Awesome-Token-Compress

🔥🔥🔥 A paper list of some recent works about Token Compress for Vit and VLM.

VLM

2024

• Don’t Look Twice: Faster Video Transformers with Run-Length Tokenization .[RLT;Video;NeurIPS 2024;Github]

• Inference Optimal VLMs Need Only One Visual Token but Larger Models .[QueCC;Github]

• Video Token Merging for Long-form Video Understandin .[Learnable VTM;Video]

• LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding .[LongVU;Video;Github]

• PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction .[PyramidDrop;Github]

• Efficient Vision-Language Models by Summarizing Visual Tokens into Compact Registers .[Victor;]

• VidCompress: Memory-Enhanced Temporal Compression for Video Understanding in Large Language Models .[VidCompress;]

• Retrieval Replace Reduction:An effective visual token reduction method via semantic match .[TRSM;]

• AVG-LLaVA: A Large Multimodal Model with Adaptive Visual Granularity .[AVG-LLaVA;Github]

• Less is More: A Simple yet Effective Token Reduction Method for Efficient Multi-modal LLMs .[TRIM]

• TC-LLaVA: Rethinking the Transfer from Image to Video Understanding with Temporal Consideration .[TC-LLaVA;Video;]

• TG-LLaVA: Text Guided LLaVA via Learnable Latent Embeddings .[TG-LLaVA]

• mPLUG-DocOwl2: High-resolution Compressing for OCR-free Multi-page Document Understanding . [mPLUG-DocOwl2;Github]

• TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval . [TempMe;Video;Github]

• Recoverable Compression: A Multimodal Vision Token Recovery Mechanism Guided by Text Information . [Recoverable Compression]

• HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models in Resource-Constrained Environments . [HiRED;Github]

• Token-level Correlation-guided Compression for Efficient Multimodal Document Understanding . [Token-level;Github]

• HiRes-LLaVA: Restoring Fragmentation Input in High-Resolution Large Vision-Language Models . [HiRes-LLaVA;]

• TokenPacker: Efficient Visual Projector for Multimodal LLM . [TokenPacker;Github]

• VoCo-LLaMA: Towards Vision Compression with Large Language Models . [VoCo-LLaMA;Github]

• DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models . [DeCo;Github]

• Matryoshka Multimodal Models . [Matryoshka;M³]Github]

• How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites . [InternVL;Pixel-Shuffle;Github]

• CATP: Cross-Attention Token Pruning for Accuracy Preserved Multimodal Model Inference . [CATP;]

• LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models . [LLaVA-PruMerge;Github]

• An Image is Worth 1/2 Tokens After Layer 2: Plug-and-PLay Acceleration for VLLM Inference . [FastV;ECCV 2024;Github]

• MobileVLM V2: Faster and Stronger Baseline for Vision Language Model . [LDP-v2;Github]

2023

• Honeybee: Locality-enhanced Projector for Multimodal LLM . [C-Abstractor;CVPR 2024;Github ]
• LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models . [LLaMA-VID;ECCV 2024;Github ]
• Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond . [Resampler;Github]
• CrossGET: Cross-Guided Ensemble of Tokens for Accelerating Vision-Language Transformers . [CrossGET; ICML 2024;Github]
• BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models . [Q-former;Github]

  Vit

  2024

• Vote&Mix: Plug-and-Play Token Reduction for Efficient Vision Transformer . [Vote&Mix;]
• Token Compensator: Altering Inference Cost of Vision Transformer without Re-Tuning . [Token Compensator;ToCom;Github]
• Dynamic and Compressive Adaptation of Transformers From Images to Videos . [InTI;]
• LookupViT: Compressing visual information to a limited number of tokens . [LookupViT;DeepMind]
• PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation . [PYRA;ECCV 2024;Github]

2023

• PPT: Token Pruning and Pooling for Efficient Vision Transformers . [PPT;Github]
• DiffRate : Differentiable Compression Rate for Efficient Vision Transformers . [DiffRate;ICCV 2023;Github]
• Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers . [TPS;CVPR 2023;Github]

  2022

• TOKEN MERGING: YOUR VIT BUT FASTER . [ToMe;Token Merging; ICLR 2023]
• Adaptive Sparse ViT: Towards Learnable Adaptive Token Pruning by Fully Exploiting Self-Attention . [Adaptive Sparse ViT]
• EViT: Expediting Vision Transformers via Token Reorganizations . [EViT;ICLR 2022;Github]
• Vision Transformer Slimming: Multi-Dimension Searching in Continuous Optimization Space . [ViT-Slim;CVPR 2022;Github]

  2021

• A-ViT: Adaptive Tokens for Efficient Vision Transformer . [A-Vit;]
• ATS: Adaptive Token Sampling For Efficient Vision Transformers . [ATS;ECCV 2022;Github]
• Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer . [Evo-ViT;AAAI 2022;Github]
• Patch Slimming for Efficient Vision Transformers . [Patch Slimming;]
• DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsificationr . [DynamicViT;NeurIPS 2021;Github]