Research map / Image generation / Text-to-image diffusion
Inserting Anybody in Diffusion Models via Celeb Basis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use celeb basis coefficients (1024 params) instead of Custom Diffusion | Stable Diffusion personalization · Synthetic StyleGAN faces | identity similarity: 0.1385 -> 0.2072 | Helped |
| use celeb basis instead of Custom Diffusion | Stable Diffusion personalization · Synthetic StyleGAN faces | Prompt CLIP score: 0.2608 -> 0.2545 | No clear effect |
| use celeb basis instead of Dreambooth | Stable Diffusion personalization · Synthetic StyleGAN faces | face detection rate: 54.76% -> 84.78% | Helped |
| use celeb basis instead of Textual Inversion | Stable Diffusion personalization · Synthetic StyleGAN faces | identity similarity / detection: 0.2958 -> 0.2072 identity; 92.86% -> 84.78% detection | Hurt |
| use celeb basis instead of baselines | Stable Diffusion personalization · User study (identity) | identity rank: 2.96 -> 3.37 (vs Custom Diffusion) | Helped |
| use celeb basis instead of Custom Diffusion | Stable Diffusion personalization · Synthetic StyleGAN faces | training time (min): 12 -> 3 | Helped |
| remove celeb basis (learn embedding directly from face encoder) (ablation) | Celeb Basis · Synthetic faces | concept combination | Hurt |
| build basis from 691 filtered names instead of 350 names (ablation) | Celeb Basis · Synthetic faces | generation quality | Helped |
| filter celeb names by SD generation quality (ablation) | Celeb Basis · Synthetic faces | interaction ability | Helped |
| use first/last name basis instead of flattened basis (ablation) | Celeb Basis · Synthetic faces | prompt understanding | Helped |
| set PCA dimension p=512 instead of 64, 256 or 768 (ablation) | Celeb Basis · Synthetic faces | identity similarity | Helped |
| add pretrained face recognition encoder F (ablation) | coefficient optimization · Synthetic faces | identity similarity | Helped |
Closest papers by meaning
- When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · similarity 0.86)
- MagicNaming: Consistent Identity Generation by Finding a “Name Space” in T2I Diffusion Models (AAAI 2025 · similarity 0.86)
- Cross Initialization for Face Personalization of Text-to-Image Models (CVPR 2024 · similarity 0.85)
- Energy-Guided Optimization for Personalized Image Editing with Pretrained Text-to-Image Diffusion Models (AAAI 2025 · similarity 0.84)
- Multi-focal Conditioned Latent Diffusion for Person Image Synthesis (CVPR 2025 · similarity 0.84)
- JeDi: Joint-Image Diffusion Models for Finetuning-Free Personalized Text-to-Image Generation (CVPR 2024 · 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.