Research map / Architectures and efficiency / Attention and transformers
Distillation-free Scaling of Large State-Space Models for Images and Videos
Hamid Suleman, Syed Talal Wasim, Muzammal Naseer, Jüergen Gall
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 |
|---|---|---|---|
| interleave one Transformer block per seven Mamba blocks (StableMamba) instead of pure Mamba, no distillation | VideoMamba-M · ImageNet-1K val | Top-1 accuracy (%): 81.4 -> 83.1 | Improved |
| replace distilled VideoMamba-B with undistilled StableMamba-B | VideoMamba-B · ImageNet-1K val | Top-1 accuracy (%): 82.7 -> 83.9 | Improved |
| add distillation-initialization to StableMamba-B | StableMamba-B · ImageNet-1K val | Top-1 accuracy (%): 83.9 -> 84.1 | Improved |
| add distillation-initialization to StableMamba-M | StableMamba-M · Kinetics-400 | Top-1 accuracy (%): 82.2 -> 82.5 | Improved |
| replace VideoMamba-M (distilled) with StableMamba-M (distilled) | VideoMamba-M · Kinetics-400 | Top-1 accuracy (%): 81.9 -> 82.5 | Improved |
| replace VideoMamba-M (distilled) with StableMamba-M (distilled) | VideoMamba-M · Something-Something-v2 | Top-1 accuracy (%): 67.3 -> 68.1 | Improved |
| replace VideoMamba-M with StableMamba-M | VideoMamba-M · ImageNet-C | mCE (%): 51.6 -> 50.5 | Improved |
| place the Transformer block in the middle rather than start or end (ablation) | StableMamba-T/S · ImageNet-1K val | Top-1 accuracy (%) | No clear effect |
| increase Mamba blocks per Transformer block up to 1:7 (ablation) | StableMamba · ImageNet-1K val | Top-1 accuracy (%): optimal at ratio 1:7 | Mixed |
| increase resolution from 224 to 448 (ablation) | StableMamba-T · ImageNet-1K val | Top-1 accuracy (%): 77.4 -> 79.9 | Improved |
| increase frames from 16 to 32 (ablation) | StableMamba-T · Kinetics-400 | Top-1 accuracy (%): 78.6 -> 79.3 | Improved |
| use more training data (25% to 100%) (ablation) | StableMamba · ImageNet-1K val | Top-1 accuracy (%) | Improved |
Closest papers by meaning
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.90)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.90)
- Mamba-Adaptor: State Space Model Adaptor for Visual Recognition (CVPR 2025 · similarity 0.89)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.89)
- Block-Biased Mamba for Long-Range Sequence Processing (NeurIPS 2025 · similarity 0.89)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.88)
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.
