- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents a visual backbone solution named Vision Fluxer to combine the idea of global filters in the frequency domain and adaptive/generated kernels. Although there are several methods working on adaptive/generated kernels, including AFFNet and CDFFormer, the proposed method emphasizes the importance of the coefficient of variation of the feature maps. The method is evaluated on widely used visual backbone benchmarks including ImageNet classification, COCO detection, and ADE20k segmentation.
Critical strengths
-
The idea is well-motivated and easy to follow.
-
The paper provides some interesting and in-depth analysis of the value of adaptive kernels for global filters.
Critical weaknesses
-
The paper highlights its performance (like Figure 1), and I believe superior performance is also key to supporting the value of the proposed method and claims. However, some important recent models are not compared/mentioned in the paper, and the proposed solution cannot outperform them. For example, the proposed method achieves similar performance on ImageNet to CAFormer and VSSD, which were published in 2024 and 2025, respectively. Interestingly, VSSD and the proposed method use similar training recipes and tricks like MESA. So it is still unknown whether the high performance comes from better architectural designs or better training strategies.
-
I think it is a good paper with a comprehensive experimental study and carefully designed analysis. However, I am a bit doubtful about the value of publishing a visual backbone paper nowadays with limited performance improvement and similar overall architectures.
Questions for authors
Please see my comments above.
After rebuttal: The authors add comparisons with more recent methods. However, I remain concerned that the improvements over existing methods are relatively marginal in terms of both technical value and empirical performance. Therefore, I will maintain my initial rating.
4 comments
Dear Reviewer SU9k,
Thank you again for your review. Even with self-attention firmly established, approaches such as Mamba and FNO have opened up new possibilities for model design by exploring different computational building blocks. This is also the motivation behind Fluxer: exploring global interaction through convolution, with a kernel generated directly from the current image features.
The paper develops this idea into a mixer that retains convolutional weight sharing and FFT evaluation, and studies the consequences of how its kernel is generated. We analyze what the construction allows and the restrictions it imposes, then examine it within a standard backbone across visual tasks. We hope this gives subsequent work a concrete alternative for global mixing, together with an understanding of its design that can guide further development.
We believe the value of this work lies in that operator and the knowledge needed to build on it, extending beyond a particular backbone configuration. We would be grateful if you could reconsider your rating in light of this contribution and the evidence in our response. We would also welcome any remaining concern you would like us to address.
Best regards,
The Authors
Thank you for the detailed feedback and additional comparisons. However, I remain concerned that the improvements over existing methods are relatively marginal in terms of both technical value and empirical performance. Therefore, I will maintain my initial rating.
Dear Reviewer SU9k,
Thank you for taking the time to consider our responses and the additional comparisons. We appreciate your careful review and your recognition of the experimental study and analysis.
We understand your remaining reservations about the significance of the contribution and would like to offer some further context. CAFormer, SPANetV2-hybrid, and VSSD combine attention with other mixers. ViF uses Fluxer as the token mixer in all four stages, with ViF-B reaching 85.4% top-1 accuracy on ImageNet-1K at 224 × 224. As reported in Section 5.3, Fig. 6 (left), and Table A8, the same ViF-T checkpoint trained at 224 × 224 is evaluated at the tested resolutions from 128 × 128 to 512 × 512 without fine-tuning or weight resampling. At 384 × 384, it achieves 82.97% top-1 accuracy, only 0.90 percentage points below its 83.87% at 224 × 224. We hope these properties, together with the analysis and controlled comparisons, offer useful additional context for evaluating Fluxer as a reusable global mixer.
We would also be grateful if you could clarify which specific experiments, comparisons, or aspects of the technical contribution remain a concern.
Best regards,
The Authors
Response to Reviewer SU9k
Thank you for recognizing the motivation and the analysis of adaptive global filters. We appreciate your suggestions on the comparisons and training recipe.
Comparisons with recent backbones
We provide comparisons with CAFormer, VSSD, and 2026 methods. Sources are CAFormer, VSSD, ICCV 2025, Random Wins All, CVPR 2026, and SPANetV2, TPAMI 2026. ViF results below are from our submission.
ImageNet-1K at 224 × 224. We include the S18/S36 variants of CAFormer and SPANetV2-hybrid. Entries are parameters (M) / FLOPs (G) / top-1 (%).
CAFormer and SPANetV2-hybrid combine different token mixers, including attention, across stages. ViF uses Fluxer as the token mixer in all four stages, providing dynamic global convolution throughout the backbone. These comparisons assess complete backbones, whereas Fig. 5 isolates our kernel-generation design within a fixed backbone and training recipe.
COCO with Mask R-CNN, 1× schedule. Parameters/FLOPs include the detector.
ADE20K with UPerNet. Parameters/FLOPs include the segmentor; SS/MS denote single-/multi-scale testing.
Token mixer and training recipe
Our contribution is the Fluxer token mixer and its analysis. We use an established four-stage hierarchy to evaluate its effectiveness as a general mixing mechanism. Developing such operators remains valuable because they provide reusable building blocks and design insights for visual representation learning beyond a particular backbone configuration.
Fluxer synthesizes a full-resolution convolution kernel from each image and applies it by FFT. It preserves convolutional weight sharing while making global interactions content-dependent. The nonlinear spatial construction permits cross-frequency interactions beyond frequency-pointwise kernel generation. We analyze this distinction and test the generator choices with the surrounding backbone and training recipe fixed.
We adopt MESA as an existing training technique. Without MESA, ViF-T/S/B reach 83.68/84.60/85.35%; with MESA, they reach 83.93/84.84/85.41%, giving gains of 0.25/0.24/0.06 points.
For Fig. 5, we retrain the generator variants with the same stem, downsampling, positional encodings, gated MLPs, widths, depths, and training recipe, including MESA. The static filter reaches 83.2%, frequency-pointwise generation 83.3%, shift-invariant generation 83.5%, and Fluxer 83.9%. The static-to-Fluxer gain is therefore measured with the surrounding architecture and recipe fixed.