Research map / Vision-language and multimodal / CLIP and zero-shot
Steerable Vision Transformers
Jona Ruthardt, Manu Gaur, Deva Ramanan, Makarand Tapaswi, Yuki Asano
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 |
|---|---|---|---|
| early fusion (gated cross-attention inside ViT) instead of late fusion (ablation) | SteerViT · PODS | PR-AUC: 36.6 -> 58.1 | Improved |
| early fusion instead of late fusion (ablation) | SteerViT · CORE | acc@1: 93.3 -> 96.0 | Improved |
| early fusion instead of late fusion (ablation) | SteerViT · FG-CLS | accuracy: 91.8 -> 87.7 | Hurt |
| add zero-initialized tanh gate (ablation) | SteerViT · PODS | PR-AUC: 47.1 -> 58.1 | Improved |
| add zero-initialized tanh gate (ablation) | SteerViT · FG-CLS | accuracy: 83.5 -> 87.7 | Improved |
| add zero-initialized tanh gate (ablation) | SteerViT · CORE | acc@1: 94.6 -> 96.0 | Improved |
| use two-layer MLP projector instead of linear (ablation) | SteerViT · PODS | PR-AUC: 56.4 -> 58.1 | Improved |
| use two-layer MLP projector instead of linear (ablation) | SteerViT · FG-CLS | accuracy: 86.7 -> 87.7 | Improved |
| apply early fusion to SigLIP backbone vs late fusion (ablation) | SigLIP · CORE | acc@1: 75.4 -> 91.3 | Improved |
| apply early fusion to MAE backbone vs late fusion (ablation) | MAE · CORE | acc@1: 41.0 -> 74.9 | Improved |
| use detailed instance descriptions instead of supercategory prompts (ablation) | SteerViT · PODS classification | PR-AUC: 27.9 -> 58.1 | Improved |
| scale cross-attention gates by 0.6 at inference (ablation) | SteerViT (MAE) · representation quality | score: 40 -> 50 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MaskCLIP — MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining (CVPR 2023) · better on 1 of 1 shared result
- WinCLIP — WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation (CVPR 2023) · better on 1 of 1 shared result
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
- SteerVTE: Seamless Video Text Editing with Style and Glyph Control (NeurIPS 2026 · similarity 0.84)
- TextRefiner: Internal Visual Feature as Efficient Refiner for Vision-Language Models Prompt Tuning (AAAI 2025 · similarity 0.84)
- Hugging Visual Prompt and Segmentation Tokens: Consistency Learning for Fine-Grained Visual Understanding in MLLMs (CVPR 2026 · similarity 0.84)
- Decomposing and Interpreting Image Representations via Text in ViTs Beyond CLIP (NeurIPS 2024 · similarity 0.84)
- Retrieval-Enhanced Contrastive Vision-Text Models (ICLR 2024 · similarity 0.84)
- ITO: Multi-View Alignment and Training-Time Fusion for Image-Text Pretraining (NeurIPS 2026 · 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.
