Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• PruneSID achieves 96.3% accuracy on LLaVA-1.5 with only 11.1% token retention.result 95%
  Evidencepartial

  achieving 96.3% accuracy on LLaVA-1.5 with only 11.1% token retention

  Implicationpartial

  Directly stated in abstract with specific numeric results

  Verificationpartial

  partial

• PruneSID achieves 92.8% accuracy at extreme compression rates (5.6%) on LLaVA-NeXT, outperforming prior methods by 2.5%.result 95%
  Evidencepartial

  92.8% accuracy at extreme compression rates (5.6%) on LLaVA-NeXT, outperforming prior methods by 2.5%

  Implicationpartial

  Directly stated in abstract with specific numeric comparisons

  Verificationpartial

  partial

• PruneSID achieves 7.8× faster prefilling speed compared to the original model.result 95%
  Evidencepartial

  7.8× faster prefilling speed compared to the original model

  Implicationpartial

  Directly stated in abstract with specific numeric comparison

  Verificationpartial

  partial

• PruneSID uses a two-stage pipeline: Principal Semantic Components Analysis (PSCA) for clustering tokens and Intra-group Non-Maximum Suppression (NMS) for pruning redundant tokens.method 90%
  Evidencepartial

  featuring a two-stage pipeline: (1) Principal Semantic Components Analysis (PSCA) for clustering tokens into semantically coherent groups... and (2) Intra-group Non-Maximum Suppression (NMS) for pruning redundant tokens

  Implicationpartial

  Directly described in abstract as the core method

  Verificationpartial

  partial

• PruneSID incorporates an information-aware dynamic compression ratio mechanism that optimizes token compression rates based on image complexity.method 90%
  Evidencepartial

  PruneSID incorporates an information-aware dynamic compression ratio mechanism that optimizes token compression rates based on image complexity

  Implicationpartial

  Directly stated in abstract as a key component of the method

  Verificationpartial

  partial

• PruneSID is a training-free approach to visual token compression.technical 90%
  Evidencepartial

  PruneSID, a training-free Synergistic Importance-Diversity approach

  Implicationpartial

  Explicitly stated in abstract as a characteristic of the method

  Verificationpartial

  partial

• PruneSID generalizes across diverse VLMs and both image and video modalities, showcasing strong cross-modal versatility.result 85%
  Evidencepartial

  Our framework generalizes across diverse VLMs and both image and video modalities, showcasing strong cross-modal versatility

  Implicationpartial

  Directly stated in abstract but without specific evidence of video results

  Verificationpartial

  partial

• Existing compression methods struggle to balance importance preservation and information diversity in visual token compression.limitation 80%
  Evidencepartial

  existing compression methods struggle to balance importance preservation and information diversity

  Implicationpartial

  Stated as motivation in abstract but presented as a general problem rather than a specific finding

  Verificationpartial

  partial

Startup potential card