LLaVA-UHD v4 reduces visual-encoding FLOPs by 55.8% for high-resolution images in MLLMs via slice-based encoding plus intra-ViT early compression while matching or exceeding baseline performance on document, OCR, and VQA benchmarks.
A-vit: Adaptive tokens for efficient vision transformer
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 2verdicts
UNVERDICTED 2roles
background 1polarities
background 1representative citing papers
FAR substitutes self-attention in pretrained DeiTs with multi-head bidirectional LSTMs via block-wise distillation and structured pruning to enable IMC-compatible inference with comparable accuracy and lower latency.
citing papers explorer
-
LLaVA-UHD v4: What Makes Efficient Visual Encoding in MLLMs?
LLaVA-UHD v4 reduces visual-encoding FLOPs by 55.8% for high-resolution images in MLLMs via slice-based encoding plus intra-ViT early compression while matching or exceeding baseline performance on document, OCR, and VQA benchmarks.
-
FAR: Function-preserving Attention Replacement for IMC-friendly Inference
FAR substitutes self-attention in pretrained DeiTs with multi-head bidirectional LSTMs via block-wise distillation and structured pruning to enable IMC-compatible inference with comparable accuracy and lower latency.