Research map / Continual learning / Online and replay-based
Dynamic Expansion Diffusion Learning for Lifelong Generative Modelling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the Dynamic Expansion Diffusion Model (DEDM), a mixture of DDPM components that grows and prunes under online task-free continual learning, with a VAE student trained by distillation for representation learning.
Problem
Diffusion models trained on a continuously changing data stream without task boundaries suffer catastrophic forgetting; existing dynamic-expansion generative continual learners rely on GANs or VAEs with mode collapse or poor generation quality.
Approach
Expand a mixture of DDPM components when a new batch's Gaussian-smoothed distribution (forward diffusion at a small step, compared via JS divergence) differs from stored components' memory distributions; prune components whose knowledge overlaps most; distill the mixture into a VAE student.
Key results
- On Split MNIST, DEDM achieves FID 34.26 versus 46.47 for the best CGKD-VAE baseline; average FID over four split datasets is 56.23 vs 84.92 for Reservoir-DDPM.
- Average reconstruction PSNR of 20.62 and SSIM of 0.90 for DEDM on class-incremental splits, exceeding the other compared methods.
- On high-resolution 256x256 generation, DEDM reaches FID 61.52 on CelebA-HQ versus 168.52 for CGKD-GAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DEDM instead of the strongest listed baseline | OTFCL image generation · Split MNIST, Split Fashion, Split SVHN, and Split CIFAR10 | FID: 34.26 vs 48.57; 53.78 vs 107.92; 54.79 vs 66.24; 82.12 vs 112.29 | Helped |
| Use DEDM instead of the strongest listed baseline | OTFCL image reconstruction · Split MNIST, Split Fashion, Split SVHN, and Split CIFAR10 | PSNR: 21.09 vs 21.03; 20.01 vs 19.90; 22.72 vs 21.14; 18.66 vs 18.52 | Mixed |
| Use DEDM instead of the strongest listed baseline | OTFCL image reconstruction · Split MNIST, Split Fashion, Split SVHN, and Split CIFAR10 | SSIM: 0.93 vs 0.94; 0.92 vs 0.92; 0.91 vs 0.86; 0.84 vs 0.81 | Mixed |
| Use DEDM instead of the strongest listed baseline | OTFCL image generation · CelebA-Chair, CelebA-ImageNet, and S-MINIImageNet | FID: 69.52 vs 131.81; 116.77 vs 123.27; 140.19 vs 177.89 | Helped |
| Decrease the expansion threshold γ (ablation) | DEDM · Split MNIST | number of components | Mixed |
| Use the proposed expansion criterion when data distributions change | DEDM · Split MNIST | number of components | Helped |
Surprising findings
- The GAN-based teacher and VAE student baselines do not perform consistently across datasets, attributed to adversarial instability.
Limitations
- Expansion and pruning rely on a threshold gamma that must be tuned, with the number of components sensitive to it.
- Evaluation is limited to small-to-medium image datasets and generation quality is measured mainly by FID.
Experimental setup
Datasets: MNIST Fashion-MNIST SVHN CIFAR-10 CelebA 3D-Chair ImageNet Mini-ImageNet CelebA-HQ CACD FFHQ
Benchmarks: Split MNIST Split Fashion Split SVHN Split CIFAR10 CelebA-Chair CelebA-ImageNet S-MiniImageNet CelebA-HQ CACD FFHQ
Metrics: FID PSNR SSIM
Baselines: CNDPM CGKD-GAN CGKD-VAE CGKD-WVAE Reservoir-VAE Reservoir-DDPM
Builds on: DDPM Improved DDPM CN-DPM CGKD VAE
Closest papers by meaning
- Lifelong Scalable Generative System via Online Maximum Mean Discrepancy (AAAI 2025 · similarity 0.90)
- Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · similarity 0.90)
- Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · similarity 0.89)
- DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture (CVPR 2025 · similarity 0.88)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.88)
- Random Conditioning for Diffusion Model Compression with Distillation (CVPR 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.