Research map / Vision-language and multimodal / Human and embodied modeling
Guiding Image Captioning Models Toward More Specific Captions
Simon Kornblith, Lala Li, Zirui Wang, Thao Nguyen
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 |
|---|---|---|---|
| apply classifier-free guidance scale 2 at decoding instead of greedy decoding | CoCa-Base bottleneck captioner · MS-COCO Karpathy test | CLIPScore: 0.775 -> 0.808 | Improved |
| apply classifier-free guidance scale 2 | CoCa-Base bottleneck captioner · MS-COCO Karpathy test | caption-to-image recall@1: 26.5 -> 44.6 | Improved |
| apply classifier-free guidance scale 2 | CoCa-Base bottleneck captioner · MS-COCO Karpathy test | CIDEr: 126.1 -> 78.6 | Hurt |
| raise guidance scale from 3.0 to 4.0 (ablation) | CoCa-Base bottleneck captioner · MS-COCO Karpathy test | caption-to-image recall@1: 49.4 -> 44.7 | Hurt |
| use moderate guidance scale (up to 2.0) instead of none (ablation) | CoCa-Base bottleneck captioner · MS-COCO Karpathy test | caption length (words): 9.6 -> 9.3 | No clear effect |
| apply classifier-free guidance scale 2 on Stanford Dogs | CoCa-Base bottleneck captioner · Stanford Dogs test | % captions containing breed name: 1.9 -> 42.4 | Improved |
| apply classifier-free guidance scale 2 on Stanford Dogs | CoCa-Base bottleneck captioner · Stanford Dogs test | % breeds correct: 61.7 -> 58.5 | Mixed |
| prompt LM guidance with 20 MS-COCO captions for web-only model | CoCa-Base pretrained on JFT-5B + ALIGN · MS-COCO Karpathy test | CIDEr: 21.8 -> 57.4 | Improved |
| increase number of LM prompt captions from 5 to 50 (ablation) | LM guidance on web-only model · MS-COCO Karpathy test | CIDEr: 48.6 -> 59.7 | Improved |
| use LM guidance (alpha=5, beta=-5/2) instead of CFG gamma=1.6 | CoCa-Base fine-tuned on MS-COCO · MS-COCO Karpathy test | recall@1 and CIDEr: R@1 39.0 -> 39.6, CIDEr 109.3 -> 114.4 | Improved |
| use LM guidance (alpha=beta=1) instead of CFG on web-only model | CoCa-Base pretrained on JFT-5B + ALIGN · MS-COCO Karpathy test | CLIPScore and recall@1: R@1 27.3 (CFG gamma 4) -> 14.4 (LM 20 captions) | Hurt |
| use attention-pooling CoCa instead of the bottleneck architecture (ablation) | classifier-free guidance captioner · MS-COCO Karpathy test | caption-to-image recall@1 (gamma 2.0): 44.6 -> 48.6 | Improved |
Closest papers by meaning
- CaptionQA: Is Your Caption as Useful as the Image Itself? (CVPR 2026 · similarity 0.89)
- Revisit Large-Scale Image-Caption Data in Pre-training Multimodal Foundation Models (ICLR 2025 · similarity 0.89)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Mining Fine-Grained Image-Text Alignment for Zero-Shot Captioning via Text-Only Training (AAAI 2024 · similarity 0.88)
- BACON: Improving Clarity of Image Captions via Bag-of-Concept Graphs (CVPR 2025 · similarity 0.88)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (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.
