Research map / Image generation / Autoregressive and latent generation
Semantic-Conditional Diffusion Networks for Image Captioning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 semantic condition to conventional diffusion (ablation) | Diffusion Transformer · COCO Karpathy test split | CIDEr: 114.5 -> 116.2 | Helped |
| Apply standard SCST to semantic-conditioned diffusion (ablation) | Diffusion Transformer · COCO Karpathy test split | CIDEr: 116.2 -> 120.8 | Helped |
| Apply standard SCST instead of no sentence-level optimization (ablation) | Semantic-conditioned Diffusion Transformer · COCO Karpathy test split | BLEU@4: 36.4 -> 34.6 | Hurt |
| Replace SCST with guided SCST using a teacher-generated sentence (ablation) | Semantic-conditioned Diffusion Transformer · COCO Karpathy test split | CIDEr: 120.8 -> 128.6 | Helped |
| Add cascaded Diffusion Transformer stages (ablation) | Semantic + GSCST · COCO Karpathy test split | CIDEr: 128.6 -> 131.6 | Helped |
| Increase single Diffusion Transformer block count from 3 to 4 (ablation) | Single Diffusion Transformer · COCO Karpathy test split | CIDEr: 128.6 -> 128.6 | No clear effect |
| Increase single Diffusion Transformer block count from 3 to 5 (ablation) | Single Diffusion Transformer · COCO Karpathy test split | CIDEr: 128.6 -> 128.1 | Hurt |
| Increase single Diffusion Transformer block count from 3 to 6 (ablation) | Single Diffusion Transformer · COCO Karpathy test split | CIDEr: 128.6 -> 128.1 | Hurt |
| Increase cascade depth from 1 to 2 Diffusion Transformers (ablation) | SCD-Net · COCO Karpathy test split | CIDEr: 128.6 -> 131.6 | Helped |
| Increase cascade depth from 2 to 3 Diffusion Transformers (ablation) | SCD-Net · COCO Karpathy test split | CIDEr: 131.6 -> 131.7 | No clear effect |
| Use semantic-conditioned diffusion instead of a standard Transformer baseline | Image captioning on COCO · COCO Karpathy test split (cross-entropy) | CIDEr: 113.3 -> 118.0 | Helped |
| Use SCD-Net with guided self-critical training instead of M² Transformer | Image captioning on COCO · COCO Karpathy test split (CIDEr optimization) | CIDEr: 131.2 -> 131.6 | Helped |
Closest papers by meaning
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.91)
- Enhancing Text-to-Image Diffusion Transformer via Split-Text Conditioning (NeurIPS 2025 · similarity 0.90)
- Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis (CVPR 2025 · similarity 0.89)
- Stochastic Conditional Diffusion Models for Robust Semantic Image Synthesis (ICML 2024 · similarity 0.89)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.89)
- Exploring the Role of Large Language Models in Prompt Encoding for Diffusion Models (NeurIPS 2024 · 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.