Research map / Video / Video generation
LibraGen: Playing a Balance Game in Subject-Driven Video Generation
Jiahao Zhu, Shanshan Lao, Lijie Liu, Gen Li, Tianhao Qi, hanwei hanwei, Bingchuan Li, FangfangLiu FangfangLiu, Zhuowei Chen, Tianxiang Ma, Qian HE, Yi Zhou and 1 more
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 |
|---|---|---|---|
| merge in-pair LoRA at 0.15 (ablation) | cross-pair SFT baseline · In-house S2V benchmark (cumulative strategy ablation) | GSB ratio: GSB 0.190 | Improved |
| merge in-pair LoRA at 0.15 (ablation) | cross-pair SFT baseline · In-house S2V benchmark (cumulative strategy ablation) | VQ: 2.794 -> 2.708 | Hurt |
| merge in-pair LoRA at 0.15 (ablation) | cross-pair SFT baseline · In-house S2V benchmark (cumulative strategy ablation) | AES: 0.6420 -> 0.6306 | Hurt |
| add Consis-DPO LoRA (0.5) (ablation) | SFT + in-pair merge · In-house S2V benchmark (cumulative strategy ablation) | GSB ratio: GSB 0.286 | Improved |
| add Consis-DPO LoRA (0.5) (ablation) | SFT + in-pair merge · In-house S2V benchmark (cumulative strategy ablation) | TA: 3.466 -> 3.490 | Improved |
| add Real-Fake DPO LoRA (0.1) (ablation) | SFT + Consis-DPO · In-house S2V benchmark (cumulative strategy ablation) | VQ: 2.714 -> 2.756 | Improved |
| add Real-Fake DPO LoRA (0.1) (ablation) | SFT + Consis-DPO · In-house S2V benchmark (cumulative strategy ablation) | AES: 0.6334 -> 0.6394 | Improved |
| add Real-Fake DPO LoRA (0.1) (ablation) | SFT + Consis-DPO · In-house S2V benchmark (cumulative strategy ablation) | GSB ratio: GSB 0.025 | No clear effect |
| add time-dependent dynamic CFG (ablation) | SFT + both DPO LoRAs · In-house S2V benchmark (cumulative strategy ablation) | TA: 3.524 -> 3.594 | Improved |
| add time-dependent dynamic CFG (ablation) | SFT + both DPO LoRAs · In-house S2V benchmark (cumulative strategy ablation) | AES: 0.6394 -> 0.6496 | Improved |
| replace MAGREF with LibraGen on subject consistency | LibraGen · In-house S2V benchmark (200 cases) | GSB ratio (4 reference images): 0.700 | Improved |
| replace Kling O1 with LibraGen on subject consistency | LibraGen · In-house S2V benchmark (200 cases) | GSB ratio (1 reference image): 0.077 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MAGREF — MAGREF: Masked Guidance for Any-Reference Video Generation with Subject Disentanglement (ICLR 2026) · better on 6 of 6 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
- OpenS2V-Nexus: A Detailed Benchmark and Million-Scale Dataset for Subject-to-Video Generation (NeurIPS 2025 · similarity 0.86)
- VBench: Comprehensive Benchmark Suite for Video Generative Models (CVPR 2024 · similarity 0.86)
- T2V-Turbo-v2: Enhancing Video Model Post-Training through Data, Reward, and Conditional Guidance Design (ICLR 2025 · similarity 0.85)
- T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback (NeurIPS 2024 · similarity 0.85)
- Video-Bench: Human-Aligned Video Generation Benchmark (CVPR 2025 · similarity 0.85)
- Generative Inbetweening through Frame-wise Conditions-Driven Video Generation (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.
