让视频模型学得更好更快：Linum公司披露数据过滤流水线演进

Linum公司的两兄弟团队目前正在从头训练文本到视频生成模型，并发布了技术博客，介绍了训练文本到视频生成模型时数据过滤流水线的演进历程 1。该团队的数据处理方案经历了从2024年使用低成本传统计算机视觉算法，到2025年初使用数千GPU运行微调大语言模型的显著升级 1。

在2024年的初始阶段，团队在CPU上运行启发式规则，使用PySceneDetect进行场景检测，使用EAST Detector进行文本检测，使用Haar-cascade进行人脸检测 1。2025年初，流水线迁移至数千GPU，部署在Nvidia A10Gs和L4s上，使用AutoShot和TransNetV2替代PySceneDetect，使用PaddleOCR+TensorRT替代EAST Detector 1。此外，团队使用SFT微调Qwen-2-VL-2B作为类别分类器，图像类别包括劣质产品图、图表截图、拼贴画和带水印图像等 1。在美学评估方面，团队使用RLVR（带有可验证奖励的强化学习）训练Qwen-2.5-VL-3B进行美学评分，采用GSPO算法，评分范围为1至4分 1。

最终，团队将约150亿张图像候选池过滤至约2.5亿张（1.7%），将约10亿个视频片段候选池过滤至约5000万个（5%）用于训练 1。同时，团队使用SOTA光流预测器WAFT过滤长尾低运动视频 1。CMU 2025年研究显示，电影摄影专家在标注镜头运动时与'ground truth'的分歧率约24%，经多次迭代才能达到96%一致率 1。

Linum, a company with a two-brother team currently training a text-to-video model from scratch, has published a technical blog detailing the evolution of its data filtering pipeline for training text-to-video generation models 1. The team has transitioned from running heuristic rules on CPUs using low-cost traditional computer vision algorithms in 2024 to deploying fine-tuned large language models on thousands of GPUs for category classification and aesthetic scoring in 2025 1.

In 2024, the pipeline relied on CPU-based heuristic rules, utilizing PySceneDetect for scene detection, the EAST Detector for text detection, and Haar-cascade for face detection 1. By early 2025, the infrastructure was upgraded to thousands of Nvidia A10G and L4 GPUs 1. This migration replaced PySceneDetect with AutoShot and TransNetV2, while the EAST Detector was swapped for PaddleOCR combined with TensorRT 1.

To further refine the data, the team employed fine-tuned large language models for advanced filtering tasks 1. They used supervised fine-tuning to adapt Qwen-2-VL-2B as a category classifier to filter out content types such as ugly product images, diagrams, screenshots, collages, and watermarked images 1. Additionally, Reinforcement Learning with Verifiable Rewards and the GSPO algorithm were utilized to train Qwen-2.5-VL-3B for aesthetic scoring on a scale of 1 to 4 1. This rigorous filtering process reduced an initial pool of approximately 15 billion images to about 250 million, or 1.7 percent, and 1 billion video clips to roughly 50 million, or 5 percent 1.

The blog also highlights challenges in video data annotation and filtering 1. A 2025 study from Carnegie Mellon University revealed that film cinematography experts had a disagreement rate of about 24 percent with the ground truth when annotating camera movements, requiring multiple iterations to reach a 96 percent agreement rate 1. Furthermore, the team utilized WAFT, a state-of-the-art optical flow predictor, to filter out long-tail low-motion videos 1.