Research map / Continual learning / Class-incremental
Breaking the Synthetic-Real Domain Shortcut for Training-Free Generative Replay-based Class Incremental Learning
Tao Zhang, Qixuan Fan, Yiyuan Liang, Yanjie Wang, Song Yan, Tian Tian, Jiahuan Zhou, Luxin Yan, Sheng Zhong, Xu Zou
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 Dynamic Domain Centering (ablation) | raw generative replay · CIFAR-100 (T=10) | accuracy: 33.27 -> 41.02 | Improved |
| add Hierarchical Subspace Rectification (HSR) (ablation) | raw generative replay · CIFAR-100 (T=10) | accuracy: 33.27 -> 58.46 | Improved |
| add HSR on top of Centering (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 41.02 -> 67.69 | Improved |
| add RAPC on top of Centering and HSR (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 67.69 -> 69.73 | Improved |
| use hard projection gamma=1.0 instead of soft gamma=0.8 (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 69.73 -> 56.92 | Hurt |
| subspace dimension k=10 instead of 5 (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 69.73 -> 62.02 | Hurt |
| lambda_KD=10.0 instead of 1.0 (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 69.73 -> 63.78 | Hurt |
| add prototype loss lambda_Proto=0.10 (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 67.69 -> 69.73 | Improved |
| add task-constraint data filtering with Qwen-Image (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 68.67 -> 69.73 | Improved |
| add task-constraint data filtering with SD1.5 (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 68.08 -> 68.84 | Improved |
| swap Qwen-Image for SD1.5 generator (ablation) | DREAM · CIFAR-100 (T=10) | accuracy: 69.73 -> 68.84 | No clear effect |
| add DREAM to multi-source generative replay | raw generative replay · CIFAR-100 (T=10) | accuracy: 39.21 -> 68.06 | Improved |
Closest papers by meaning
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.87)
- Generative Multi-modal Models are Good Class Incremental Learners (CVPR 2024 · similarity 0.87)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.87)
- Autoencoder-Based Hybrid Replay for Class-Incremental Learning (ICML 2025 · similarity 0.87)
- Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · similarity 0.87)
- Dual-Consistency Model Inversion for Non-Exemplar Class Incremental Learning (CVPR 2024 · similarity 0.86)
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.
