# LibraGen: Playing a Balance Game in Subject-Driven Video Generation

Source: [LibraGen: Playing a Balance Game in Subject-Driven Video Generation](<https://www.sota2.com/research/paper/69b8e00e0636e5bf4afbea6e>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-19T05:00:05.030Z

Authors: Jiahao Zhu, Shanshan Lao, Lijie Liu, Gen Li, Tianhao Qi, Wei Han, Bingchuan Li, Fangfang Liu, Zhuowei Chen, Tianxiang Ma, Qian HE, Yi Zhou, Xiaohua Xie

Published: 2026-03-13T18:36:23.000Z

## Abstract

With the advancement of video generation foundation models (VGFMs), customized generation, particularly subject-to-video (S2V), has attracted growing attention. However, a key challenge lies in balancing the intrinsic priors of a VGFM, such as motion coherence, visual aesthetics, and prompt alignment, with its newly derived S2V capability. Existing methods often neglect this balance by enhancing one aspect at the expense of others. To address this, we propose LibraGen, a novel framework that views extending foundation models for S2V generation as a balance game between intrinsic VGFM strengths and S2V capability. Specifically, guided by the core philosophy of "Raising the Fulcrum, Tuning to Balance," we identify data quality as the fulcrum and advocate a quality-over-quantity approach. We construct a hybrid pipeline that combines automated and manual data filtering to improve overall data quality. To further harmonize the VGFM's native capabilities with its S2V extension, we introduce a Tune-to-Balance post-training paradigm. During supervised fine-tuning, both cross-pair and in-pair data are incorporated, and model merging is employed to achieve an effective trade-off. Subsequently, two tailored direct preference optimization (DPO) pipelines, namely Consis-DPO and Real-Fake DPO, are designed and merged to consolidate this balance. During inference, we introduce a time-dependent dynamic classifier-free guidance scheme to enable flexible and fine-grained control. Experimental results demonstrate that LibraGen outperforms both open-source and commercial S2V models using only thousand-scale training data.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Subject-driven Video Generation on Subject-to-Video (S2V) (test)](<https://www.sota2.com/research/sota/subject-driven-video-generation-on-subject-to-video-s2v-test.md>): Subject-driven video generation; Subject-to-Video (S2V) (test); MS: 0.538.

- [Subject-driven Video Generation on S2V 1 reference image (val)](<https://www.sota2.com/research/sota/subject-driven-video-generation-on-s2v-1-reference-image-val.md>): Subject-driven video generation; S2V 1 reference image (val); GSB Ratio: 42.3.

- [Subject-driven Video Generation on S2V 2 reference images (evaluation set)](<https://www.sota2.com/research/sota/subject-driven-video-generation-on-s2v-2-reference-images-evaluation-set.md>): Subject-driven video generation; S2V 2 reference images (evaluation set); GSB Ratio: 0.62.

- [Subject-driven Video Generation on S2V 3 reference images (test)](<https://www.sota2.com/research/sota/subject-driven-video-generation-on-s2v-3-reference-images-test.md>): Subject-driven video generation; S2V 3 reference images (test); GSB Ratio: 0.429.

- [Subject-driven Video Generation on S2V 4 reference images (test)](<https://www.sota2.com/research/sota/subject-driven-video-generation-on-s2v-4-reference-images-test.md>): Subject-driven video generation; S2V 4 reference images (test); GSB Ratio: 70.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.13506>)

[PDF](<https://arxiv.org/pdf/2603.13506.pdf>)
