👁️ Evaluation Tool for Vision Encoder Explainability Methods

A tool to evaluate different explainability methods for Vision Transformer Architectures.

University Project (Internship @M3S, SMART & SMU)

We created this tool to evaluate different explainability methods for different Vision Transformer Architectures:

  • LeGrad (ICCV 2025 – to be implemented)
  • CheferCAM (CVPR 2021)
  • Attention Rollout (ACL 2020)
  • Grad-CAM (IJCV 2019)

You can perturb the image (+ ve/ - ve) and see how it affects the final prediction.

Supported model families include ViT, DeiT, DinoV2, and ViTs with Registers (ICLR 2024).

You can check the live demo here: Hugging Face Space