Video generation has advanced rapidly from short, low-quality clips to high-resolution, long-duration sequences with complex spatiotemporal dynamics. However, despite strong generative priors learned through large-scale pretraining, pretrained video models often struggle to reliably follow human intent, maintain temporal coherence, or satisfy physical and safety constraints, according to a new survey published in TMLR.
The survey, titled Video Generation Models: A Survey of Post-Training and Alignment, presents what its authors describe as the first comprehensive review of post-training and alignment in video generation. It frames post-training as a unifying framework and distinguishes between implicit and explicit alignment based on how alignment signals are enforced.
Existing approaches are organized into four broad categories: supervised fine-tuning methods, self-training and distillation methods, preference- and reward-based methods, and inference-time methods. This taxonomy aims to provide a coherent view of how alignment signals shape model behavior across both training and deployment.
The authors note that alignment in video generation poses unique challenges compared with image and text generation, including error accumulation over time, motion-appearance coupling, multi-objective trade-offs, and limited supervision for temporal properties. These challenges motivate systematic post-training strategies that adapt pretrained models without retraining them from scratch.
Beyond methodological advances, the survey reviews commonly used datasets, benchmarks, and evaluation practices, and discusses open challenges such as scalable reward design, long-horizon temporal consistency, stability-expressiveness trade-offs, and safety-aware generation. A companion repository with a continuously updated collection of papers and resources is also maintained.