Research map / Image generation / Text-to-image diffusion
Generating Multi-Image Synthetic Data for Text-to-Image Customization
Nupur Kumari, Xi Yin, Jun-Yan Zhu, Ishan Misra, Samaneh Azadi
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 normalized image/text guidance inference (ablation) | IP-Adapter Plus · DreamBooth (1 reference image) | TIFA: 0.615 -> 0.816 | Improved |
| add normalized guidance inference (ablation) | IP-Adapter Plus · DreamBooth (1 reference image) | MDINOv2-I (property change): 0.737 -> 0.668 | Hurt |
| add normalized guidance instead of guidance rescale (ablation) | IP-Adapter Plus · DreamBooth (1 reference image) | Geometric score: 0.710 -> 0.756 | Improved |
| fine-tune on SynCD dataset (ablation) | IP-Adapter Plus · DreamBooth (1 reference image) | Geometric score: 0.756 -> 0.819 | Improved |
| add masked shared attention reference conditioning (ablation) | IP-Adapter Plus · DreamBooth (1 reference image) | Geometric score: 0.819 -> 0.825 | Improved |
| use 3 reference images instead of 1 at inference (ablation) | Ours (3B) · DreamBooth | Geometric score: 0.825 -> 0.838 | Improved |
| add masked shared attention in dataset generation (ablation) | SynCD deformable · SynCD dataset curation | DINOv2-I: 0.626 -> 0.700 | Improved |
| add detailed object descriptions to prompts (ablation) | SynCD deformable · SynCD dataset curation | DINOv2-I: 0.564 -> 0.700 | Improved |
| add feature warping with 3D assets (ablation) | SynCD rigid · SynCD dataset curation | DINOv2-I: 0.572 -> 0.598 | Improved |
| add MSA on top of depth guidance (ablation) | SynCD rigid · SynCD dataset curation | DINOv2-I: 0.495 -> 0.572 | Improved |
| increase training dataset from 100 to 95K samples (ablation) | Ours (12B) · DreamBooth | MDINOv2-I: 0.736 -> 0.774 | Improved |
| increase training dataset from 100 to 95K samples (ablation) | Ours (3B) · DreamBooth | MDINOv2-I: 0.790 -> 0.813 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OminiControl (12B) — OminiControl: Minimal and Universal Control for Diffusion Transformer (ICCV 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Customization Assistant for Text-to-Image Generation (CVPR 2024 · similarity 0.87)
- Image Captioning with Multi-Context Synthetic Data (AAAI 2024 · similarity 0.87)
- Direct Consistency Optimization for Robust Customization of Text-to-Image Diffusion models (NeurIPS 2024 · similarity 0.87)
- JeDi: Joint-Image Diffusion Models for Finetuning-Free Personalized Text-to-Image Generation (CVPR 2024 · similarity 0.87)
- ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models (CVPR 2024 · similarity 0.86)
- PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding (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.
