Research map / Image generation / Text-to-image diffusion
Dataset Diffusion: Diffusion-based Synthetic Data Generation for Pixel-Level Semantic Segmentation
Quang Ho Nguyen, Truong Tuan Vu, Anh Tuan Tran, Khoi 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 |
|---|---|---|---|
| append class labels to captions (class-prompt appending) vs captions only (ablation) | Dataset Diffusion · VOC | mIoU: 50.8 -> 62.0 | Improved |
| append class labels to captions vs class labels only (ablation) | Dataset Diffusion · VOC | mIoU: 57.4 -> 62.0 | Improved |
| add self-attention exponentiation refinement to cross-attention (ablation) | Dataset Diffusion · VOC | mIoU: 44.8 -> 61.0 | Improved |
| add uncertainty-aware loss (ablation) | Dataset Diffusion · VOC | mIoU: 61.0 -> 62.0 | Improved |
| add self-training (ablation) | Dataset Diffusion · VOC | mIoU: 62.0 -> 62.7 | Improved |
| add test-time augmentation (ablation) | Dataset Diffusion · VOC | mIoU: 62.7 -> 64.3 | Improved |
| set self-attention exponent tau to 4 instead of 5 (ablation) | Dataset Diffusion · VOC | mIoU: 60.5 -> 62.0 | Improved |
| use uncertainty range 0.5-0.6 instead of 0.4-0.6 (ablation) | Dataset Diffusion · VOC | mIoU: 60.7 -> 62.0 | Improved |
| aggregate attention over all timesteps instead of subranges (ablation) | Dataset Diffusion · VOC | mIoU: 61.5-61.9 -> 62.0 | No clear effect |
| increase synthetic images from 10k to 40k (ablation) | Dataset Diffusion · VOC | mIoU: 63.8 -> 64.8 | Improved |
| use Dataset Diffusion synthetic data instead of DiffuMask | Mask2Former R50 · VOC val | mIoU: 57.4 -> 60.2 | Improved |
| add 20 labeled images to 5k synthetic data | Dataset Diffusion · CelebA-HQ-Mask | mIoU: 78.2 -> 89.9 | Improved |
Later papers that compared with this method
5 later accepted papers report Dataset Diffusion as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- JoDiffusion: Jointly Diffusing Image with Pixel-Level Annotations for Semantic Segmentation Promotion (AAAI 2026 · better on 10 of 10 shared results)
- MMDiff: Extending Diffusion Transformers for Multi-Modal Generation (ECCV 2026 · better on 2 of 2 shared results)
- FreeGen: Bridging Visual-Linguistic Discrepancies Towards Diffusion-based Pixel-level Data Synthesis (AAAI 2025 · better on 2 of 2 shared results)
- SeeDiff: Off-the-Shelf Seeded Mask Generation from Diffusion Models (AAAI 2025 · better on 1 of 1 shared result)
- Exploring Phrase-Level Grounding with Text-to-Image Diffusion Model (ECCV 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- $\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation (NeurIPS 2026 · similarity 0.90)
- EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models (ICLR 2024 · similarity 0.90)
- Open-Vocabulary Attention Maps with Token Optimization for Semantic Segmentation in Diffusion Models (CVPR 2024 · similarity 0.89)
- SeeDiff: Off-the-Shelf Seeded Mask Generation from Diffusion Models (AAAI 2025 · similarity 0.89)
- Text-Image Alignment for Diffusion-Based Perception (CVPR 2024 · similarity 0.89)
- Seg4Diff: Unveiling Open-Vocabulary Semantic Segmentation in Text-to-Image Diffusion Transformers (NeurIPS 2025 · 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.
