Research map / Vision-language and multimodal / CLIP and zero-shot
MoRA: Missing Modality Low-Rank Adaptation for Visual Recognition
Shu Zhao, Nilesh Ahuja, Tan Yu, Tianyi Shen, Vijaykrishnan Narayanan
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 Gram matrices for cross-modal sharing (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 80.31 -> 83.77 | Improved |
| add modality-specific parameters (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 81.32 -> 83.77 | Improved |
| share parameters across modalities instead of independent ones (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy | Improved |
| use Gram matrices instead of linear projection (Align) (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 81.14 -> 83.77 | Improved |
| use Gram matrices instead of concatenation fusion (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 81.01 -> 83.77 | Improved |
| use fixed Gram matrix instead of learnable one (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 83.37 -> 83.77 | Improved |
| add frozen W0 into the shared branch (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 83.77 -> 83.59 | No clear effect |
| attach MoRA to rear layers instead of front layers (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 81.08 -> 83.77 | Improved |
| attach MoRA to rear layers instead of middle layers (ablation) | MoRA · Food101 (70% missing, both modalities missing) | accuracy: 82.48 -> 83.77 | Improved |
| use MoRA instead of LoRA | LoRA · Food101 (70% missing, both modalities missing) | accuracy: 82.14 -> 83.77 | Improved |
| use MoRA instead of full fine-tuning | FFT · Food101 (70% missing, both modalities missing) | accuracy: 14.05 -> 83.77 | Improved |
| add MoRA to CLIP4CIR | CLIP4CIR · MS-COCO image-to-image retrieval | Recall@1: 43.34 -> 60.50 | Improved |
Closest papers by meaning
- Deep Correlated Prompting for Visual Recognition with Missing Modalities (NeurIPS 2024 · similarity 0.89)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.88)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.88)
- MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · similarity 0.88)
- Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion (ICLR 2025 · similarity 0.88)
- MMRL: Multi-Modal Representation Learning for Vision-Language Models (CVPR 2025 · similarity 0.87)
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.
