Research map / Continual learning / Online and replay-based
SCNS: Continual Personalization of Diffusion Models via Submodular Concept Neuron Selection
Zijie Peng, Enneng Yang, Yifei Cheng, Hongliang Yuan, Fei Ma, Xiaochun Cao, Li Shen
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 neuron selection (sparse update) instead of dense fine-tuning (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 26.4386 -> 0.3439 | Improved |
| use SCNS selection instead of random selection at matched sparsity (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 9.0304 -> 0.3439 | Improved |
| add cost-aware stopping to score-based selection (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 0.5954 -> 0.3439 | Improved |
| add cost-aware stopping to score-based selection (ablation) | SCNS · 8-concept continual personalization (final) | TFR-CLIP: 0.8116 -> 0.1183 | Improved |
| add cost-aware stopping to score-based selection (ablation) | SCNS · 8-concept continual personalization (final) | CLIP-T: 0.2830 -> 0.3041 | Improved |
| add task relevance score (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 5.3487 -> 0.3439 | Improved |
| add preservation score (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 2.4477 -> 0.3439 | Improved |
| add submodular span term (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 1.9114 -> 0.3439 | Improved |
| add zero-shot protection score (ablation) | SCNS · 8-concept continual personalization (final) | TFR-DINO: 0.5933 -> 0.3439 | Improved |
| add zero-shot protection score (ablation) | SCNS · 8-concept continual personalization (final) | CLIP-T: 0.2952 -> 0.3041 | Improved |
| add task relevance score (ablation) | SCNS · 8-concept continual personalization (final) | CLIP-T: 0.3026 -> 0.3041 | No clear effect |
| SCNS vs CNS magnitude-based neuron selection | SCNS · 8-concept continual personalization (final) | DINO-I: 0.5837 -> 0.6216 | Improved |
Closest papers by meaning
- How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization? (NeurIPS 2024 · similarity 0.87)
- Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.87)
- ConceptGuard: Continual Personalized Text-to-Image Generation with Forgetting and Confusion Mitigation (CVPR 2025 · similarity 0.86)
- APT: Adaptive Personalized Training for Diffusion Models with Limited Data (CVPR 2025 · similarity 0.85)
- Finding NeMo: Localizing Neurons Responsible For Memorization in Diffusion Models (NeurIPS 2024 · similarity 0.84)
- Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · similarity 0.84)
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.
