Research map / Video / Video generation
Phantom-Data: Towards a General Subject-Consistent Video Generation Dataset
Zhuowei Chen, Bingchuan Li, Tianxiang Ma, Lijie Liu, Mingcong Liu, Yunsheng Jiang, Gen Li, Xinghui Li, Liyang Chen, SiYu Zhou, Qian HE, Xinglong Wu
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 |
|---|---|---|---|
| train on Phantom-Data cross-pair instead of in-pair | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | Reward-TA: 2.074 -> 3.827 | Improved |
| train on Phantom-Data cross-pair instead of in-pair | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | GPT-4o subject score: 2.481 -> 3.041 | Improved |
| train on Phantom-Data cross-pair instead of in-pair | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | DINO: 0.478 -> 0.416 | Hurt |
| train on Phantom-Data cross-pair instead of in-pair | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | VBench IQ: 0.725 -> 0.739 | Improved |
| use general-subject cross-pair instead of face-only cross-pair | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | DINO: 0.354 -> 0.416 | Improved |
| add data augmentation to in-pair training | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | Reward-TA: 2.074 -> 2.427 | Improved |
| add data augmentation to in-pair training | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | DINO: 0.478 -> 0.473 | No clear effect |
| enrich training set with diverse subject types (humans, animals, products, multi-subject) (ablation) | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | subject consistency and prompt following | Improved |
| increase training set from 100k to 1 million samples (ablation) | Phantom-wan 1.3B · Subject-to-video test suite (100 cases) | all metrics | Improved |
| sample reference frames at minute-level instead of second-level intervals (ablation) | Phantom-Data pipeline · reference diversity | visual diversity | Improved |
| add LAION image retrieval to the video-only bank (ablation) | Phantom-Data pipeline · retrieval | recall and candidate diversity | Improved |
| add prior filtering and VLM identity verification (ablation) | Phantom-Data pipeline · pair quality | false positives | Improved |
Closest papers by meaning
- PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement (NeurIPS 2025 · similarity 0.87)
- OpenS2V-Nexus: A Detailed Benchmark and Million-Scale Dataset for Subject-to-Video Generation (NeurIPS 2025 · similarity 0.86)
- FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.86)
- Identity-Text Video Corpus Grounding (AAAI 2025 · similarity 0.86)
- VidProM: A Million-scale Real Prompt-Gallery Dataset for Text-to-Video Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Multi-subject Open-set Personalization in Video Generation (CVPR 2025 · similarity 0.84)
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.
