, tracking temporal trends in accessibility, estimating environmental impact, establishing reproducibility standards, and understanding how computational advantages affect research velocity. These insights will inform infrastructure initiatives (like NSF's NAIRR, OECD GPAI, and EU programs), promote algorithmic efficiency research, and support more equitable access to computational resources. All analyses use anonymized data with appropriate aggregation.

Q. What if my research used a lot of computational resources?

A. Using significant computational resources is not penalized in any way. Many important advances require substantial compute. We're interested in understanding the relationship between compute and performance gains across different approaches. High-compute papers can still be recognized for their contributions and may provide valuable insights about what's possible at different resource scales.

Q. Is there an example of a completed form I can reference?

A. Yes! We've provided a prefilled example form showing how to complete each section. You can download it here: CVPR26 CRF prefilled example.pdf

This example walks through a sample submission with realistic values and shows how to:

• Report hardware specifications
• Calculate efficiency metrics
• Provide context for your computational choices
• Fill out optional sections

Reviewing this example before starting your own submission can save time and clarify any questions about what information to provide.

Q. How should I save and submit the completed CRF?

A. Important: Save the original PDF file directly after filling it out. Do not flatten, print to PDF, or otherwise manipulate the form, as this removes the interactive form fields needed for automated data processing.

Common mistakes to avoid:

• Using "Print to PDF" or "Export as PDF" in macOS Preview
• Flattening the PDF in Adobe Acrobat
• Converting to another format and back to PDF

Correct approach: Simply fill out the form fields in your PDF reader and save the file using "Save" or "Save As" - this preserves the form data. Submit this saved PDF through the supplementary materials submission system.

Q. What sections are required vs. optional?

A. Only Section 1 (Hardware Specifications) and Section 5 (Verification and Submission) are mandatory. These sections take approximately 5 minutes to complete and cover basic information about your hardware infrastructure, as well as verify your submission.

Section 2 (Task and Compute Reporting) and all subsequent sections are optional. However, papers that complete Sections 1-3 may be eligible for some recognition awards (such as the "Efficient CVPR" Badge and "CVPR Compute Gold Star"), while completing all sections (including Section 4 with W&B logs) is required for eligibility for all awards, including the "CVPR Compute Transparency Champion" award. If you cannot or prefer not to report computational data in the optional sections, you may provide an explanation in Section 5.2 (Additional Comments).

Q. My compute environment is heterogeneous (multiple GPU types, different machines). How should I fill out Section 1?

A. For heterogeneous environments, report the primary or most-used hardware configuration in Section 1. You can provide additional details about your heterogeneous setup in one of these ways:

• Use the "Additional Comments" section (5.2) to describe the full range of hardware used
• For Section 1 dropdown fields, select the predominant hardware type and note "mixed environment - see comments" in any available text fields
• If using multiple distinct configurations equally, report the one used for your main results

The goal is to capture a representative profile of your computational infrastructure. Perfect precision across all hardware variations is not required.

Q. I don't have exact numbers for everything. What should I do?

A. We understand that tracking every computational detail isn't always feasible. For fields you choose to complete, provide your best estimates based on:

• Cloud provider bills or cluster logs
• Training time multiplied by number of GPUs used
• Records from your experiment tracking tools

Remember that only Section 1 is required. Optional sections can be left incomplete if you don't have the data. The goal is to capture a reasonable computational profile of your work, not to achieve perfect precision.

Q. What's the difference between reporting just my final model vs. total development compute?

A. Section 2 (Optional) asks for the compute needed to reproduce your paper's main result - training and evaluating your final model versus your strongest baseline. This is what readers would need to replicate your work.

Section 3 (Optional) asks for total development compute, including all experiments, failed attempts, and iterations. This provides insight into the full research process but is entirely optional.

Q. Why are you asking for Weights & Biases logs, and do I have to submit them?

A. W&B log submission is completely optional. Anonymized logs provide more precise efficiency calculations based on actual GPU utilization patterns, which can strengthen your eligibility for recognition awards. We provide an anonymization tool to protect your privacy. Many researchers find it valuable to contribute this data, but there's no penalty for not submitting logs.

Q. What are the recognition awards, and how are they determined?

A. Papers that complete sections of the CRF may receive recognition including:

• "Efficient CVPR" Badge - for strong efficiency within your task category (requires completing Sections 1-3)
• "CVPR Compute Gold Star" - for outstanding efficiency achievements (requires completing Sections 1-3)
• "CVPR Compute Transparency Champion" - for high-quality W&B log submissions (requires completing all sections including Section 4)

Awards are determined by the Compute Reporting Committee (to be nominated by the Program and General Chairs) using objective metrics (such as compute per performance improvement) along with other criteria as determined by the committee, with comparisons made within similar task categories to ensure fair comparison. This is about celebrating transparency and efficiency, not creating a ranking system.

Q. How long will this form take to complete?

A. Section 1 (Hardware Specifications) and Section 5 (Verification and Submission), which are mandatory, take approximately 5 minutes to complete. If you choose to complete the optional sections:

• Section 2 (Task and Compute Reporting) adds approximately 10-15 minutes
• Section 3 (Additional Computational Details) adds additional time if you choose to complete it
• Section 4 (W&B Logs) requires time for anonymization if you choose to submit logs

You can complete as much or as little beyond Section 1 as you wish. Papers completing Sections 1-3 are eligible for some recognition awards, while completing all sections (including Section 4) is required for eligibility for all awards.

Q. What if my paper doesn't fit neatly into your categories?

A. We've tried to accommodate diverse research areas, but we know computer vision research is varied. If your work doesn't fit standard categories:

• Use "Other" options and provide brief descriptions in text fields
• Focus on reporting what's most relevant to your specific contribution
• Add context in the "Additional Comments" section (5.2)

The goal is to capture meaningful information about your work, not to force it into rigid categories.

Q. Who can I contact if I have questions or concerns?

A. For questions about compute reporting: crf_cvpr2026@computer.org

For W&B anonymization technical issues: Visit the anonymization tool repository

For data protection and privacy questions: Review our Data Protection and Retention Policy or contact crf_cvpr2026@computer.org

We're here to help make this process as smooth as possible. As a pilot initiative, your feedback will help us improve this program for future conferences.