Jingren Liu, Zhong Ji, Shuning Xu, Yun Wang, Xiangyu Chen
- [2026-03-13] Repository initialized.
- [2026-03-13] README released.
- [2026-xx-xx] Training code will be released.
- [2026-xx-xx] Evaluation code will be released.
- [2026-xx-xx] Reproducibility guide will be released.
This repository provides the official implementation for:
CCD: Continual Consistency Diffusion for Lifelong Generative Modeling
Diffusion models have shown remarkable performance in high-fidelity generation, but they remain highly vulnerable to Generative Catastrophic Forgetting (GCF) in continual learning scenarios. When trained on non-stationary task streams, newly learned generative abilities often overwrite previously acquired knowledge, causing severe degradation or even generative collapse.
To address this problem, we propose a principled continual diffusion framework consisting of:
- a standardized Continual Diffusion Generation (CDG) pipeline for training and evaluation,
- a Continual Consistency Diffusion (CCD) framework for stable lifelong generative modeling,
- a theoretical cross-task diffusion analysis that identifies the core factors behind forgetting in diffusion models.
Our method introduces three key consistency principles:
- Inter-task Knowledge Consistency (IKC) for aligning score predictions across tasks,
- Unconditional Knowledge Consistency (UKC) for preserving reverse-time denoising trajectories,
- Prior Knowledge Consistency (PKC) for maintaining semantic alignment in label/prior space.
By jointly enforcing these consistency objectives with rehearsal, CCD improves long-term generative retention and delivers strong continual generation performance across multiple benchmarks.
- A standardized continual diffusion generation pipeline for fair and reproducible evaluation
- A theory-driven continual diffusion framework grounded in cross-task SDE analysis
- Three complementary consistency objectives: IKC, UKC, and PKC
- A practical CCD training framework for mitigating generative catastrophic forgetting
- Strong continual generation results on MNIST, OxfordPets, CIFAR100, Flowers102, and CUB200
- A replay-based extension with Hierarchical Diversity Buffer (HDB) for more stable sample retention