Research map / Image generation / Text-to-image diffusion
Preserve and Personalize: Personalized Text-to-Image Diffusion Models without Distributional Drift
Gihoon Kim, Hyungjin Park, Taesup Kim
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 Lipschitz-based L2 regularization toward pretrained weights (replacing prior preservation) | DreamBooth (SD-1.5) · Personalization benchmark (30 subjects) | DINO: 0.6028 -> 0.6394 | Improved |
| add Lipschitz-based regularization | DreamBooth (SD-1.5) · Personalization benchmark (30 subjects) | CLIP-T: 0.2793 -> 0.2976 | Improved |
| add Lipschitz-based regularization | DB-LoRA (SD-1.5) · Personalization benchmark (30 subjects) | CLIP-T: 0.3108 -> 0.2975 | Hurt |
| add Lipschitz-based regularization | DB-LoRA (SD-1.5) · Personalization benchmark (30 subjects) | DINO: 0.5776 -> 0.6038 | Improved |
| add Lipschitz-based regularization | Custom Diffusion (SD-1.5) · Personalization benchmark (30 subjects) | DINO: 0.5568 -> 0.5638 | Improved |
| add Lipschitz-based regularization | Custom Diffusion (SD-XL) · Personalization benchmark (30 subjects) | DINO: 0.5337 -> 0.5493 | Improved |
| add Lipschitz-based regularization | DB-LoRA (SD-XL) · Personalization benchmark (30 subjects) | CLIP-T: 0.3099 -> 0.3014 | Hurt |
| add Lipschitz-based regularization | Custom Diffusion (SD-3.0) · Personalization benchmark (30 subjects) | CLIP-T: 0.3031 -> 0.2955 | Hurt |
| add Lipschitz-based regularization | DB-LoRA (SD-3.0) · Personalization benchmark (30 subjects) | CLIP-I: 0.7823 -> 0.7949 | Improved |
| increase regularization weight lambda (ablation) | Lipschitz-regularized DreamBooth · Personalization benchmark (30 subjects) | DINO/CLIP-I vs CLIP-T: higher lambda raises CLIP-T, lower lambda raises DINO and CLIP-I | Mixed |
| lower DreamBooth learning rate to 2e-6 instead of using our regularizer (ablation) | DreamBooth (SD-1.5) · Personalization benchmark (30 subjects) | DINO: 0.6028 -> 0.5707 | Hurt |
| add DreamBooth prior-preservation loss on top of Lipschitz regularization (ablation) | Lipschitz-regularized DreamBooth · Personalization benchmark (30 subjects) | quality | No clear effect |
Closest papers by meaning
- PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction (ICLR 2025 · similarity 0.93)
- Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.91)
- APT: Adaptive Personalized Training for Diffusion Models with Limited Data (CVPR 2025 · similarity 0.91)
- Direct Consistency Optimization for Robust Customization of Text-to-Image Diffusion models (NeurIPS 2024 · similarity 0.91)
- Efficient Personalization of Quantized Diffusion Model without Backpropagation (CVPR 2025 · similarity 0.90)
- How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization? (NeurIPS 2024 · similarity 0.90)
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.
