Hello organizers, I would like to clarify the runtime requirement for Task 3. According to the FAQ, the total runtime for each task should not exceed 1 hour. For Task 3, we understood that there are 300 cases and up to 4 target field-strength generations per input image (excluding the source field strength), which would result in approximately 1200 inference runs. Given the image size (364 × 436 × 364), this seems to imply only a few seconds per volume including I/O and preprocessing/postprocessing, which feels extremely challenging for full-volume 3D models. Could you clarify the expected runtime requirement per case/volume, and whether the test phase will use cropped regions (similar to validation) or full-volume scans? Also, does the runtime budget include file I/O and preprocessing/postprocessing? We want to make sure we correctly understand the intended evaluation setting. Thank you.

Created by 창선 박 changseon
Thank you for your question. For the test phase, participants are expected to perform inference on the entire volume (all slices). No cropping will be applied during test evaluation. Please also note that the runtime limit has been updated to 20 hours per team for each task, corresponding to an average of 1 hour per case for the 20 test cases. This runtime includes the complete inference pipeline. We hope this clarifies the evaluation setting. Best regards, MRIxFields2026 Organizing Committee
Hi Jiahao, Thank you for the clarification on the central axial slab submission. I would like to confirm my understanding of the runtime requirement. If Task 3 has a 1 hour limit for 300 cases, this would mean about 12 seconds per case. With up to 4 target field strengths per case, this would be approx. 3 seconds per source-to-target generation for the 30-slice slab. Could you please confirm if this is correct? And does the 1 hour limit include data loading, preprocessing, inference, postprocessing, and saving outputs? For diffusion-based methods, inference can be slower because of the iterative sampling, so this seems like a tight runtime budget. Do you have any recommendations or best practices for meeting the runtime requirement? Thank you!
Dear Changseon, Thank you for your questions. In the **validation state**, for each case, the required submission is the central axial slab only: z in [150, 180), i.e. the (364, 436, 30) region instead of the full 364-slice volume. The amount of data to generate and submit per case is therefore much smaller than the full 364x436x364 grid. In the **testing state**, for each case, inference will be conducted in full 364-slice volume, i.e. the (364, 436, 364). Please feel free to reach out if anything remains unclear. Thanks, Jiahao

Clarification on Task 3 Test Phase (Runtime / Full Volume vs Crop) page is loading…