Research map / Architectures and efficiency / Attention and transformers
S4ND: Modeling Images and Videos as Multidimensional Signals with State Spaces
Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, Christopher Ré
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace self-attention with S4ND layers and drop positional encodings | ViT-B · ImageNet-1k | top-1 accuracy: 78.9 -> 80.4 | Improved |
| Replace local Conv2D with S4ND layers | ConvNeXt-T · ImageNet-1k | top-1 accuracy: 82.1 -> 82.2 | No clear effect |
| Replace Conv2D with S4ND in isotropic backbone | Conv2D-ISO · CIFAR-10 | top-1 accuracy: 93.7 -> 94.1 | Improved |
| Replace Conv2D with S4ND in small ConvNeXt | ConvNeXt-M · Celeb-A | top-1 accuracy: 91.0 -> 91.3 | Improved |
| Use temporal S4 kernel instead of I3D inflation | ConvNeXt-I3D · HMDB-51 RGB | top-1 accuracy: 58.1 -> 62.1 | Improved |
| Use separable S3D kernels instead of I3D inflation | ConvNeXt · HMDB-51 RGB | top-1 accuracy: 58.1 -> 58.6 | No clear effect |
| Shorten initial temporal kernel length from 20.0 to 1.0 (ablation) | S4ND-ConvNeXt-3D · HMDB-51 RGB | top-1 accuracy: 53.74 -> 62.07 | Improved |
| S4ND versus Conv2D, trained on mid resolution and tested on base | S4ND · CIFAR-10 | accuracy: 73.1 -> 88.77 | Improved |
| S4ND versus Conv2D, trained on low resolution and tested on base | S4ND · CIFAR-10 | accuracy: 33.1 -> 73.71 | Improved |
| S4ND versus Conv2D, trained on mid and tested on base | S4ND · Celeb-A | accuracy: 80.52 -> 90.14 | Improved |
| S4ND versus FlexNet-16, trained on mid and tested on base | S4ND · CIFAR-10 | accuracy: 82.7 -> 88.77 | Improved |
| Progressive resizing low->base (80-20 epochs) instead of training only at base | S4ND · CIFAR-10 | accuracy and speed: 93.40 -> 92.32 | Mixed |
Later papers that compared with this method
1 later accepted paper reports S4ND as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SONIC: Spectral Oriented Neural Invariant Convolutions (ICLR 2026 · better on 4 of 5 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Selective Structured State-Spaces for Long-Form Video Understanding (CVPR 2023 · similarity 0.87)
- GG-SSMs: Graph-Generating State Space Models (CVPR 2025 · similarity 0.85)
- Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video (CVPR 2025 · similarity 0.85)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.85)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.85)
- GSDD: Generative Space Dataset Distillation for Image Super-resolution (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
