Research map / Image generation / Fast diffusion sampling
Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models
Xiao Liu, Xiaoliu Guan, Yu Wu, Jiaxu Miao
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 IET-AGC training to DDPM | DDPM · CIFAR-10 | MQ0.4: 111 -> 14 | Improved |
| add IET-AGC training to DDPM | DDPM · CIFAR-10 | FID: 8.81 -> 8.34 | Improved |
| train on images with added Gaussian noise | DDPM · CIFAR-10 | FID: 8.81 -> 94.61 | Hurt |
| use Ambient Diffusion instead of IET-AGC | DDPM · CIFAR-10 | FID: 8.81 -> 11.7 | Hurt |
| add iterative ensemble training (IET) with IID shards (ablation) | DDPM · CIFAR-10 | MQ0.5 and FID: MQ0.5 465 -> 382; FID 8.81 -> 7.61 | Improved |
| split data in a non-IID way for IET (ablation) | IET · CIFAR-10 | FID: 7.61 -> 12.5 | Hurt |
| add anti-gradient control (AGC) alone (ablation) | DDPM · CIFAR-10 | MQ0.5 and FID: MQ0.5 465 -> 154; FID 8.81 -> 11.36 | Mixed |
| combine IET with AGC (ablation) | AGC · CIFAR-10 | FID and MQ0.5: FID 11.36 -> 8.34; MQ0.5 154 -> 117 | Improved |
| increase skipping threshold lambda from 0.5 to 0.8 (ablation) | IET-AGC · CIFAR-10 | MQ0.5 and FID: MQ0.5 117 -> 45; FID 8.34 -> 10.43 | Mixed |
| increase number of shards from 10 to 15 (ablation) | IET-AGC · CIFAR-10 | MQ0.5 and FID: MQ0.5 117 -> 169; FID 8.34 -> 9.1 | Hurt |
| apply IET-AGC when fine-tuning Stable Diffusion | Stable Diffusion · LAION-10k | similarity score: 0.64 -> 0.37 | Improved |
| apply IET-AGC when fine-tuning a pre-trained DDPM for 2 epochs | pre-trained DDPM · CIFAR-10 | MQ0.4 and FID: MQ0.4 111 -> 83; FID 8.81 -> 7.93 | Improved |
Closest papers by meaning
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.90)
- Towards Memorization-Free Diffusion Models (CVPR 2024 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.88)
- How Diffusion Models Memorize (NeurIPS 2026 · 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.
