Research map / Image generation / Text-to-image diffusion
MUSE: Multi-Subject Unified Synthesis via Explicit Layout Semantic Expansion
Fei Peng, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, Huiyuan Fu
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 |
|---|---|---|---|
| replace decoupled cross-attention with concatenated cross-attention for layout control (ablation) | MUSE layout stage on SDXL · MIG Bench | layout success rate (avg): 0.781 -> 0.884 | Improved |
| use concatenated cross-attention instead of DCA (ablation) | MUSE · MIG Bench | CLIP-T: 0.322 -> 0.331 | Improved |
| use progressive two-stage training instead of DCA for both layout and subject (ablation) | MUSE · MS-Bench-Random | SR-0.65: -0.410 relative to full method | Improved |
| use progressive two-stage training instead of single-stage CCA+DCA (ablation) | MUSE · MS-Bench-Random | CLIP-I-local: -0.085 relative to full method | Improved |
| train layout first then subject (progressive) instead of jointly in stage 2 (ablation) | MUSE · MS-Bench-Random | SR-0.65: -0.043 relative to full method | Improved |
| train layout control before subject synthesis instead of reversed order (ablation) | MUSE · MS-Bench-Random | CLIP-I-local: -0.056 relative to full method | Improved |
| add bounding box to image tokens rather than concatenating (ablation) | MUSE grounding tokens · MS-Bench-Random | CLIP-I-local: -0.015 relative to full method | Improved |
| increase subject synthesis strength lambda from 0.8 to 1.0 (ablation) | MUSE · MS-Bench-Random | CLIP-T: 0.321 -> 0.310 | Mixed |
| lower subject synthesis strength lambda from 0.8 to 0.6 (ablation) | MUSE · MS-Bench-Random | SR-0.65: 0.755 -> 0.712 | Hurt |
| randomize subject layout instead of fixed layout | MS-Diffusion · MS-Bench | SR-0.65: 0.706 -> 0.289 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- MS-Diffusion — MS-Diffusion: Multi-subject Zero-shot Image Personalization with Layout Guidance (ICLR 2025) · better on 5 of 5 shared results
- InstanceDiffusion — InstanceDiffusion: Instance-level Control for Image Generation (CVPR 2024) · better on 3 of 3 shared results
- MIGC — MIGC: Multi-Instance Generation Controller for Text-to-Image Synthesis (CVPR 2024) · better on 2 of 3 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
- PLACE: Adaptive Layout-Semantic Fusion for Semantic Image Synthesis (CVPR 2024 · similarity 0.90)
- Zero-Painter: Training-Free Layout Control for Text-to-Image Synthesis (CVPR 2024 · similarity 0.89)
- SSMG: Spatial-Semantic Map Guided Diffusion Model for Free-Form Layout-to-Image Generation (AAAI 2024 · similarity 0.88)
- XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiT Modulation (NeurIPS 2025 · similarity 0.88)
- Grounded Text-to-Image Synthesis with Attention Refocusing (CVPR 2024 · similarity 0.88)
- SpotActor: Training-Free Layout-Controlled Consistent Image Generation (AAAI 2025 · similarity 0.88)
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.
