Updated
Updated · arxiv.org · Jul 23
Visual Contrastive Self-Distillation
Updated
Updated · arxiv.org · Jul 23

Visual Contrastive Self-Distillation

1 articles · Updated · arxiv.org · Jul 23

Summary

  • Researchers have introduced Visual Contrastive Self-Distillation (VCSD), a new technique to improve vision-language models without external teachers or privileged information.
  • VCSD leverages the contrast between predictions on original and content-erased images to provide a more informative self-distillation signal for model training.
  • Tested on Qwen3 models, VCSD consistently outperformed traditional methods, suggesting a simpler and more effective approach to enhancing visual grounding in AI systems.