Research map / Image generation / Fast diffusion sampling
Accelerating Diffusion Model Training under Minimal Budgets: A Condensation-Based Perspective
Rui Huang, Shitong Shao, zikai zhou, Pukun Zhao, Hangyu Guo, Tian Ye, Lichen Bai, Shuo Yang, Zeke Xie
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 |
|---|---|---|---|
| add the Select stage (diffusion difficulty plus interval sampling) to the Attach-only data (ablation) | D2C on DiT-L/2 · ImageNet 256 10K budget | gFID: 37.07 -> 14.96 | Improved |
| add visual representation alignment to Select (ablation) | D2C on DiT-L/2 · ImageNet 256 10K budget | gFID: 14.96 -> 10.37 | Improved |
| add DC embedding to Select (ablation) | D2C on DiT-L/2 · ImageNet 256 10K budget | gFID: 14.96 -> 9.01 | Improved |
| add both DC embedding and visual embedding to Select (ablation) | D2C on DiT-L/2 · ImageNet 256 10K budget | gFID: 14.96 -> 7.62 | Improved |
| add Select on top of Attach (ablation) | D2C on DiT-L/2 · ImageNet 256 10K budget | gFID: 8.79 -> 7.62 | Improved |
| replace one-hot class embedding alone with text plus class dual embedding (ablation) | D2C · ImageNet 256 10K budget | gFID | Improved |
| use interval sampling instead of lowest-loss or highest-loss selection (ablation) | D2C Select · ImageNet 10K budget | gFID-10K | Improved |
| replace MAE-L visual encoder with DINOv2-L (ablation) | D2C · ImageNet 256 10K budget | gFID: 9.23 -> 7.62 | Improved |
| use a pretrained scorer instead of a scorer trained from scratch (ablation) | D2C Select · ImageNet-1K SiT-XL/2 | gFID: 4.9 -> 4.3 | Improved |
| add Select to Attach-only training (ablation) | D2C on SiT-XL/2 · ImageNet-1K 40k iter | gFID: 5.6 -> 4.3 | Improved |
| replace random sampling with D2C | D2C · ImageNet 256 0.8% budget, 100k iter | gFID-50K: 35.86 -> 4.20 | Improved |
| replace random sampling with D2C at long training | D2C · ImageNet 256 0.8% budget, 300k iter DiT-L/2 | gFID-50K: 4.19 -> 4.13 | No clear effect |
Closest papers by meaning
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.91)
- Efficient Dataset Distillation via Minimax Diffusion (CVPR 2024 · similarity 0.91)
- DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents (ICML 2024 · similarity 0.91)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.90)
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.
