Research map / Image generation / Text-to-image diffusion
Customizable Image Synthesis with Multiple Subjects
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 |
|---|---|---|---|
| learn residual token embedding instead of a directly learned token embedding (ablation) | Cones 2 · Single subject | Text Alignment: 0.324 -> 0.330 | Helped |
| learn residual token embedding instead of a directly learned token embedding (ablation) | Cones 2 · Two subjects | Text Alignment: 0.291 -> 0.309 | Helped |
| learn residual token embedding instead of a directly learned token embedding (ablation) | Cones 2 · Three subjects | Image Alignment: 0.669 -> 0.689 | Helped |
| learn residual token embedding instead of a directly learned token embedding (ablation) | Cones 2 · Four subjects | Image Alignment: 0.651 -> 0.673 | Helped |
| replace joint-training Cones with composition of single-subject residual embeddings plus layout guidance | Cones · Four subjects | Text Alignment: 0.271 -> 0.299 | Helped |
| replace DreamBooth with the proposed method on single subjects | DreamBooth · Single subject | Image Alignment: 0.727 -> 0.725 | No clear effect |
| replace Cones with the proposed method on storage for four subjects | Cones · Four subjects | Storage and complexity: (7.75$\pm$ 0.56) MB -> 19.2 KB; $O(n^4)$ -> $O(n)$ | Helped |
| replace Cones with the proposed method in user study of image alignment (four subjects) | Cones · User study, four subjects | Image Alignment (% positive): 28.75$\%$ -> 59.08$\%$ | Helped |
| replace Cones with the proposed method in user study of text alignment (two subjects) | Cones · User study, two subjects | Text Alignment (% positive): 62.55$\%$ -> 77.87$\%$ | Helped |
| replace DreamBooth with the proposed method in user study of text alignment (four subjects) | DreamBooth · User study, four subjects | Text Alignment (% positive): 36.42$\%$ -> 77.35$\%$ | Helped |
| strengthen target subject only, without weakening irrelevant subjects (ablation) | layout guidance · Qualitative multi-subject generation | attribute binding | Hurt |
| weaken irrelevant subjects only, without strengthening target subject (ablation) | layout guidance · Qualitative multi-subject generation | subject presence | Hurt |
Closest papers by meaning
- Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects (CVPR 2025 · similarity 0.87)
- Multi-party Collaborative Attention Control for Image Customization (CVPR 2025 · similarity 0.87)
- MS-Diffusion: Multi-subject Zero-shot Image Personalization with Layout Guidance (ICLR 2025 · similarity 0.87)
- CustomContrast: A Multilevel Contrastive Perspective for Subject-Driven Text-to-Image Customization (AAAI 2025 · similarity 0.87)
- Harmonizing Visual and Textual Embeddings for Zero-Shot Text-to-Image Customization (AAAI 2025 · similarity 0.87)
- DynASyn: Multi-Subject Personalization Enabling Dynamic Action Synthesis (AAAI 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.