Research map / Vision-language and multimodal / Human and embodied modeling
Deep Visual-Semantic Alignments for Generating Image Descriptions
Andrej Karpathy, Li Fei-Fei
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 |
|---|---|---|---|
| use the simpler ranking cost instead of DeFrag loss (dependency tree edges) (ablation) | DeFrag alignment · Flickr30K ranking | Image annotation R@1: 19.2 -> 20.0 | Improved |
| replace dependency tree edges with BRNN (ablation) | alignment model · Flickr30K ranking | Image annotation R@1: 20.0 -> 22.2 | Improved |
| replace dependency tree edges with BRNN (ablation) | alignment model · Flickr30K ranking | Image search R@1: 15.0 -> 15.2 | Improved |
| replace dependency tree edges with BRNN (ablation) | alignment model · Flickr8K ranking | Image annotation R@1: 14.8 -> 16.5 | Improved |
| simpler cost function with dependency edges (ablation) | DeFrag alignment · Flickr8K ranking | Image annotation R@1: 13.8 -> 14.8 | Improved |
| use region-level model instead of full-frame model | multimodal RNN · Region description | BLEU-1: 14.2 -> 35.2 | Improved |
| use multimodal RNN instead of nearest neighbor retrieval | caption generator · MSCOCO 2014 captioning | CIDEr: 38.3 -> 66.0 | Improved |
| use multimodal RNN instead of nearest neighbor retrieval | caption generator · MSCOCO 2014 captioning | BLEU-4: 10.0 -> 23.0 | Improved |
| use beam size 7 instead of beam size 1 (ablation) | multimodal RNN · MSCOCO 2014 captioning | CIDEr: 0.61 -> 0.66 | Improved |
| use LSTM instead of RNN (ablation) | multimodal RNN · captioning | quality | Improved |
| train word vectors from random init instead of word2vec (ablation) | multimodal RNN · captioning | quality | No clear effect |
| clip gradients (ablation) | multimodal RNN · captioning | validation performance | Improved |
Closest papers by meaning
- Single-Stream Multi-level Alignment for Vision-Language Pretraining (ECCV 2022 · similarity 0.90)
- Aligning Where to See and What to Tell: Image Captioning with Region-Based Attention and Scene-Specific Contexts (TPAMI 2016 · similarity 0.90)
- Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning (NeurIPS 2026 · similarity 0.90)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.89)
- Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models (NeurIPS 2025 · similarity 0.89)
- Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding (NeurIPS 2026 · similarity 0.88)
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.
