Clarification on experimental setup for DetailCLIP-style comparison in NeurIPS 2025 rebuttal

[showstarpro]

Hi SuperCLIP team,

Thank you for the insightful work and for adding additional comparisons in the NeurIPS 2025 rebuttal. I’m particularly interested in the table where SuperCLIP is compared with SLIP, MaskCLIP, A-CLIP, and DetailCLIP under the DetailCLIP-style pretraining setup (15M samples, batch size 4K, 25 epochs).
To better understand and potentially reproduce these results, I would like to ask for clarification on a few experimental details regarding this table:

Training data

Is the 15M-sample pretraining set based on YFCC15M, or another subset / filtering of YFCC?

Were the same images and captions used across all compared methods?

Optimization details

What learning rate (and schedule) was used for SuperCLIP in this setting （1e-3 or 3e-3）?

Were other hyperparameters (optimizer type, weight decay, warmup strategy) aligned with DetailCLIP’s official setup?

These details would be very helpful for ensuring a fair understanding of the reported gains and for future comparisons. Thanks again for the clear rebuttal and strong results.

Best regards

[showstarpro]

Following up on the previous issue regarding the DetailCLIP-style comparison in your NeurIPS 2025 rebuttal, we tried to reproduce the zero-shot ImageNet-1K results using the following settings:

｜Dataset | Model | Batch_size | lr | Epochs | Warmup | eps | wd｜
｜YFCC15M | ViT-B | 4096 | 5e-4 | 25 | 1 epoch | 1e-8 | 0.5｜

With these hyperparameters, our reproduced zero-shot accuracy on ImageNet-1K is 37.7, which is significantly lower than the 44.8 reported in the rebuttal table.

Could you please clarify which parts of the experimental setup might differ from ours? Specifically:

Any differences in data preprocessing / filtering for the 15M samples

Learning rate schedule or optimizer details

Any additional tricks applied during pretraining or evaluation (e.g., data augmentations, gradient clipping, normalization)

We’d appreciate guidance so we can better understand and fairly compare to your results.

Thanks in advance for your help!