Research map / Vision-language and multimodal / Multimodal LLMs
LLaVA-SP: Enhancing Visual Representation with Visual Spatial Tokens for MLLMs
Haoran Lou, Chunxiao Fan, Ziyan Liu, Yuexin Wu, Xinliang Wang
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 SP spatial tokens (Pooling variant) to LLaVA-1.5 trained with LoRA | LLaVA-1.5 Vicuna-7B · VizWiz | accuracy: 45.7 -> 51.6 | Improved |
| add six visual spatial tokens (SFE, Cropping) (ablation) | LLaVA-1.5 Vicuna-7B · 11-benchmark average (AvgN) | AvgN: 63.4 -> 64.6 | Improved |
| add six visual spatial tokens (SFE, Pooling) (ablation) | LLaVA-1.5 Vicuna-7B · 11-benchmark average (AvgN) | AvgN: 63.4 -> 64.9 | Improved |
| use convolution instead of transformer blocks in SFE (Cropping) (ablation) | SFE Cropping · 11-benchmark average (AvgN) | AvgN: 64.5 -> 64.6 | Improved |
| use convolution instead of transformer blocks in SFE (Pooling) (ablation) | SFE Pooling · 11-benchmark average (AvgN) | AvgN: 64.4 -> 64.9 | Improved |
| use cropping tokens instead of sliding-window tokens (ablation) | SFE · Spatial understanding subset | AvgN: 44.7 -> 46.5 | Improved |
| increase spatial tokens from 1 to 6 (ablation) | SFE Cropping · 11-benchmark average (AvgN) | AvgN: 63.4 -> 64.6 | Improved |
| increase spatial tokens from 6 to 12 (ablation) | SFE Cropping · 11-benchmark average (AvgN) | AvgN: 64.6 -> 64.6 | No clear effect |
| add DFI cross-attention (5x5 Zs-big) to SFE Cropping (ablation) | LLaVA-SP-Cropping · 11-benchmark average (AvgN) | AvgN: 64.6 -> 64.8 | Improved |
| add DFI cross-attention (5x5 Zs-big) to SFE Pooling (ablation) | LLaVA-SP-Pooling · 11-benchmark average (AvgN) | AvgN: 64.9 -> 65.1 | Mixed |
| try multi-layer CNN blocks or small convolution kernels in SFE (ablation) | SFE · training stability | training | Hurt |
| add SP tokens (Cropping) for visual grounding | LLaVA-1.5 · RefCOCO val | accuracy: 54.7 -> 60.3 | Improved |
Closest papers by meaning
- LLaVA-ST: A Multimodal Large Language Model for Fine-Grained Spatial-Temporal Understanding (CVPR 2025 · similarity 0.92)
- iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models (ICLR 2026 · similarity 0.90)
- TC-LLaVA: Rethinking the Transfer of LLava from Image to Video Understanding with Temporal Considerations (AAAI 2025 · similarity 0.90)
- LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token (ICLR 2025 · similarity 0.90)
- HiRes-LLaVA: Restoring Fragmentation Input in High-Resolution Large Vision-Language Models (CVPR 2025 · similarity 0.90)
- Feast Your Eyes: Mixture-of-Resolution Adaptation for Multimodal Large Language Models (ICLR 2025 · similarity 0.89)
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.
