Research map / Architectures and efficiency / Attention and transformers
Stitched ViTs are Flexible Vision Backbones
Zizheng Pan, Jing Liu, Haoyu He, Jianfei Cai, Bohan Zhuang
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 |
|---|---|---|---|
| add Two-way Stitching (TWS) to SN-Netv1 (ablation) | SN-Netv1 (DeiT3-S/L) · ADE20K, 34G FLOPs | mIoU (%): 40.4 -> 42.8 | Improved |
| add Two-way Stitching (TWS) to SN-Netv1 (ablation) | SN-Netv1 (DeiT3-S/L) · ADE20K, 363G FLOPs | mIoU (%): 51.6 -> 50.5 | Hurt |
| add Resource-constrained Sampling (ROS) on top of TWS (ablation) | SN-Netv2 (DeiT3-S/L) · ADE20K, 34G FLOPs | mIoU (%): 42.8 -> 43.7 | Improved |
| add LoRA on stitching layers on top of TWS and ROS (ablation) | SN-Netv2 (DeiT3-S/L) · ADE20K, 34G FLOPs | mIoU (%): 43.7 -> 45.6 | Improved |
| enlarge stitching space with TWS alone on ImageNet-1K (ablation) | SN-Netv1 (DeiT3-S/L) · ImageNet-1K | Top-1 accuracy (%): -17.2% Top-1 at small anchor | Hurt |
| use SN-Netv2 instead of training individual anchors separately | DeiT3-B/L on ADE20K · ADE20K | training cost (A100 GPU hours): 140 vs 174 + 90 | Improved |
| use SN-Netv2 with LoRA on backbone instead of DeiT3-L with LoRA | COCO-Stuff-10K backbone PEFT · COCO-Stuff-10K | mIoU (%): 44.8 -> 45.8 | Improved |
| apply LoRA PEFT to the entire backbone instead of full finetuning (ablation) | SN-Netv2 · COCO-Stuff-10K | mIoU (%): 48.2 -> 45.8 (top end) | Hurt |
| train SN-Netv2 on detection with heavy ViTDet decoder | SN-Netv2 DeiT3-S/L · COCO-2017 detection | bbox AP: 51.2 vs 51.9 bbox AP at DeiT3-L FLOPs | Mixed |
| remove low-rank update and fully finetune stitching layers (ablation) | SN-Netv2 (DeiT3-S/L) · COCO-Stuff-10K | mIoU | Hurt |
Closest papers by meaning
- Efficient Stitchable Task Adaptation (CVPR 2024 · similarity 0.90)
- Revisiting Model Stitching In the Foundation Model Era (CVPR 2026 · similarity 0.90)
- Elastic ViTs from Pretrained Models without Retraining (NeurIPS 2025 · similarity 0.87)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.87)
- Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model (NeurIPS 2024 · similarity 0.87)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.86)
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.
