Research map / Image generation / Text-to-image diffusion
Directional Textual Inversion for Personalized Text-to-Image Generation
Kunhee Kim, NaHyeon Park, Kibeom Hong, Hyunjung Shim
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 |
|---|---|---|---|
| fix in-distribution norm and optimize only direction (DTI vs TI) | Textual Inversion on SDXL · DreamBooth, SigLIP text alignment | text alignment: 0.292 -> 0.522 | Improved |
| DTI vs TI on image similarity | Textual Inversion on SDXL · DreamBooth, DINOv2 | image similarity: 0.561 -> 0.450 | Hurt |
| DTI vs TI on SANA 1.5-4.8B | Textual Inversion on SANA · DreamBooth | text alignment: 0.646 -> 0.757 | Improved |
| DTI vs TI on SANA 1.5-1.6B image similarity | Textual Inversion on SANA · DreamBooth, DINOv2 | image similarity: 0.480 -> 0.479 | No clear effect |
| rescale TI embeddings to in-distribution norm after training (ablation) | Textual Inversion on SDXL · DreamBooth | text alignment: 0.292 -> 0.466 | Improved |
| rescale TI embeddings to in-distribution norm after training (ablation) | Textual Inversion on SDXL · DreamBooth | image similarity: 0.561 -> 0.243 | Hurt |
| use RSGD instead of AdamW with projection (ablation) | DTI · DreamBooth (SDXL) | text alignment: 0.463 -> 0.522 | Improved |
| use mean vocabulary norm instead of minimum norm (ablation) | DTI · DreamBooth (SDXL) | text alignment: 0.074 -> 0.522 | Improved |
| use mean vocabulary norm instead of OOD norm 5.0 (ablation) | DTI · DreamBooth (SDXL) | text alignment: 0.373 -> 0.522 | Improved |
| add vMF directional prior (kappa 1e-4 vs 0) (ablation) | DTI · DreamBooth (SDXL) | text alignment: 0.436 -> 0.522 | Improved |
| add vMF directional prior (kappa 1e-4 vs 0) (ablation) | DTI · DreamBooth (SDXL) | image similarity: 0.507 -> 0.450 | Hurt |
| increase vMF concentration to 0.5e-3 (ablation) | DTI · DreamBooth (SDXL) | image similarity: 0.450 -> 0.278 | Hurt |
Closest papers by meaning
- AttnDreamBooth: Towards Text-Aligned Personalized Text-to-Image Generation (NeurIPS 2024 · similarity 0.90)
- CoRe: Context-Regularized Text Embedding Learning for Text-to-Image Personalization (AAAI 2025 · similarity 0.88)
- PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction (ICLR 2025 · similarity 0.87)
- Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions (CVPR 2025 · similarity 0.86)
- Discriminative Probing and Tuning for Text-to-Image Generation (CVPR 2024 · similarity 0.86)
- Visually Guided Decoding: Gradient-Free Hard Prompt Inversion with Language Models (ICLR 2025 · 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.
