砍掉90%视觉token性能不掉

重点关注

• Enhancing Spatial Understanding in Image Generation via Reward Modeling score 10
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；有代码实现；关键词(1): text-to-image；顶会接收: CVPR
• Accelerating Masked Image Generation by Learning Latent Controlled Dynamics score 7
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；有代码实现；关键词(2): lightweight, text-to-image
• SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching score 5
  入选 HF Daily Papers；有代码实现

也值得关注

• DLEBench: Evaluating Small-scale Object Editing Ability for Instruction-based Image Editing Model score 4
  入选 HF Daily Papers；关键词(1): reasoning
• Towards Source-Aware Object Swapping with Initial Noise Perturbation score 4
  关键词(2): lightweight, finetuning；顶会接收: CVPR
• HiDrop: Hierarchical Vision Token Reduction in MLLMs via Late Injection, Concave Pyramid Pruning, and Early Exit score 4
  关键词(1): pruning；顶会接收: ICLR
• From Static Benchmarks to Dynamic Protocol: Agent-Centric Text Anomaly Detection for Evaluating LLM Reasoning score 4
  关键词(1): reasoning；顶会接收: ICLR
• UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking score 4
  关键词(3): pruning, deployment, real-time；顶会接收: CVPR
• U-Mind: A Unified Framework for Real-Time Multimodal Interaction with Audiovisual Generation score 4
  关键词(3): real-time, reasoning, embodied；顶会接收: CVPR
• Reasoning-Driven Multimodal LLM for Domain Generalization score 4
  关键词(2): fine-tuning, reasoning；顶会接收: ICLR
• FluoCLIP: Stain-Aware Focus Quality Assessment in Fluorescence Microscopy score 4
  关键词(1): vision-language；顶会接收: CVPR
• EMO-R3: Reflective Reinforcement Learning for Emotional Reasoning in Multimodal Large Language Models score 4
  关键词(2): fine-tuning, reasoning；顶会接收: CVPR
• Interpretable Debiasing of Vision-Language Models for Social Fairness score 4
  关键词(2): reasoning, vision-language；顶会接收: CVPR
• Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection score 4
  关键词(2): scaling, pre-training；顶会接收: ICLR
• Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization score 4
  关键词(4): quantization, post-training, mixture of experts, vision-language；顶会接收: CVPR
• FoV-Net: Rotation-Invariant CAD B-rep Learning via Field-of-View Ray Casting score 4
  关键词(1): lightweight；顶会接收: CVPR
• DiffusionHarmonizer: Bridging Neural Reconstruction and Photorealistic Simulation with Online Diffusion Enhancer score 4
  机构: NVIDIA；关键词(2): production, data curation
• Fixed Anchors Are Not Enough: Dynamic Retrieval and Persistent Homology for Dataset Distillation score 4
  关键词(1): distillation；顶会接收: CVPR