Research map / Vision-language and multimodal / Human and embodied modeling
Grad-ECLIP: Gradient-Based Visual and Textual Explanations for CLIP
Chenyang Zhao, Kun Wang, Janet H. Hsiao, Antoni B. Chan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add loosened 0-1 normalized spatial weight lambda_i (ablation) | Grad-ECLIP · ImageNet val (GT, top-1) | IMD: 0.118 -> 0.137 | Improved |
| add loosened spatial weight lambda_i (ablation) | Grad-ECLIP · MS COCO image explanation IR@1 | IMD: 0.001 -> 0.033 | Improved |
| add loosened spatial weight lambda_i (ablation) | Grad-ECLIP · MS COCO text explanation IR | IMD: 0.001 -> 0.076 | Improved |
| add loosened spatial weight lambda_i (ablation) | Grad-ECLIP · ImageNet-S | Point Game: 0.8356 -> 0.8899 | Improved |
| add loosened spatial weight lambda_i (ablation) | Grad-ECLIP · ImageNet-S | maskIoU: 0.3314 -> 0.2869 | Hurt |
| replace softmax attention with 0-1 normalized similarity as spatial weight (ablation) | Grad-ECLIP · qualitative heat maps | explanation quality | Improved |
| aggregate only the last layer instead of all 12 layers for image explanation (ablation) | Grad-ECLIP image encoder · MS COCO IR | IMD: -0.1465 -> 0.0330 | Improved |
| aggregate last 8 layers instead of 1 layer for text explanation (ablation) | Grad-ECLIP text encoder · MS COCO IR | IMD: -0.0059 -> 0.0296 | Improved |
| aggregate last 8 layers instead of all 12 for text explanation (ablation) | Grad-ECLIP text encoder · MS COCO IR | IMD: 0.0125 -> 0.0296 | Improved |
| use single attention head instead of multi-head for explanation (ablation) | Grad-ECLIP · qualitative heat maps | explanation focus | Improved |
| use predicted class instead of ground-truth class as text prompt | Grad-ECLIP · ImageNet val top-1 | IMD: 0.137 -> 0.113 | Hurt |
| compare Grad-ECLIP with CLIPSurgery on insertion | Grad-ECLIP vs CLIPSurgery · ImageNet val (GT, top-1) | Insertion AUC: 0.384 vs 0.383 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports Grad-ECLIP as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Concept Regions Matter: Benchmarking CLIP with a New Cluster-Importance Approach (CVPR 2026 · better on 8 of 8 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MaskCLIP — MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining (CVPR 2023) · better on 3 of 3 shared results
- Grad-CAM — Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization (ICCV 2017) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Gradient-based Visual Explanation for Transformer-based CLIP (ICML 2024 · similarity 0.97)
- FineLIP: Extending CLIP's Reach via Fine-Grained Alignment with Longer Text Inputs (CVPR 2025 · similarity 0.90)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.90)
- FG-CLIP: Fine-Grained Visual and Textual Alignment (ICML 2025 · similarity 0.90)
- SuperCLIP: CLIP with Simple Classification Supervision (NeurIPS 2025 · similarity 0.89)
- CLIP in Mirror: Disentangling text from visual images through reflection (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
