Research map / Vision-language and multimodal / Multimodal LLMs
UltraViT: Latency-Optimized On-device Vision Encoder for Large Vision-Language Models
Ioannis Maniadis Metaxas, Adrian Bulat, Alberto Baldrati, Anestis Zaganidis, Yassine Ouali, Hyeonuk Kim, Georgios Tzimiropoulos
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 |
|---|---|---|---|
| use heterogeneous mixers (C, C+S, CP, A) instead of all-convolutional (ablation) | UltraViT · LLaVA-OV 1.5B avg of 7 benchmarks | avg score: 56.4 -> 58.5 | Improved |
| channel-partitioned attention in stage 3 instead of full single-head attention (ablation) | UltraViT · LLaVA-OV 1.5B avg | avg score / throughput: 58.5 vs 58.5; 148.9 vs 126.7 inf/s | No clear effect |
| dense distillation instead of contrastive training (ablation) | UltraViT · DocVQA | accuracy: 58.9 -> 63.9 | Improved |
| dense distillation instead of contrastive training (ablation) | FastViT · DocVQA | accuracy: 60.7 -> 62.1 | Improved |
| use larger teacher for dense distillation (ablation) | UltraViT · DocVQA | accuracy: 63.9 -> 67.5 | Improved |
| add generative pre-training with frozen 1.5B LLM (ablation) | UltraViT · DocVQA | accuracy: 67.5 -> 72.0 | Improved |
| generative pre-training with 0.5B instead of 1.5B LLM (ablation) | UltraViT · DocVQA | accuracy: 72.0 -> 70.0 | Hurt |
| dynamic 0.5B/1.5B LLM switching vs pure 1.5B (ablation) | UltraViT · InfoVQA | accuracy: 46.8 -> 47.9 | No clear effect |
| choice of teacher for dense pretraining (ablation) | UltraViT · DocVQA | accuracy: 63.9 (SigLIP2-NF) vs 65.3 (SigLIP2) | Mixed |
| UltraVLM vs re-trained FastVLM | UltraViT · TextVQA | accuracy: 62.9 -> 68.9 | Improved |
| UltraVLM vs re-trained FastVLM | UltraViT · MMMU | accuracy: 41.1 -> 38.1 | Hurt |
| UltraVLM vs token reduction methods | UltraViT · DocVQA | accuracy: 55.1 (VisionZip) -> 71.8 | Improved |
Closest papers by meaning
- FastVLM: Efficient Vision Encoding for Vision Language Models (CVPR 2025 · similarity 0.88)
- Unveiling Encoder-Free Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · similarity 0.88)
- iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models (ICLR 2026 · similarity 0.88)
- InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD (NeurIPS 2024 · similarity 0.87)
- VLsI: Verbalized Layers-to-Interactions from Large to Small 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.
