Papers
Topics
Authors
Recent
Search
2000 character limit reached

CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large VIsion-Language Models

Published 13 May 2026 in cs.CV and cs.AI | (2605.13178v1)

Abstract: In large vision-LLMs, visual tokens typically constitute the majority of input tokens, leading to substantial computational overhead. To address this, recent studies have explored pruning redundant or less informative visual tokens for image understanding tasks. However, these methods struggle with pixel grounding tasks, where token importance is highly contingent on the input text. Through an in-depth analysis of CLIP, we observe that visual tokens located within referent regions often exhibit low similarity to the textual representation. Motivated by this insight, we introduce LiteLVLM, a training-free, text-guided token pruning strategy for efficient pixel grounding inference. By reversing the ranking of CLIP's visual-text similarity, LiteLVLM effectively retains visual tokens covering the referent regions, while recovering context tokens to enable clear foreground-background separation. Extensive experiments demonstrate that LiteLVLM significantly outperforms existing methods by over 5% across diverse token budgets. Without any training or fine-tuning, LiteLVLM maintains 90\% of the original performance with a 22% speedup and a 2.3x memory reduction. Our code is available at https://github.com/sejong-rcv/LiteLVLM.

Authors (2)

Summary

  • The paper introduces LiteLVLM, a training-free, text-guided token pruning method that improves pixel grounding in LVLMs.
  • It leverages insights from CLIP’s similarity reversal to retain referent-specific tokens, achieving 90.3% performance with just one-third of the tokens.
  • Empirical results show significant speedups, memory reductions, and robust performance across both image and video grounding tasks.

Training-Free Token Pruning for Efficient Pixel Grounding in LVLMs: An Expert Analysis of LiteLVLM

Motivation and Problem Definition

Large Vision-LLMs (LVLMs), such as LLaVA and CLIP-based architectures, have rapidly advanced the integration of visual and linguistic reasoning, unlocking capabilities in pixel-level grounding and region-based interpretation. However, these models are computationally burdensome due to the dominance of visual tokens in the input sequence—especially when processing high-resolution images or video. Existing token pruning strategies are mostly text-agnostic or maximize visual-textual similarity, showing limited efficacy on pixel grounding tasks where the relevance of tokens is context-dependent and text-guided. This work introduces LiteLVLM, a training-free, text-guided token pruning approach that preserves performance and drastically enhances inference efficiency for pixel grounding.

Analysis of CLIP Token Dynamics and Similarity Reversal

The authors present a detailed investigation of CLIP's internals, identifying the phenomenon of visual-text similarity reversal, especially pronounced at the token level. Crucially, visual tokens spatially aligned with the target referent (e.g., an object described in the prompt) tend to exhibit lower similarity to the [EOS] textual embedding than context or background tokens. This counterintuitive property is traced to two CLIP artifacts:

  • Contrastive Pretraining Bias: The CLIP loss exclusively aligns the CLS and EOS embeddings, propagating gradients more strongly to globally salient, background tokens rather than referent-specific ones.
  • Text Attention Sink: The [EOS] token, rather than aggregating semantics, overwhelmingly attends to the [SOS] token, neglecting the referring tokens in the prompt. This limits the semantic informativeness of the [EOS] representation.

Consequently, pruning strategies that favor high global similarity or disregard text altogether fail to retain referent-specific information, leading to a sharp degradation on pixel-level grounding.

The LiteLVLM Approach: Methodology

Pruning Pipeline and Token Selection

LiteLVLM implements a text-guided, dual-prioritization approach without requiring additional training or fine-tuning:

  • Similarity-Aware Selection: Retain visual tokens with minimal similarity to the [EOS] text embedding; these capture referent cues essential for grounding and segmentation.
  • Context-Aware Recovery: Restore a subset of globally salient tokens by measuring each visual token’s contribution to [CLS], thus ensuring that background and semantic context are preserved to prevent spurious boundaries or incomplete masks.
  • Adaptive Ratioing: The split between similarity-aware and context tokens is dynamically determined based on the prompt and task, optimizing empirical performance across varying token budgets.

The method is implemented as a pre-filter prior to sequence encoding by the LLM, ensuring compatibility with acceleration tools like FlashAttention and facilitating scalability.

Architectural Integration

LiteLVLM is deployed atop frameworks such as GLaMM for image tasks and VideoGLaMM for video, leveraging pretrained SAM encoders for pixel-level mask prediction and using state-of-the-art large language backbones (e.g., Vicuna-7B, Qwen2.5-VL).

Empirical Results

Image Pixel Grounding

Across the RefCOCO, RefCOCO+, and RefCOCOg datasets, LiteLVLM consistently outperforms both text-agnostic and similarity-maximizing baselines (e.g., TRIM, LLaVA-PruMerge, VisionZip, VisPruner):

  • Pruning to one-third the original tokens (192/576) maintains 90.3% of full-model performance and yields a 2.3x memory reduction and 22% speedup in inference.
  • At extreme pruning rates (64 tokens, ~11%), LiteLVLM surpasses other methods by more than 10% absolute in mIoU/cIoU.
  • Text-guided methods that rely strictly on high text-visual similarity exhibit severe performance collapse, validating the necessity of the similarity-reversal insight.

Video Pixel Grounding

On referring video object segmentation (Ref-DAVIS-17, Refer-YouTube-VOS, MeViS), LiteLVLM:

  • Preserves up to 99.5% of original performance at substantial (66.7%) token reduction, outperforming the strongest competitors.
  • Demonstrates robustness across architectures and datasets, with minimal degradation under highly limited token budgets.

Ablation and Generalization Studies

  • Component analysis confirms that the combination of low-similarity selection and context-aware recovery outperforms either individually or random pruning.
  • Adaptive token ratioing is empirically superior to fixed splits, underscoring the need for prompt-conditioned selection.
  • The method generalizes across alternative backbones (Qwen2.5-VL, SigLIP-2) and vision encoders, validating its architectural agnosticism.

Efficiency

  • End-to-end FLOPs, latency, and memory usage are substantially reduced, enabling practical LVLM deployment on edge and resource-constrained environments.
  • Quantization further compounds memory advantages with negligible performance degradation.

Theoretical and Practical Implications

LiteLVLM’s findings challenge assumptions about token informativeness in vision-language fusion, demonstrating that global visual-text alignment objectives can inadvertently mute referent-specific features. The presented method offers a blueprint for efficient, training-free adaptation of LVLMs to real-world pixel grounding applications, rendering them viable at scale in interactive, low-latency settings.

On the theoretical front, the similarity-reversal identified in CLIP and extended to other CLIP-family vision encoders signals the necessity of rethinking attention/contrastive objectives when fine-grained spatial reasoning is required. This may inspire alternative pretraining regimes or dual-representation approaches better suited to dense grounding.

Future Directions

Potential developments include:

  • Extension to dense captioning, VQA with spatial constraints, and multimodal tracking tasks.
  • Investigation of training-time loss functions or architecture modifications to directly address the semantic bottleneck at the [EOS] token.
  • Dynamic or confidence-aware budget strategies integrating uncertainty estimation for further computational gains.

Conclusion

LiteLVLM establishes a training-free, text-adaptive token pruning paradigm that overcomes the inherent global-bias limitations of CLIP pretraining and existing pruning heuristics. By integrating reversed similarity insights with context-aware recovery, it achieves strong pixel grounding performance with major efficiency gains—a critical step toward scalable, deployable LVLMs for both static imagery and video under real-world constraints.

Reference: "CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large Vision-LLMs" (2605.13178)

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.