Research map / Image generation / Text-to-image diffusion
MultiCrafter: High-Fidelity Multi-Subject Generation via Disentangled Attention and Identity-Aware Preference Alignment
Tao Wu, Yibo Jiang, Yehao Lu, Zhizhong Wang, Zeyi Huang, Zequn Qin, Xi Li
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 identity-disentangled attention regularization L_attn (single LoRA) (ablation) | UNO baseline on FLUX · Multi-human benchmark | Face-Sim: 0.1474 -> 0.4983 | Improved |
| add L_attn with single LoRA (ablation) | UNO baseline on FLUX · Multi-human benchmark | CLIP-T: 0.2645 -> 0.2637 | Hurt |
| add L_attn with single LoRA (ablation) | UNO baseline on FLUX · Multi-human benchmark | AES (HPSv2): 0.2954 -> 0.2653 | Hurt |
| add MoE-LoRA on top of L_attn (ablation) | MultiCrafter pre-training · Multi-human benchmark | Face-Sim: 0.4983 -> 0.5154 | Improved |
| add MoE-LoRA on top of L_attn (ablation) | MultiCrafter pre-training · Multi-human benchmark | CLIP-T: 0.2637 -> 0.2674 | Improved |
| add IPPO post-training (online RL with Multi-ID reward) (ablation) | MultiCrafter · Multi-human benchmark | AES (HPSv2): 0.2661 -> 0.2915 | Improved |
| add IPPO post-training (ablation) | MultiCrafter · Multi-human benchmark | CLIP-T: 0.2674 -> 0.2753 | Improved |
| add IPPO post-training (ablation) | MultiCrafter · Multi-human benchmark | Face-Sim: 0.5154 -> 0.5284 | Improved |
| retrain UNO on the authors' video-derived multi-human data (ablation) | UNO · Multi-human benchmark | AES (HPSv2): 0.2954 -> 0.2656 | Hurt |
| retrain UNO on the authors' video-derived multi-human data (ablation) | UNO · Multi-human benchmark | Face-Sim: 0.1474 -> 0.2879 | Improved |
| MultiCrafter vs next-best method on multi-human fidelity | MultiCrafter · Multi-human benchmark | Face-Sim: 0.4117 -> 0.5284 | Improved |
| MultiCrafter vs OmniGen2 on aesthetics | MultiCrafter · Multi-human benchmark | AES (HPSv2): 0.3103 -> 0.2915 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- XVerse — XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiT Modulation (NeurIPS 2025) · better on 3 of 3 shared results
- OmniGen — OmniGen: Unified Image Generation (CVPR 2025) · better on 1 of 1 shared result
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
- Learning from Noisy Preferences: A Semi-Supervised Learning Approach to Direct Preference Optimization (ICLR 2026 · similarity 0.87)
- FocusDPO: Dynamic Preference Optimization for Multi-Subject Personalized Image Generation via Adaptive Focus (AAAI 2026 · similarity 0.87)
- OneActor++: Consistent Multi-Subject Generation via Cluster-Conditioned Guidance (TPAMI 2026 · similarity 0.87)
- PrefPaint: Aligning Image Inpainting Diffusion Model with Human Preference (NeurIPS 2024 · similarity 0.87)
- Preference-Enhanced Reinforcement Learning for Pluralistic Image Inpainting (ICML 2026 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 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.
