Research map / Image generation / Editing and style transfer
DECap: Towards Generalized Explicit Caption Editing via Diffusion Mechanism
Zhen Wang, Xinyun Jiang, Jun Xiao, Tao Chen, Long Chen
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 |
|---|---|---|---|
| train on word-level noised unpaired captions with edit-based diffusion instead of paired ref-GT data (DECap vs TIger-N) | Explicit caption editing · COCO-EE test (in-domain) | CIDEr-D: 145.0 -> 173.7 | Improved |
| DECap vs TIger on in-domain CIDEr-D at 4 steps | DECap · COCO-EE test (in-domain) | CIDEr-D: 176.7 -> 173.7 | Mixed |
| increase DECap editing steps from 4 to 6 (ablation) | DECap · COCO-EE test (in-domain) | CIDEr-D: 173.7 -> 177.0 | Improved |
| predict edit operations and content words in parallel rather than via three sequential modules | DECap · COCO-EE test (in-domain) | Inference time (ms): 614.23 -> 277.30 | Improved |
| apply DECap editing to BLIP-generated captions | BLIP captions · COCO test | CIDEr-D: 136.7 -> 138.1 | Improved |
| apply TIger editing to BLIP-generated captions | BLIP captions · COCO test | CIDEr-D: 136.7 -> 104.0 | Hurt |
| edit pure random reference captions (DECap vs TIger-N) | Caption editing models · COCO test, pure random 10-word references | CIDEr-D: 23.5 -> 102.5 | Improved |
| increase random input words from 10 to 12 (ablation) | DECap caption generation · COCO test | CIDEr-D: 119.0 -> 109.3 | Hurt |
| use alpha > beta = gamma noising instead of even edit-type distribution (ablation) | DECap · COCO test | CIDEr-D: 117.4 -> 119.0 | Improved |
| noise with replacement only (no deletion or insertion) (ablation) | DECap · COCO test | CIDEr-D: 119.0 -> 116.6 | Hurt |
| increase generation steps from 10 to 15 (ablation) | DECap caption generation · COCO Karpathy test | CIDEr-D: 119.0 -> 121.2 | Improved |
Closest papers by meaning
- Enhancing Text-to-Image Diffusion Transformer via Split-Text Conditioning (NeurIPS 2025 · similarity 0.86)
- An Item Is Worth a Prompt: Versatile Image Editing with Disentangled Control (AAAI 2025 · similarity 0.86)
- Noise-Aware Image Captioning with Progressively Exploring Mismatched Words (AAAI 2024 · similarity 0.85)
- Diffusion Bridge: Leveraging Diffusion Model to Reduce the Modality Gap Between Text and Vision for Zero-Shot Image Captioning (CVPR 2025 · similarity 0.85)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.85)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · 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.
