Research map / Image generation / Autoregressive and latent generation
Customized Image Narrative Generation via Interactive Visual Question Generation and Answering
Andrew Shin, Yoshitaka Ushiku, Tatsuya Harada
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 |
|---|---|---|---|
| Generate narratives using regional self-Q&A rather than standard image captioning | Image narrative generation · MS COCO human evaluation | Overall: 3.201 -> 3.391 | Improved |
| Generate narratives using regional self-Q&A rather than standard image captioning | Image narrative generation · MS COCO test 2014 | BLEU1: 20.77 -> 20.87 | Improved |
| Generate narratives using regional self-Q&A rather than standard image captioning | Image narrative generation · MS COCO test 2014 | BLEU4: 1.90 -> 1.41 | Hurt |
| Use VQG and VQA together to select questions with multiple plausible answers | Interactive question generation · Interactive question evaluation | votes for multiple responses: 217 -> 664 | Improved |
| Include question context in preference learning and narrative generation (ablation) | Same-image user choice reflection · Same-image user choice reflection | Avg. Score: 2.636$\pm$1.01 -> 3.851$\pm$1.12 | Improved |
| Use learned user preferences to customize narratives on new images | New-image interest reflection · New-image user interest reflection | Avg. Score: 2.238$\pm$1.24 -> 3.455$\pm$0.93 | Improved |
| Train preference learning with choices from the same user rather than random user matches (ablation) | New-image preference learning · New-image choice prediction | average score: 45.17 -> 68.72 | Improved |
| Train preference learning with randomly matched user choices rather than image features only (ablation) | New-image interest reflection · New-image user interest reflection | Avg. Score: 2.238$\pm$1.24 -> 2.772$\pm$0.79 | Improved |
| Replace the regional question-answer approach with directly training on collected long narratives (ablation) | Conventional image narrative captioning · MS COCO human-written narratives | descriptive accuracy | Hurt |
| Use DenseCap instead of Ours | Image narrative generation · MS COCO human evaluation | Accuracy: 3.132 -> 3.188 | Improved |
Closest papers by meaning
- ConVQG: Contrastive Visual Question Generation with Multimodal Guidance (AAAI 2024 · similarity 0.88)
- Image Content Generation with Causal Reasoning (AAAI 2024 · similarity 0.86)
- Explicitly Guided Difficulty-Controllable Visual Question Generation (AAAI 2025 · similarity 0.85)
- Instruct-Imagen: Image Generation with Multi-modal Instruction (CVPR 2024 · similarity 0.85)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.85)
- Davidsonian Scene Graph: Improving Reliability in Fine-grained Evaluation for Text-to-Image Generation (ICLR 2024 · similarity 0.85)
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.
