Research map / Video / Video generation
ID-Crafter: VLM-Grounded Online RL for Compositional Multi-Subject Video Generation
Panwang Pan, Jingjing Zhao, Yuchen Lin, Chenguo Lin, Chenxin Li, Hengyu Liu, Tingting Shen, Yadong Mu
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 hierarchical identity-preserving attention (vs plain cross-attention) (ablation) | ID-Crafter 1.3B · OpenS2V | FaceSim: 51.34% -> 58.12% | Improved |
| add Qwen2.5-VL encoder alongside T5 (ablation) | ID-Crafter 1.3B · OpenS2V | Q-Align: 0.287 -> 0.351 | Improved |
| add curated training data pipeline (ablation) | ID-Crafter 1.3B · OpenS2V | Total Score: 48.78% -> 54.33% | Improved |
| add online GRPO post-training instead of SFT only (ablation) | ID-Crafter 1.3B · OpenS2V | FaceSim: 58.12% -> 66.10% | Improved |
| online GRPO instead of offline DPO (ablation) | ID-Crafter 1.3B · OpenS2V | FaceSim: 62.35% -> 66.10% | Improved |
| add fidelity reward R_fid (ablation) | online GRPO · OpenS2V | FaceSim: 45.32% -> 66.10% | Improved |
| add natural reward R_nat (ablation) | online GRPO · OpenS2V | Q-Align: Q-Align 0.361 -> 0.410; FaceSim 69.30% -> 66.10% | Mixed |
| scale ID-Crafter from 1.3B to 14B | ID-Crafter · OpenS2V | Total Score: 55.16% -> 57.05% | Improved |
Closest papers by meaning
- PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement (NeurIPS 2025 · similarity 0.90)
- ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation (NeurIPS 2026 · similarity 0.89)
- CustomCrafter: Customized Video Generation with Preserving Motion and Concept Composition Abilities (AAAI 2025 · similarity 0.86)
- Illuminating Visual Identity in Universal Multimodal Embeddings (CVPR 2026 · similarity 0.86)
- Identity-Preserving Text-to-Video Generation by Frequency Decomposition (CVPR 2025 · similarity 0.86)
- VLOGGER: Multimodal Diffusion for Embodied Avatar Synthesis (CVPR 2025 · similarity 0.85)
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.
