Research map / Video / Video generation
CustomX: Unified Character, Action, and Scene Customization in Video World Models
Yitong Wang, Fangyun Wei, Hongyang Zhang, Bo Dai, Yan Lu
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 back view to the front-view character condition (ablation) | CustomX · Character consistency | DINOv2 score: 0.556 -> 0.613 | Improved |
| add back view to the front-view character condition (ablation) | CustomX · Character consistency | CLIP score: 0.628 -> 0.678 | Improved |
| add left and right views to front-back character condition (ablation) | CustomX · Character consistency | DINOv2 score: 0.613 -> 0.698 | Improved |
| add left and right views to front-back character condition (ablation) | CustomX · Character consistency | CLIP score: 0.678 -> 0.721 | Improved |
| add character anchor (first-frame only) (ablation) | CustomX · Character consistency | DINOv2 score: 0.477 -> 0.597 | Improved |
| propagate anchor to per-frame anchors (ablation) | CustomX · Character consistency | DINOv2 score: 0.597 -> 0.698 | Improved |
| propagate anchor to per-frame anchors (ablation) | CustomX · Character consistency | CLIP score: 0.645 -> 0.721 | Improved |
| add per-frame anchor versus no anchor (ablation) | CustomX · Character consistency | CLIP score: 0.529 -> 0.721 | Improved |
| add real-world videos to GTA-V training data (ablation) | CustomX · Character consistency (real-world characters) | DINOv2 score: 0.686 -> 0.755 | Improved |
| add real-world videos to GTA-V training data (ablation) | CustomX · Character consistency (real-world characters) | CLIP score: 0.718 -> 0.729 | Improved |
| distill 30-step model to 4 steps with DMD2 | CustomX · Character consistency | DINOv2 score: 0.698 -> 0.669 | Hurt |
| distill 30-step model to 4 steps with DMD2 | CustomX · Aesthetic quality | CLIP-Aesthetic: 5.665 -> 5.583 | Hurt |
Closest papers by meaning
- From Zero to Hero: Training-Free Custom Concept Spawning in World Models (NeurIPS 2026 · similarity 0.90)
- Genie: Generative Interactive Environments (ICML 2024 · similarity 0.88)
- Exploration-Driven Generative Interactive Environments (CVPR 2025 · similarity 0.88)
- LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE (NeurIPS 2026 · similarity 0.87)
- World Models as Group Actions (NeurIPS 2026 · similarity 0.87)
- WorldWeaver: Generating Long-Horizon Video Worlds via Rich Perception (NeurIPS 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.
